Making AI Inclusive: 4 Guiding Principles for Ethical Engagement

Tina Park

Introduction

While the concept of “human-centered design” is hardly new to the technology sector, recent years have seen growing efforts to build inclusive artificial intelligence (AI) and machine learning (ML) products. Broadly, inclusive AI/ML refers to algorithmic systems which are created with the active engagement of and input from people who are not on AI/ML development teams. This includes both end users of the systems and non-users who are impacted by the systems.“Impacted non-user” refers to people who are impacted by the deployment of an AI/ML system, but are not the direct user or customer of that system. For example, in the case of students in the United Kingdom in 2020 whose A-level grades were determined by an algorithm, the “user” of the algorithmic system is Ofqual, the official exam regulator in England, and the students are “impacted non-users.” To collect this input, practitioners are increasingly turning to engagement practices like user experience (UX) research and participatory design.

Amid rising awareness of structural inequalities in our society, embracing inclusive research and design principles helps signal a commitment to equitable practices. As many proponents have pointed out, it also makes for good business: Understanding the needs of a more diverse set of people expands the market for a given product or service. Once engaged, these people can then further improve an AI/ML product, identifying issues like bias in algorithmic systems.

Despite these benefits, however, there remain significant challenges to greater adoption of inclusive development in the AI/ML field. There are also important opportunities. For AI practitioners, AI ethics researchers, and others interested in learning more about responsible AI, this Partnership on AI (PAI) white paper provides guidance to help better understand and overcome the challenges related to engaging stakeholders in AI/ML development.

Ambiguities around the meaning and goals of “inclusion” present one of the central challenges to AI/ML inclusion efforts. To make the changes needed for a more inclusive AI that centers equity, the field must first find agreement on foundational premises regarding inclusion. Recognizing this, this white paper provides four guiding principles for ethical engagement grounded in best practices:

  1. All participation is a form of labor that should be recognized
  2. Stakeholder engagement must address inherent power asymmetries
  3. Inclusion and participation can be integrated across all stages of the development lifecycle
  4. Inclusion and participation must be integrated to the application of other responsible AI principles

To realize ethical participatory engagement in practice, this white paper also offers three recommendations aligned with these principles for building inclusive AI:

  1. Allocate time and resources to promote inclusive development
  2. Adopt inclusive strategies before development begins
  3. Train towards an integrated understanding of ethics

This white paper’s insights are derived from the research study “Towards An Inclusive AI: Challenges and Opportunities for Public Engagement in AI Development.” That study drew upon discussions with industry experts, a multidisciplinary review of existing research on stakeholder and public engagement, and nearly 70 interviews with AI practitioners and researchers, as well as data scientists, UX researchers, and technologists working on AI and ML projects, over a third of whom were based in areas outside of the US, EU, UK, or Canada. Supplemental interviews with social equity and Diversity, Equity, and Inclusion (DEI) advocates contributed to the development of recommendations for individual practitioners, business team leaders, and the field of AI and ML more broadly.

This white paper does not provide a step-by-step guide for implementing specific participatory practices. It is intended to renew discussions on how to integrate a wider range of insights and experiences into AI/ML technologies, including those of both users and the people impacted (either directly or indirectly) by these technologies. Such conversations — between individuals, inside teams, and within organizations — must be had to spur the changes needed to develop truly inclusive AI.

Making AI Inclusive: 4 Guiding Principles for Ethical Engagement

Introduction

Guiding Principles for Ethical Participatory Engagement

Principle 1: All Participation Is a Form of Labor That Should Be Recognized

Principle 2: Stakeholder Engagement Must Address Inherent Power Asymmetries

Principle 3: Inclusion and Participation Can Be Integrated Across All Stages of the Development Lifecycle

Principle 4: Inclusion and Participation Must Be Integrated to the Application of Other Responsible AI Principles

Recommendations for Ethical Engagement in Practice

Recommendation 1: Allocate Time and Resources to Promote Inclusive Development

Recommendation 2: Adopt Inclusive Development Strategies Before Development Begins

Recommendation 3: Train Towards an Integrated Understanding of Ethics

Conclusion

Acknowledgements

Sources Cited

  1. Jean-Baptiste, A. (2020). Building for Everyone: Expand Your Market with Design Practices from Google’s Product Inclusion Team. John Wiley and Sons, Inc.
  2. Romao, M. (2019, June 27). “A vision for AI: Innovative, Trusted and Inclusive.” Policy@Intel. https://community.intel.com/t5/Blogs/Intel/Policy-Intel/A-vision-for-AI-Innovative-Trusted-and-Inclusive/post/1333103
  3. Zhou, A., Madras, D., Raji, D., Milli, S., Kulynych, B. and Zemel, R. (2020, July 17). “Participatory Approaches to Machine Learning.” (Workshop). International Conference on Machine Learning 2020.
  4. Lewis, J. E., Abdilla, A., Arista, N., Baker, K., Benesiinaabandan, S., Brown, M., ... and Whaanga, H. (2020). Indigenous protocol and artificial intelligence position paper. Indigenous AI. https://www.indigenous-ai.net/position-paper
  5. Costanza-Chock, S. (2020). Design justice: Community-led practices to build the worlds we need. The MIT Press.
  6. Hamraie, A., and Fritsch, K. (2019). “Crip technoscience manifesto.” Catalyst: Feminism, Theory, Technoscience, 5(1), 1-33. https://catalystjournal.org/index.php/catalyst/article/view/29607
  7. Taylor, L. (2017). “What is data justice? The case for connecting digital rights and freedoms globally.” Big Data and Society, 4(2), 2053951717736335. https://doi.org/10.1177/2053951717736335
  8. Benjamin, Ruha. (2019). Race After Technology: Abolitionist Tools for the New Jim Code. Polity. https://www.ruhabenjamin.com/race-after-technology
  9. Hanna, A., Denton, E., Smart, A., and Smith-Loud, J. (2020). “Towards a critical race methodology in algorithmic fairness.” In Proceedings of the 2020 conference on fairness, accountability, and transparency (pp. 501-512). https://arxiv.org/abs/1912.03593
  10. Sloane, M., Moss, E., Awomolo, O. and Forlano, L. (2020). ''Participation is not a design fix for machine learning.'' arXiv. https://arxiv.org/abs/2007.02423
  11. Cifor, M., Garcia, P., Cowan, T.L., Rault, J., Sutherland, T., Chan, A., Rode, J., Hoffmann, A.L., Salehi, N. and Nakamura, L. (2019). “Feminist Data Manifest-No.” Feminist Data Manifest-No. Retrieved October 1, 2020 from https://www.manifestno.com/home
  12. Harrington, C., Erete, S. and Piper, A.M. (2019). “Deconstructing Community-Based Collaborative Design: Towards More Equitable Participatory Design Engagements.” In Proceedings of the ACM on Human-Computer Interaction 3(CSCW):1–25. https://doi.org/10.1145/3359318
  13. Freimuth V.S., Quinn, S.C., Thomas, S.B., Cole, G., Zook, E and Duncan, T. (2001). “African Americans’ Views on Research and the Tuskegee Syphilis Study.” Social Science and Medicine 52(5):797–808. https://doi.org/10.1016/S0277-9536(00)00178-7
  14. George, S., Duran, N. and Norris, K. (2014). “A Systematic Review of Barriers and Facilitators to Minority Research Participation Among African Americans, Latinos, Asian Americans, and Pacific Islanders.” American Journal of Public Health 104(2):e16–31. https://doi.org/10.2105/AJPH.2013.301706
  15. Barabas, C., Doyle, C., Rubinovitz, J.B., and Dinakar, K. (2020). “Studying Up: Reorienting the Study of Algorithmic Fairness around Issues of Power.” In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (pp. 167-176).
  16. Harrington, C., Erete, S. and Piper, A.M.. (2019). “Deconstructing Community-Based Collaborative Design: Towards More Equitable Participatory Design Engagements.” In Proceedings of the ACM on Human-Computer Interaction 3(CSCW):1–25. https://dl.acm.org/doi/10.1145/3359318.
  17. Chan, A., Okolo, C. T., Terner, Z., and Wang, A. (2021). “The Limits of Global Inclusion in AI Development.” arXiv. https://arxiv.org/abs/2102.01265
  18. Sanders, E. B. N. (2002). “From user-centered to participatory design approaches.” In Design and the social sciences (pp. 18-25). CRC Press. https://www.taylorfrancis.com/chapters/edit/10.1201/9780203301302-8/user-centered-participatory-design-approaches-elizabeth-sanders
  19. Leslie, D., Katell, M., Aitken, M., Singh, J., Briggs, M., Powell, R., ... and Burr, C. (2022). “Data Justice in Practice: A Guide for Developers.” arXiv. https://arxiv.org/ftp/arxiv/papers/2205/2205.01037.pdf
  20. Zdanowska, S., and Taylor, A. S. (2022). “A study of UX practitioners roles in designing real-world, enterprise ML systems.” In CHI Conference on Human Factors in Computing Systems (pp. 1-15). https://dl.acm.org/doi/abs/10.1145/3491102.3517607
  21. Leslie, D., Katell, M., Aitken, M., Singh, J., Briggs, M., Powell, R., ... and Burr, C. (2022). “Data Justice in Practice: A Guide for Developers.” arXiv. https://arxiv.org/ftp/arxiv/papers/2205/2205.01037.pdf
  22. Saulnier, L., Karamcheti, S., Laurençon, H., Tronchon, L., Wang, T., Sanh, V., Singh, A., Pistilli, G., Luccioni, S., Jernite, Y., Mitchell, M. and Kiela, D. (2022). “Putting Ethical Principles at the Core of the Research Lifecycle.” Hugging Face Blog. Retrieved from https://huggingface.co/blog/ethical-charter-multimodal
  23. Ada Lovelace Institute. (2021). “Participatory data stewardship: A framework for involving people in the use of data.” Ada Lovelace Institute. https://www.adalovelaceinstitute.org/report/participatory-data-stewardship/
  24. Delgado, F., Yang, S., Madaio, M., and Yang, Q. (2021). “Stakeholder Participation in AI: Beyond ‘Add Diverse Stakeholders and Stir.’” arXiv. https://arxiv.org/pdf/2111.01122.pdf
  25. Sloane, M., Moss, E., Awomolo, O. and Forlano, L. (2020). ''Participation is not a design fix for machine learning.'' arXiv. https://arxiv.org/abs/2007.02423
Table of Contents
1
2
3
4
5
6

After the Offer: The Role of Attrition in AI’s ‘Diversity Problem’

Jeffrey Brown

Executive Summary

As a field, AI struggles to retain team members from diverse backgrounds. Given the far-reaching effects of algorithmic systems and the documented harms to marginalized communities, the fact that these communities are not represented on AI teams is particularly troubling. Why is this such a widespread phenomenon and what can be done to close the gap? This research paper, “After the Offer: The Role of Attrition in AI’s ‘Diversity Problem’” seeks to answer these questions, providing four recommendations for how organizations can make the AI field more inclusive. Click the button below to download a summary of these recommendations or continue on to read the paper in full.

Summary of Recommendations

Amid heightened attention to society-wide racial and social injustice, organizations in the AI space have been urged to investigate the harmful effects that AI has had on marginalized populations. It’s an issue that engineers, researchers, project managers, and various leaders in both tech companies and civil society organizations have devoted significant time and resources to in recent years. In examining the effects of AI, organizations must consider who exactly has been designing these technologies.

Diversity reports have revealed that the people working at the organizations that develop and deploy AI lack diversity across several dimensions. While organizations have blamed pipeline problems in the past, research has increasingly shown that once workers belonging to minoritized identities get hired in these spaces, systemic difficulties affect their experiences in ways that their peers from dominant groups do not have to worry about.

Attrition in the tech industry is a problem that disproportionately affects minoritized workers. In AI, where technologies already have a disproportionately negative impact on these communities, this is especially troublesome.

Minoritized Workers

This report uses minoritized workers as an umbrella term to refer to people whose identities (in categories such as race, ethnicity, gender, or ability) have been historically marginalized by those in dominant social groups. The minoritized workers in this study include people who identified as minoritized within the identity categories of race and ethnicity, gender identity, sexual orientation, ability, and immigration status. Because this study was international in scope, it is important to note that these categories are relative to their social context.

We are left wondering: What leads to these folks leaving their teams, organizations, or even the AI field more broadly? What about the AI field in particular influences these people to stay or leave? And what can organizations do to stem this attrition to make their environments more inclusive?

The current study uses interviews with folks belonging to minoritized identities across the AI field, managers, and DEI (diversity, equity, and inclusion)- leaders in tech to get rich information about what aspects of cultures within an organization promote inclusion or contribute to attrition. Themes that emerged during these interviews formed 3 key takeaways:

  1. Diversity makes for better team climates
  2. Systemic supports are difficult but necessary to undo the current harms to minoritized workers
  3. Individual efforts to change organizational culture fall disproportionately on minoritized folks who are usually not professionally rewarded for their efforts

In line with these takeaways, the study makes 4 recommendations about what can be done to make the AI field more inclusive for workers:

  1. Organizations must systemically support ERGs
  2. Organizations must intentionally diversify leadership and managers
  3. DEI trainings must be specific in order to be effective and be more connected to the content of AI work
  4. Organizations must interrogate their values as practiced and fundamentally alter them to include the perspectives of people who are not White, cis, or male

These takeaways and recommendations are explored in more depth below.

Key Takeaways

Key Takeaways

1. Diversity makes for better team climates

Across interviews, participants consistently expressed that managers who belonged to minoritized identities or who took the time to learn about working with diverse identities were more supportive of their needs and career goals. Such efforts reportedly resulted in teams that were also more diverse, inclusive, interdisciplinary, and engendering of a positive team culture/climate. In these environments, workers belonging to minoritized identities thrived. A diversity in backgrounds and perspectives was particularly important for AI teams that needed to solve interdisciplinary problems.

Conversely, the negative impact of work environments that were sexist or where participants experienced acts of prejudice such as microaggressions was also a recurring theme.

While collaborative or positive work environments were also a common theme, such environments did not in themselves negate predominant cultures which deprioritized “DEI-focused” work, work that was highly interdisciplinary, or work that did not serve the dominant group. Negative organizational cultures seemed to exacerbate experiences of prejudice or discrimination on AI teams.

2. Systemic supports are difficult but necessary to undo the current harms to minoritized workers

Participants belonging to minoritized identities said that they either left or intended to leave organizations that did not support their continued career growth or possessed values that did not align with their own. Consistent with this, participants described examples of their organizations not valuing the content of their work.

Participants also tied their desires to leave with instances of prejudice or discrimination, which may also be related to “toxic” work environments. Some participants reported instances of being tokenized or being subject to negative stereotypes about their identity groups, somewhat reflective of wider contexts in tech beyond AI.

Systemic supports include incentive structures that allow minoritized workers to succeed at every level, from the teams that they work with actively validating their experiences to their managers finding the best ways for them to deliver work products in accordance with both individual and institutional needs. Guidelines for promotion that recognize the barriers these workers face in environments mostly occupied by dominant group norms are another important support.

3. Individual efforts to change organizational culture fall disproportionately on minoritized folks who are usually not professionally rewarded for their efforts

Individuals discussed ways in which they tried to make their workplaces or teams more inclusive or otherwise sought to incorporate diverse perspectives into their work around AI. Participants sometimes had to contend with bias against DEI efforts, reporting that other workers in their organizations would dismiss their efforts as lacking rigor or focus on the product.

There were some institutional efforts to foster a more inclusive culture, most commonly DEI trainings. DEI trainings that were very specific to some groups (e.g., gender diverse folks, Black people) were reported as being the most effective. However, even when they were specific, DEI trainings seemed to be disconnected from some aspects of the workplace climate or the content of what teams were working on.

Participants who mentioned Employee Resource Groups (ERGs) uniformly praised them, discussing the huge positive impact they had on a personal level, forming the bases of their social support networks in their organizations and having a strong impact on their ability to integrate aspects of their identities or other “DEI topics” they were passionate about into their work.

Recommendations

Recommendations

1. Organizations must systemically support ERGs

Employees specifically named ERGs as one of their main sources of support even in work environments that were otherwise toxic.. Additionally, ERGs provided built-in mentorship for those who did not have ready access to mentors or whose supervisors had not done the work to understand the kinds of support needed for those of minoritized identities to thrive in predominantly White and male environments.

What makes this recommendation work?

Within these ERGs, there existed other grass-roots initiatives that supported workers, such as informal talking circles and networks of employees that essentially provided peer mentoring that participants found crucial to navigating White- and male-dominated spaces. The mentorship provided by ERGs was also essential when HR failed to provide systemic support for staff and instead prioritized protecting the organization.

What must be in place?

While participants uniformly praised ERGs, they required large amounts of time from staff members that detracted from their work. Such groups also ran the risk of getting taken over by leadership and having their original mission derailed. Institutions should seek a balance between supporting these groups and giving them the freedom to organize in pursuit of their own best interests.

What won’t this solve?

ERGs will not necessarily make an organization’s AI or tech more inclusive. Rather, systematically supporting ERGs will provide more support and community for minoritized workers, which is meant to promote a more inclusive workplace in general.

2. Organizations must intentionally diversify leadership and managers
What makes this recommendation work?

Participants repeatedly pointed to managers and upper-level leaders who belonged to minoritized identities (especially racial ones) as important influences, changing policy that permeated through various levels of their organizations. A diverse workforce may also bring with it multiple perspectives, including those belonging to people from different disciplines who may be interested in working in the AI field due to the opportunity for interdisciplinary collaboration, research, and product development. Bringing in folks from various academic, professional, and technical backgrounds to solve problems is especially crucial for AI teams.

What must be in place?

There must be understanding about the reasons behind the lack of diversity and the “bigger picture” of how powerful groups more easily perpetuate power structures already in place. Participants spoke of managers who did not belong to minoritized identities themselves but who took the time to learn in depth about differences in power and privilege in the tech ecosystem, appreciating the diverse perspectives that workers brought. These managers, while not perfect, tended to take advocating for their reports very seriously, particularly female reports who often went overlooked.

What won’t this solve?

Intentionally diversifying leadership and managers will not automatically create a pipeline for diversity at the leadership level, nor will it automatically override institutional culture or policies that ignore DEI best practices.

3. DEI trainings must be specific in order to be effective and be more connected to the content of AI work
What makes this recommendation work?

Almost all participants reported that their organizations mandated some form of DEI training for all staff. These ranged widely, from very general ones to very specific trainings that discussed cultural competency about more specific groups of people (e.g., participants reported that there were trainings on anti-Black racism). Participants discussed that the more specific trainings tended to be more impactful.

What must be in place?

Organizations must invest in employees who see the importance of inclusive values in AI research and product design. Participants pointed to the importance of managers who had an ability to foster inclusive team values, which was not something that HR could mandate.

What won’t this solve?

As several participants observed, DEI trainings will not uproot or counteract institutional stigmas against DEI. It would take sustained effort and deliberate alignment of values for an organization to emphasize DEI in its work.

4. Organizations must interrogate their values as practiced and fundamentally alter them to include the perspectives of people who are not White, cis, or male
What makes this recommendation work?

Participants frequently reported that a misalignment of values was a primary reason for them leaving their organizations or wanting to leave their organizations. Participants in this sample discussed joining the AI field to create a positive impact while growing professionally. This led them to feeling disappointed when their organizations did not prioritize these goals (despite them being among their stated values).

What must be in place?

Participants found it frustrating when organizations stated that they valued diversity and then failed to live up to this value with hiring, promotion, and day-to-day operations, ignoring the voices of minoritized individuals. If diversity is truly a value, organizations may have to investigate their systems of norms and expectations that are fundamentally male, Eurocentric, and do not make space for those from diverse backgrounds. They then must take additional steps to consider how such systems influence their work in AI.

What won’t this solve?

Because achieving a fundamental re-alignment like this is a more comprehensive solution, it cannot satisfy the most immediate and urgent needs for reform. Short-term, organizations must work with DEI professionals to recognize how they are perpetuating potentially harmful norms of the dominant group and work to create policies that are more equitable. Longer term fixes may not, for instance, satisfy the immediate and urgent need for more diversity in leadership and teams in general.

After the Offer: The Role of Attrition in AI’s ‘Diversity Problem’

Executive Summary

Key Takeaways

Recommendations

Introduction

Why Study Attrition of Minoritized Workers in AI?

Background

Problems Due to Lack of Diversity of AI Teams

More Diverse Teams Yield Better Outcomes

Current Level of Diversity in Tech

Diversity in AI

What Has Been Done

What Has Been Done

What Has Been Done

Attrition in Tech

Current Study and Methodology

Recruitment

Participants

Measure

Procedure

Analysis

Results

Attrition

Culture

Efforts to Improve Inclusivity

Summary and the Path Forward

Acknowledgements

Appendices

Appendix 1: Recruitment Document

Appendix 2: Privacy Document

Appendix 3: Research Protocol

Appendix 4: Important Terms

Sources Cited

  1. Buolamwini, J., u0026amp; Gebru, T. (2018, January). Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on fairness, accountability and transparency (pp. 77-91). PMLR.
  2. Zhao, D., Wang, A., u0026amp; Russakovsky, O. (2021). Understanding and Evaluating Racial Biases in Image Captioning. arXiv preprint arXiv:2106.08503.
  3. Feldstein, S. (2021). The Global Expansion of AI Surveillance. Carnegie Endowment for International Peace. Retrieved 17 September 2019, from https://carnegieendowment.org/2019/09/17/global-expansion-of-ai-surveillance-pub-79847.
  4. Firth, N. (2021). Apple Card is being investigated over claims it gives women lower credit limits. MIT Technology Review. Retrieved 23 November 2021, from https://www.technologyreview.com/2019/11/11/131983/apple-card-is-being-investigated-over-claims-it-gives-women-lower-credit-limits/.
  5. Howard, A., u0026amp; Isbell, C. (2021). Diversity in AI: The Invisible Men and Women. MIT Sloan Management Review. Retrieved 21 September 2020, from https://sloanreview.mit.edu/article/diversity-in-ai-the-invisible-men-and-women/.
  6. AI Now. (2019). Discriminating Systems: Gender, Race, and Power in AI (Ebook). Retrieved 23 November 2021.
  7. Swauger, S. (2021). Opinion | What's worse than remote school? Remote test-taking with AI proctors. NBC News. Retrieved 7 November 2020, from https://www.nbcnews.com/think/opinion/remote-testing-monitored-ai-failing-students-forced-undergo-it-ncna1246769
  8. Belani, G. (2021). AI Paving the Way for Remote Work | IEEE Computer Society. Computer.org. Retrieved 26 July 2021, from https://www.computer.org/publications/tech-news/trends/remote-working-easier-with-ai
  9. Scott, A., Kapor Klein, F., and Onovakpuri, U. (2017). Tech Leavers Study (Ebook). Retrieved 24 November 2021, from https://www.kaporcenter.org/wp-content/uploads/2017/08/TechLeavers2017.pdf
  10. Women in the Workplace (2021). 2021. Retrieved 23 November 2021, from https://www.mckinsey.com/featured-insights/diversity-and-inclusion/women-in-the-workplace
  11. Silicon Valley Bank. (2021). 2020 Global Startup Outlook: Key insights from the Silicon Valley Bank startup outlook survey (Ebook). Retrieved 23 November 2021, from https://www.svb.com/globalassets/library/uploadedfiles/content/trends_and_insights/reports/startup_outlook_report/suo_global_report_2020-final.pdf
  12. Firth, N. (2021). Apple Card is being investigated over claims it gives women lower credit limits. MIT Technology Review. Retrieved 23 November 2021, from https://www.technologyreview.com/2019/11/11/131983/apple-card-is-being-investigated-over-claims-it-gives-women-lower-credit-limits/.
  13. Tomasev, N., McKee, K.R., Kay, J., u0026amp; Mohamed, S. (2021). Fairness for Unobserved Characteristics: Insights from technological impacts on queer communities. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society (AIES ’21), Retrieved October 1, 2021 from https://doi.org/10.1145/3461702.3462540
  14. Martinez, E., u0026amp; Kirchner, L. (2021). The secret bias hidden in mortgage-approval algorithms | AP News. AP News. Retrieved 24 November 2021, from https://apnews.com/article/lifestyle-technology-business-race-and-ethnicity-mortgages-2d3d40d5751f933a88c1e17063657586
  15. Turner Lee, N., Resnick, P., u0026amp; Barton, G. (2021). Algorithmic bias detection and mitigation: Best practices and policies to reduce consumer harms. Brookings. Retrieved 24 November 2021, from https://www.brookings.edu/research/algorithmic-bias-detection-and-mitigation-best-practices-and-policies-to-reduce-consumer-harms/.
  16. Rock, D., u0026amp; Grant, H. (2016). Why diverse teams are smarter. Harvard Business Review, 4(4), 2-5.
  17. Wang, J., Cheng, G. H. L., Chen, T., u0026amp; Leung, K. (2019). Team creativity/innovation in culturally diverse teams: A meta‐analysis. Journal of Organizational Behavior, 40(6), 693-708.
  18. Lorenzo, R., Voigt, N., Tsusaka, M., Krentz, M., u0026amp; Abouzahr, K. (2018). How Diverse Leadership Teams Boost Innovation. BCG Global. Retrieved 24 November 2021, from https://www.bcg.com/publications/2018/how-diverse-leadership-teams-boost-innovation
  19. Hoobler, J. M., Masterson, C. R., Nkomo, S. M., u0026amp; Michel, E. J. (2018). The business case for women leaders: Meta-analysis, research critique, and path forward. Journal of Management, 44(6), 2473-2499.
  20. Chakravorti, B. (2020). To Increase Diversity, U.S. Tech Companies Need to Follow the Talent. Harvard Business Review. Retrieved 24 November 2021, from https://hbr.org/2020/12/to-increase-diversity-u-s-tech-companies-need-to-follow-the-talent.
  21. Accenture. (2018). Getting to Equal 2018: The Disability Inclusion Advantage. Retrieved from https://www.accenture.com/_acnmedia/pdf-89/accenture-disability-inclusion-research-report.pdf
  22. Whittaker, M., Alper, M., Bennett, C. L., Hendren, S., Kaziunas, L., Mills, M., ... u0026amp; West, S. M. (2019). Disability, bias, and AI. AI Now Institute.
  23. Heater, B. (2020). Tech companies respond to George Floyd’s death, ensuing protests and systemic racism. Techcrunch.com. Retrieved 24 November 2021, from https://techcrunch.com/2020/06/01/tech-co-protests/.
  24. Google (2021). 2021 Diversity Annual Report. Retrieved 24 November 2021, from https://static.googleusercontent.com/media/diversity.google/en//annual-report/static/pdfs/google_2021_diversity_annual_report.pdf?cachebust=2e13d07.
  25. Facebook. (2021). Facebook Diversity Update: Increasing Representation in Our Workforce and Supporting Minority-Owned Businesses | Meta. Meta. Retrieved 24 November 2021, from https://about.fb.com/news/2021/07/facebook-diversity-report-2021/.
  26. Amazon Staff. (2020). Our workforce data. US About Amazon. Retrieved 24 November 2021, from https://www.aboutamazon.com/news/workplace/our-workforce-data
  27. Adobe. (2021). Adobe Diversity By the Numbers. Adobe. Retrieved 24 November 2021, from https://www.adobe.com/diversity/data.html
  28. National Center for Women in Tech. (2020). NCWIT Scorecard: The Status of Women in Computing (2020 Update). Retrieved https://ncwit.org/resource/scorecard/
  29. Center for American Progress (2012). The State of diversity in Today’s workforce. Retrieved from https://www.americanprogress.org/article/the-state-of-diversity-in-todays-workforce/
  30. Gillenwater, S. (2020). Meet the CIOs of the Fortune 500 — 2021 edition. Boardroom Insiders. Retrieved from https://www.boardroominsiders.com/blog/meet-the-cios-of-the-fortune-500-2021-edition
  31. Stack Overflow. (2020). 2020 Developer Survey. Retrieved from https://insights.stackoverflow.com/survey/2020#developer-profile-disability-status-mental-health-and-differences
  32. Stanford HAI. (2021). The AI Index Report: Measuring Trends in Artificial intelligence (Ebook). Retrieved 24 November 2021, from https://aiindex.stanford.edu/wp-content/uploads/2021/03/2021-AI-Index-Report-_Chapter-6.pdf.
  33. Chi, N., Lurie, E., u0026amp; Mulligan, D. K. (2021). Reconfiguring Diversity and Inclusion for AI Ethics. arXiv preprint arXiv:2105.02407.
  34. Selyukh, A. (2016). Why Some Diversity Thinkers Aren't Buying The Tech Industry's Excuses. NPR. Retrieved 24 November 2021, from https://www.npr.org/sections/alltechconsidered/2016/07/19/486511816/why-some-diversity-thinkers-arent-buying-the-tech-industrys-excuses.
  35. National Association for Educational Progress. (2020). NAEP Report Card: Mathematics. Retrieved from https://www.nationsreportcard.gov/mathematics/nation/achievement/?grade=4
  36. Ladner, R. (2021). Expanding the pipeline: The status of persons with disabilities in the Computer Science Pipeline. Retrieved February 1, 2022, from https://cra.org/cra-wp/expanding-the-pipeline-the-status-of-persons-with-disabilities-in-the-computer-science-pipeline/
  37. Center for Evaluating the Research Pipeline (2021). “Data Buddies Survey 2019 Annual Report”. Computing Research Association, Washington, D.C.
  38. Code.org. (2021). Code.org's Approach to Diversity u0026amp; Equity in Computer Science. Code.org. Retrieved 24 November 2021, from https://code.org/diversity
  39. Zweben, S., u0026amp; Bizot, B. (2021). 2020 Taulbee Survey: Bachelor’s and Doctoral Degree Production Growth Continues but New Student Enrollment Shows Declines (Ebook). Computing Research Association. Retrieved 24 November 2021, from https://cra.org/wp-content/uploads/2021/05/2020-CRA-Taulbee-Survey.pdf
  40. Computing Research Association (2017). Generation CS: Computer Science Undergraduate Enrollments Surge Since 2006
  41. The Higher Education Statistics Agency (2021). Higher Education Student Statistics. Retrieved from https://www.hesa.ac.uk/news/16-01-2020/sb255-higher-education-student-statistics/subjects
  42. BCS. (2014). Women in IT Survey (Ebook). BCS: The Chartered Institute for IT. Retrieved 24 November 2021, from https://www.bcs.org/media/4446/women-it-survey.pdf
  43. Inclusive Boards. (2018). Inclusive Tech Alliance Report 2018 (Ebook). Retrieved 24 November 2021, from https://www.inclusivetechalliance.co.uk/wp-content/uploads/2019/07/Inclusive-Tech-Alliance-Report.pdf.
  44. Atomico. (2020). The State of European Tech 2020. 2020.stateofeuropeantech.com. Retrieved 24 November 2021, from https://2020.stateofeuropeantech.com/chapter/diversity-inclusion/article/diversity-inclusion/.
  45. Chung-Yan, G. A. (2010). The nonlinear effects of job complexity and autonomy on job satisfaction, turnover, and psychological well-being. Journal of occupational health psychology, 15(3), 237.
  46. McKnight, D. H., Phillips, B., u0026amp; Hardgrave, B. C. (2009). Which reduces IT turnover intention the most: Workplace characteristics or job characteristics?. Information u0026amp; Management, 46(3), 167-174.
  47. Vaamonde, J. D., Omar, A., u0026amp; Salessi, S. (2018). From organizational justice perceptions to turnover intentions: The mediating effects of burnout and job satisfaction. Europe's journal of psychology, 14(3), 554.
  48. Instructure (2019). How to get today's employees to stay and engage? Develop their careers. PR Newswire. Retrieved from https://www.prnewswire.com/news-releases/how-to-get-todays-employees-to-stay-and-engage-develop-their-careers-300860067.html
  49. McCarty, E. (2021). Integral and The Harris Poll Find Employees are giving Employers a Performance Review - Integral. Integral. Retrieved 24 November 2021, from https://www.teamintegral.com/2021/news-release-integral-employee-activation-index/
  50. McCarty, E. (2021). Integral and The Harris Poll Find Employees are giving Employers a Performance Review - Integral. Integral. Retrieved 24 November 2021, from https://www.teamintegral.com/2021/news-release-integral-employee-activation-index/
  51. Bureau of Labor Statistics. (2021). News Release - The Employment Situation - October 2021 (Ebook). Retrieved 24 November 2021, from https://www.bls.gov/news.release/pdf/empsit.pdf
  52. Scott, A., Kapor Klein, F., u0026amp; Onovakpuri, U. (2017). Tech Leavers Study (Ebook). Retrieved 24 November 2021, from https://www.kaporcenter.org/wp-content/uploads/2017/08/TechLeavers2017.pdf.
  53. Young, E., Wajcman, J. and Sprejer, L. (2021). Where are the Women? Mapping the Gender Job Gap in AI. Policy Briefing: Full Report. The Alan Turing Institute.
  54. Metz, C. (2021). A second Google A.I. researcher says the company fired her.. Nytimes.com. Retrieved 24 November 2021, from https://www.nytimes.com/2021/02/19/technology/google-ethical-artificial-intelligence-team.html
  55. Myrow, R. (2021). Pinterest Sounds A More Contrite Tone After Black Former Employees Speak Out. Npr.org. Retrieved 24 November 2021, from https://www.npr.org/2020/06/23/881624553/pinterest-sounds-a-more-contrite-tone-after-black-former-employees-speak-out
  56. Scheer, S. (2021). The Tech Sector’s Big Disability Inclusion Problem. ERE. Retrieved from https://www.ere.net/the-tech-sectors-big-disability-inclusion-problem/
  57. Robinson, O. C. (2014). Sampling in interview-based qualitative research: A theoretical and practical guide. Qualitative research in psychology, 11(1), 25-41.
  58. Yancey, A. K., Ortega, A. N., u0026amp; Kumanyika, S. K. (2006). Effective recruitment and retention of minority research participants. Annu. Rev. Public Health, 27, 1-28.
  59. Hill, C. E., Knox, S., Thompson, B. J., Williams, E. N., Hess, S. A., u0026amp; Ladany, N. (2005). Consensual qualitative research: An update. Journal of counseling psychology, 52(2), 196.
  60. Gunaratnam, Y. (2003). Researching'race'and ethnicity: Methods, knowledge and power. Sage.
  61. Race and Ethnicity. American Sociological Association. (2022). Retrieved 29 January 2022, archived at https://web.archive.org/web/20190821170406/https://www.asanet.org/topics/race-and-ethnicity
  62. University of Minnesota Libraries (2022). 10.2 The Meaning of Race and Ethnicity. Open.lib.umn.edu. Retrieved 29 January 2022, from https://open.lib.umn.edu/sociology/chapter/10-2-the-meaning-of-race-and-ethnicity/.
  63. Sue, Derald Wing, Christina M. Capodilupo, Gina C. Torino, Jennifer M. Bucceri, Aisha Holder, Kevin L. Nadal, and Marta Esquilin.
  64. https://adata.org/glossary-terms#D
Table of Contents
1
2
3
4
5
6
7
8
9

ABOUT ML Foundational Resource

Overview


ABOUT ML (Annotation and Benchmarking on Understanding and Transparency of Machine learning Lifecycles) is a multi-year, multi-stakeholder initiative aimed at building transparency into the AI development process, industry-wide, through full lifecycle documentation. On this page, you will find the collected outputs of ABOUT ML, a library of resources designed to help organizations and individuals begin implementing transparency at scale. To further increase the usability of these resources, recommended reading plans for different readers are provided below.

Learn more about the origins of ABOUT ML and contributors to the project here.

Recommended Reading Plans

At the foundation of these resources lies the newly revised ABOUT ML Reference Document, which both identifies transparency goals and offers suggestions on how they might be achieved. Using principles provided by the Reference Document and insights about implementation gathered through our research, PAI plans to release additional ML documentation guides, templates, recommendations, and other artifacts. These future artifacts will also be available on this page.

Read the full ABOUT ML Reference Document

 

Recommended Reading Plans for…


ML System Developers/Deployers

ML system developers/deployers are encouraged to do a deep dive exploration of Section 3: Preliminary Synthesized Documentation Suggestions and use it to highlight gaps in their current understanding of both data- and model-related documentation and planning needs. This group will most benefit from further participation in the ABOUT ML effort by engaging with the community in the forthcoming online forum and by testing the efficacy and applicability of templates and specifications to be published in the PLAYBOOK and PILOTS, which will be developed based on use cases as an opportunity to implement ML documentation processes within an organization.


ML System Procurers

ML system procurers might explore Section 2.2: Documentation to Operationalize AI Ethics Goals to get ideas about what concepts to include as requirements for models and data in future requests for proposals relevant to ML systems. Additionally, they could use Section 2.3: Research Themes on Documentation for Transparency to shape conversations with the business owners and requirements writers to further elicit detailed key performance indicators and measures for success for any procured ML systems.


Users of ML System APIs and/or Experienced End Users of ML Systems

Users of ML system APIs and/or experienced end users of ML systems might skim the document and review all of the coral Quick Guides to get a better understanding of how ML concepts are relevant to many of the tools they regularly use. A review of Section 2.1: Demand for Transparency and AI Ethics in ML Systems will provide insight into conditions where it is appropriate to use ML systems. This section also explains how transparency is a foundation for both internal accountability among the developers, deployers, and API users of an ML system and external accountability to customers, impacted non-users, civil society organizations, and policymakers.


Internal Compliance Teams

Internal compliance teams are encouraged to explore Section 4: Current Challenges of Implementing Documentation and use it to shape conversations with developer/deployment teams to find ways to measure compliance throughout the Machine Learning Lifecycle (MLLC).


External Auditors

External auditors could skim Appendix: Compiled List of Documentation Questions and familiarize themselves with high-level concepts as well as tactically operationalized tenets to look for in their determination of whether or not an ML System is well-documented.


Lay Users of ML Systems and/or Members of Low-Income Communities

Lay users of ML systems and/or members of low-income communities might skim the document and review all of the blue “How We Define” boxes in order to get an overarching understanding of the text’s contents. These users are encouraged to continue learning ABOUT ML systems by exploring how they might impact their everyday lives. Additional insights can be gathered from the Glossary section of the ABOUT ML Reference Document.

Managing the Risks of AI Research: Six Recommendations for Responsible Publication

PAI Staff

Once a niche research interest, artificial intelligence (AI) has quickly become a pervasive aspect of society with increasing influence over our lives. In turn, open questions about this technology have, in recent years, transformed into urgent ethical considerations. The Partnership on AI’s (PAI) new white paper, “Managing the Risks of AI Research: Six Recommendations for Responsible Publication,” addresses one such question: Given AI’s potential for misuse, how can AI research be disseminated responsibly?

Many research communities, such as biosecurity and cybersecurity, routinely work with information that could be used to cause harm, either maliciously or accidentally. These fields have thus established their own norms and procedures for publishing high-risk research. Thanks to breakthrough advances, AI technology has progressed rapidly in the past decade, giving the AI community less time to develop similar practices.

Recent pilots, such as OpenAI’s “staged release” of GPT-2 and the “broader impact statement” requirement at the 2020 NeurIPS conference, demonstrate a growing interest in responsible AI publication norms. Effectively anticipating and mitigating the potential negative impacts of AI research, however, will require a community-wide effort. As a first step towards developing responsible publication practices, this white paper provides recommendations for three key groups in the AI research ecosystem:

  • Individual researchers, who should disclose and report additional information in their papers and normalize discussion about the downstream consequences of research.
  • Research leadership, which should review potential downstream consequences earlier in the research pipeline and commend researchers who identify negative downstream consequences.
  • Conferences and journals, which should expand peer review criteria to include engagement with potential downstream consequences and establish separate review processes to evaluate papers based on risk and downstream consequences.

Additionally, this white paper includes an appendix which seeks to disambiguate a variety of terms related to responsible research which are often conflated: “research integrity,” “research ethics,” “research culture,” “downstream consequences,” and “broader impacts.”

This document represents an artifact that can be used as a basis for further discussion, and we seek feedback on it to inform future iterations of the recommendations it contains. Our aim is to help build our capacity as a field to anticipate downstream consequences and mitigate potential risks.

To read “Managing the Risks of AI Research: Six Recommendations for Responsible Publication” in full, click here.

Framework for Promoting Workforce Well-being in the AI-Integrated Workplace

PAI Staff

Executive Summary

Executive Summary

The Partnership on AI’s “Framework for Promoting Workforce Well-being in the AI- Integrated Workplace” provides a conceptual framework and a set of tools to guide employers, workers, and other stakeholders towards promoting workforce well-being throughout the process of introducing AI into the workplace.

As AI technologies become increasingly prevalent in the workplace, our goal is to place  workforce well-being at the center of this technological change and resulting metamorphosis in work, well-being, and society, and provide a starting point to discuss and create pragmatic solutions.

The paper categorizes aspects of workforce well-being that should be prioritized and protected throughout AI integration into six pillars. Human rights is the first pillar, and supports all aspects of workforce well-being. The five additional pillars of well-being include physical, financial, intellectual, emotional well-being, as well as purpose and meaning. The Framework presents a set of recommendations that organizations can use to guide organizational thinking about promoting well-being throughout the integration of AI in the workplace.

The Framework is designed to initiate and inform discussions on the impact of AI that strengthen the reciprocal obligations between workers and employers, while grounding that discourse in six pillars of worker well-being.

We recognize that the impacts of AI are still emerging and often difficult to distinguish from the impact of broader digital transformation, leading to organizations being challenged to address the unknown and potentially fundamental changes that AI may bring to the workplace. We strongly advise that management collaborate with workers directly or with worker representatives in the development, integration, and use of AI systems in the workplace, as well as in the discussion and implementation of this Framework.

We acknowledge that the contexts for having a dialogue on worker well-being may differ. For instance, in some countries there are formal structures in place such as workers’ councils that facilitate the dialogue between employers and workers. In other cases, countries or sectors do not have these institutions in place, nor a tradition for dialogue between the two parties. In all cases, the aim of this Framework is to be a useful tool for all parties to collaboratively ensure that the introduction of AI technologies goes hand in hand with a commitment to worker well-being. The importance of making such commitment in earnest has been highlighted by the COVID-19 public health and economic crises which exposed and exacerbated the long-standing inequities in the treatment of workers. Making sure those are not perpetuated further with the introduction of AI systems into the workplace requires deliberate efforts and will not happen automatically.

Recommendations

Recommendations

This section articulates a set of recommendations to guide organizational approaches and thinking on what to promote, what to be cognizant of, and what to protect against, in terms of worker and workforce well-being while integrating AI into the workplace. These recommendations are organized along the six well-being pillars identified above, and are meant to serve as a starting place for organizations seeking to apply the present Framework to promote workforce well-being throughout the process of AI integration. Ideally, these can be recognized formally as organizational commitments at the board level and subsequently discussed openly and regularly with the entire organization.

The “Framework for Promoting Workforce Well-Being in the AI-Integrated Workplace” is a product of the Partnership on AI’s AI, Labor, and the Economy (AILE) Expert Group, formed through a collaborative process of research, scoping, and iteration. In August 2019, at a workshop called “Workforce Well-being in the AI-Integrated Workplace” co-hosted by PAI and the Ford Foundation, this work received additional input from experts, academics, industry, labor unions, and civil society. Though this document reflects the inputs of many PAI Partner organizations, it should not under any circumstances be read as representing the views of any particular organization or individual within this Expert Group, or any specific PAI Partner.

Acknowledgements

The Partnership on AI is deeply grateful for the input of many colleagues and partners, especially Elonnai Hickok, Ann Skeet, Christina Colclough, Richard Zuroff, Jonnie Penn as well as the participants of the August 2019 workshop co-hosted by PAI and the Ford Foundation. We thank Arindrajit Basu, Pranav Bidaire, and Saumyaa Naidu for the research support.

Visa Laws, Policies, and Practices: Recommendations for Accelerating the Mobility of Global AI/ML Talent

PAI Staff


Executive Summary

Executive Summary

Immigration laws, policies, and practices are challenging the ability of many communities, including the artificial intelligence and machine learning (AI/ML) community, to incorporate diverse voices in their work. As a global, multi-stakeholder non profit committed to the creation and dissemination of best practices in artificial intelligence, the Partnership on AI (PAI) is uniquely positioned to address the impacts of immigration laws, policies, and practices on the AI/ML community.

PAI believes that bringing together experts from countries around the world that represent different cultures, socio-economic experiences, backgrounds, and perspectives is essential for AI/ML to flourish and help create the future we desire. In order to fulfill their talent goals and host conferences of international caliber, countries around the world will need to devise laws, policies, and practices that enable people around the world to contribute to these conversations.

Based on input from PAI Partners, and PAI’s own research, this paper offers recommendations to address these specific challenges. It highlights the importance of conferences and convenings for a variety of disciplines that are making important contributions to AI/ML, and makes recommendations for participants and organizers that may facilitate ease of travel for these events. It also presents recommendations for governments to improve the accessibility, evaluation and processing of visas for all types of potential visitors, including students, interns, and accompanying families. Appendices to the paper respond to potential questions, and provide an overview of the global demand for AI talent, as well as additional details on technical or expert visa, residence and work permit laws, policies and practices.

PAI’s recommendations are based on our area of expertise, and have been developed to help advance the mobility of innovative global AI/ML talent from a variety of disciplines. Many countries have already created visa classifications for other specialized occupations, including medical professionals, professional athletes, entertainers, religious workers, and entrepreneurs.

At the same time, we acknowledge the complex immigration debates taking place in countries around the world, and the challenges posed by global migration and the quest for basic human rights and dignity. These recommendations are in no way intended to minimize or replace opportunities for those affected by the ongoing immigration discussions and policymakers actions. We hope policymakers can create a path towards permanent residency or citizenship for these groups. In fact, while our recommendations target our field of expertise, we hope our paper can serve as a useful resource for the broader community, in support of balancing government public safety responsibilities with the benefits of immigration, freedom of movement, and collaboration.

Though this document incorporated suggestions from many of PAI’s partner organizations, it should not under any circumstances be read as representing the views of any specific member of the Partnership. Instead, it is an attempt to report the views of the artificial intelligence community as a whole.

Recommendations

Recommendations

Based on our investigations, PAI has developed the policy recommendations below for the global AI/ML community and policymakers around the world. Additional details on each of these recommendations are provided in the full text of the report.

I. Recommendations for the Global AI/ML Community:

  1. Use Plain Language Where Possible
    Consular and immigration officials may not be trained or familiar with the language used in the AI/ML community. PAI recommends that visa applicants explain technical terms using as much plain language as possible to describe the purpose of their visit and areas of expertise to facilitate the review of application documents and forms.
  2. Share Relevant Information with Host Countries in Advance
    Many governments evaluate visa applications on the basis of the applicant’s nationality and other factors, rather than the skills they will bring to the convening. Conference organizers will have to take extraordinary steps to facilitate the entry of their invited participants until laws, policies, and practices change in countries around the world. Conference organizers should contact host country government officials far in advance of the conference to share relevant information and facilitate government review of visa applications. Useful information includes a description of the conference, number of invited participants, and copies of invitation letter templates and other necessary paperwork.

II. Recommendations for Policymakers:

  1. Accelerate Reviews of Visa Applications
    Pass and implement laws, policies, and practices that accelerate review and favorably consider applications for visas, permits, and permanent legal status from highly skilled individuals. Visas should not be numerically limited or “capped.”
  2. Create AI/ML Visa Classifications within Existing Groups
    Members of existing intergovernmental groups, such as the Organization for Economic Cooperation and Development (OECD), should create visa classifications that enable AI/ML multidisciplinary experts to meet, convene, study, and work across member countries. The terms of the visa should be reciprocal across all countries.
  3. Publish Accessible Visa Application Information
    Visa application rules, processes & timelines should be clear, easily understood and accessible – published in plain language, in the applicants’ native languages on websites and in other publicly available locations. These processes should be fair, transparent, and clearly demonstrate that determinations for sponsor visas are based on skills.
  4. Establish Just Standards for Evaluating Visa Applications
    Eliminate nationality-based barriers in evaluating visa and permanent residence applications from highly skilled individuals. Security-based denials of applications should not be nationality based, but rather should be founded on specific and credible security and public safety threats, evidence of visa fraud, or indications of human trafficking.
  5. Train Officials in the Language of Emerging Technologies
    Train consular and immigration officials in the language of emerging technologies so they can quickly recognize and adjudicate applications from highly skilled experts.
  6. Assist Visa Applicants
    Empower select officials to assist applicants in correctly filling out visa paperwork, as well as clarifying and resolving any questions or discrepancies that may otherwise lead to a denial or delay in approval. Beneficiaries would include startups, small- and medium-sized enterprises, smaller colleges and universities, less affluent applicants, and students and interns.
  7. Students and Interns are the Future
    Pass laws that establish special categories of visas or permits for AI/ML students and interns. These laws should clearly identify a path for graduates to obtain a work permit (as necessary), or to obtain permanent legal status or citizenship.
  8. Redefine “Families”
    Adopt visa permissions that reflect a comprehensive definition of “family,” modeled on the Finnish Aliens Act and similar definitions in other European nations. Family visas should not be numerically limited. Legal spouses, partners, and those with family ties should also be permitted to work or study in the host country. Long-term caregivers should be permitted to accompany and remain with the main visa applicant and their family while employed in that capacity.
  9. Rely on Effective Policies and Systems to Protect Information
    Immigration restrictions do not adequately protect information and intellectual property rights. For example, trade negotiations can strengthen intellectual property laws and establish courts to protect and enforce intellectual property rights owned by individual rights holders, whereas implementing immigration policies and practices that broadly apply to all applicants from a particular country do not.

READ THE FULL PAPER

Frequently Asked Questions

Frequently Asked Questions

Why would PAI tackle a subject such as visas and immigration? This topic is not really related to artificial intelligence research.

PAI believes that bringing together experts from countries around the world that represent different cultures, socio-economic experiences, backgrounds, and perspectives is essential for AI/ML to flourish and help create the future we desire. Artificial intelligence is projected to affect all facets of society, and in some ways it already is having those effects. PAI’s work addresses a number of topics related to AI, such as criminal justice and labor and economy. Our work to address immigration challenges affecting the AI community is quite similar.

How does this document pertain to PAI’s mission and work?

This document makes visa policy recommendations that would improve the mobility of global AI/ML talent and enable companies, organizations and countries to benefit from their diverse perspectives. Fostering, cultivating, and preserving a culture of diversity and belonging in our work and in the people and organizations who contribute to our work is essential to our mission, and embedded in our Tenets. These include: committing to open research and dialogue on the ethical, social, economic, and legal implications of AI, ensuring that AI technologies benefit and empower as many people as possible, and striving to create a culture of cooperation, trust, and openness among AI scientists and engineers to help better achieve these goals.

Who benefits from this policy paper?

Unlike large, multinational companies and prominent, well-funded universities and colleges,  startups, small- and medium-sized enterprises, individuals traveling to conferences, less affluent applicants, students, and interns often lack the resources to hire experts to ensure their preferred candidates have the greatest chance to obtain visas for internships, to study, or to work in their organizations. These groups and individuals  often cannot successfully compete for visas, especially those that are numerically limited. They would be the greatest beneficiaries should governments implement these recommendations.

Why is PAI uniquely suited to address this issue?

As a multi-stakeholder non profit, PAI convenes over  100 global Partners, originating from 12 countries and four continents, and representing industry, civil society, and academic and research institutes. As such, we are uniquely qualified to describe the impacts of immigration laws, policies, and practices on the AI/ML community. The impetus for this document came from many of PAI’s Partners and colleagues, who have shared how certain visa laws, policies, and practices negatively affect their organizations’ abilities to benefit from global representatives and perspectives in their work.

Why is PAI focused on incorporating diverse voices in AI/ML?

Diverse perspectives are necessary to ensure that AI is developed in a responsible manner,  thoughtfully benefiting all people in society. Voices and contributions from global talent are also essential to reducing the unintended consequences that can arise from AI/ML development and deployment, including those related to safety and security. Due to the emergent and rapidly evolving nature of AI technology, AI in particular engenders high impact AI safety and security risks, which can be mitigated by increasing the diversity of participating voices Han, T. A., Pereira, L. M., Santos, F. C., & Lenaerts, T. (2019). Modelling the Safety and Surveillance of the AI Race. arXiv preprint. Diverse representation also serves to promote the safety of key members of the AI/ML community. Underrepresented voices, such as those of minorities and the LGBTQ community, are important as we design AI/ML systems to be inclusive of all populations.

Is PAI suggesting that AI/ML practitioners should be treated differently than other skilled workers? How is this different from other visa categories?

PAI’s recommendations would enable AI/ML practitioners, from a variety of disciplines, to travel and work more freely. In some cases, this could entail special visa classifications, similar to those that already exist for skilled workers in other specialized occupations, such as medical professionals, professional athletes, entertainers, religious workers, entrepreneurs, skilled laborers and trades workers.

This paper also highlights the many disciplines involved in the development and operations of AI/ML systems, above and beyond what is sometimes defined as “skilled technology work.” Responsible AI/ML systems involve input from researchers and practitioners in social sciences such as economics, sociology, philosophy, ethics, linguistics, and communications, and the “experiential expertise” offered by those working in labor and workers’ rights See discussion of “experiential expertise” in: Young, M., Magassa, L., & Friedman, B. (2019). Toward inclusive tech policy design: a method for underrepresented voices to strengthen tech policy documents. Ethics and Information Technology, 21(2), 89-103., in addition to technical fields such as mathematics, statistics, computer science, data science, neuroscience, and biology.

How does this work? Unlike medical professionals or engineers, AI/ML practitioners don’t have a certificate or license for governments to determine that they are experts.

Countries establish criteria for evaluating applications, whether for technical talent, a professional athlete, or someone skilled in trades or labor. Established eligibility criteria, and the process for evaluating this criteria, vary greatly from country to country. The PAI paper offers models for countries to consider and draw upon if they decide to create a classification for AI/ML practitioners.

For example, some countries require letters from a potential employer, or to have someone in the field attest to the applicant’s particular skills, or other supporting documentation that proves the applicant has the desired skills. Some examples:

  • An independent review board: The UK Tech Nation Visa, also known as the Tier 1 Exceptional Talent Visa, assigned an independent, “designated competent body,” to review and endorse applications. The Tech Nation Visa Guide outlines the skills and specialties typically exhibited in applications reviewed by this independent body, and the eligibility criteria.
  • Points-based systemCanada’s Express Entry Program, like other Canadian visas, evaluates applicants on the basis of the types of occupations and levels of skills they hope to attract. Certain occupations and skills, among other criteria, garner greater numbers of points. The higher the overall point total, the greater the likelihood of being admitted entry.
  • Government review: Japan’s Skilled Labor Visa program seeks documentation to support the visa application, and that documentation must prove, among other elements, that the applicant has a certain number of years of experience. The government will review the documentation, and issue a Certificate of Eligibility (COE) if they think the applicant possesses the necessary experience and skills. The existence of the COE in the application can accelerate the visa processing time.
  • Additional examples can be found in Recommendations for Policymakers #1 and Appendix C of the paper.

Visa Laws, Policies, and Practices: Recommendations for Accelerating the Mobility of Global AI/ML Talent

Executive Summary

Recommendations

Frequently Asked Questions

Recommendations

Frequently Asked Questions

Sources Cited

  1. Han, T. A., Pereira, L. M., Santos, F. C., & Lenaerts, T. (2019). Modelling the Safety and Surveillance of the AI Race. arXiv preprint.
  2. See discussion of “experiential expertise” in: Young, M., Magassa, L., & Friedman, B. (2019). Toward inclusive tech policy design: a method for underrepresented voices to strengthen tech policy documents. Ethics and Information Technology, 21(2), 89-103.
  3. Han, T. A., Pereira, L. M., Santos, F. C., & Lenaerts, T. (2019). Modelling the Safety and Surveillance of the AI Race. arXiv preprint.
  4. See discussion of “experiential expertise” in: Young, M., Magassa, L., & Friedman, B. (2019). Toward inclusive tech policy design: a method for underrepresented voices to strengthen tech policy documents. Ethics and Information Technology, 21(2), 89-103.
Table of Contents
1
2
3
4

Report on Algorithmic Risk Assessment Tools in the U.S. Criminal Justice System

PAI Staff

Overview

Overview 

This report was written by the staff of the Partnership on AI (PAI) and many of our Partner organizations, with particularly  input from the members of PAI’s Fairness, Transparency, and Accountability Working Group. Our work on this topic was initially prompted by California’s Senate Bill 10 (S.B. 10), which would mandate the purchase and use of statistical and machine learning risk assessment tools for pretrial detention decisions, but our work has subsequently expanded to assess the use of such software across the United States.

Though this document incorporated suggestions or direct authorship from around 30-40 of our partner organizations, it should not under any circumstances be read as representing the views of any specific member of the Partnership. Instead, it is an attempt to report the widely held views of the artificial intelligence research community as a whole.

The Partnership on AI is a 501(c)3 nonprofit organization established to study and formulate best practices on AI technologies, to advance the public’s understanding of AI, and to serve as an open platform for discussion and engagement about AI and its influences on people and society.

The Partnership’s activities are determined in collaboration with its coalition of over 80 members, including civil society groups, corporate developers and users of AI, and numerous academic artificial intelligence research labs. PAI aims to create a space for open conversation, the development of best practices, and coordination of technical research to ensure that AI is used for the benefit of humanity and society. Crucially, the Partnership is an independent organization; though supported and shaped by our Partner community, the Partnership is ultimately more than the sum of its parts and makes independent determinations to which its Partners collectively contribute, but never individually dictate. PAI provides administrative and project management support to Working Groups, oversees project selection, and provides financial resources or direct research support to projects as needs dictate.

The Partnership on AI is deeply grateful for the collaboration of so many colleagues in this endeavor and looks forward to further convening and undertaking the multi-stakeholder research needed to build best practices for the use of AI in this critical domain.

Report on Algorithmic Risk Assessment Tools in the U.S. Criminal Justice System

Overview

Executive Summary

Introduction

Minimum Requirements for the Responsible Deployment of Criminal Justice Risk Assessment Tools

Requirement 1: Training datasets must measure the intended variables

Requirement 2: Bias in statistical models must be measured and mitigated

Requirement 3: Tools must not conflate multiple distinct predictions

Requirement 4: Predictions and how they are made must be easily interpretable

Requirement 5: Tools should produce confidence estimates for their predictions

Requirement 6: Users of risk assessment tools must attend trainings on the nature and limitations of the tools

Requirement 7: Policymakers must ensure that public policy goals are appropriately reflected in these tools

Requirement 8: Tool designs, architectures, and training data must be open to research, review and criticismRequirement 8: Tool designs, architectures, and training data must be open to research, review and criticism

Requirement 9: Tools must support data retention and reproducibility to enable meaningful contestation and challenges

Requirement 10: Jurisdictions must take responsibility for the post-deployment evaluation, monitoring, and auditing of these tools

Conclusion

Sources Cited

  1. For example, many risk assessment tools assign individuals to decile ranks, converting their risk score into a rating from 1-10 which reflects whether they’re in the bottom 10% of risky individuals (1), the next highest 10% (2), and so on (3-10). Alternatively, risk categorization could be based on thresholds labeled as “low,” “medium,” or “high” risk.
  2. Whether this is the case depends on how one defines AI; it would be true under many but not all of the definitions surveyed for instance in Stuart Russell & Peter Norvig, Artificial Intelligence: A Modern Approach, Prentice Hall, 2010, at 2. PAI considers more expansive definitions, that include any automation of analysis and decision making by humans, to be most helpful.
  3. In California, the recently enacted California Bail Reform Act (S.B. 10) mandates the implementation of risk assessment tools while eliminating money bail in the state, though implementation of the law has been put on hold as a result of a 2020 ballot measure funded by the bail bonds industry to repeal it; see https://ballotpedia.org/California_Replace_Cash_Bail_with_Risk_Assessments_Referendum_(2020); Robert Salonga, Law ending cash bail in California halted after referendum qualifies for 2020 ballot, San Jose Mercury News (Jan. 17, 2019), https://www.mercurynews.com/2019/01/17/law-ending-cash-bail-in-california-halted-after-referendum-qualifies-for-2020-ballot/. In addition, a new federal law, the First Step Act of 2018 (S. 3649), requires the Attorney General to review existing risk assessment tools and develop recommendations for “evidence-based recidivism reduction programs” and to “develop and release” a new risk- and needs- assessment system by July 2019 for use in managing the federal prison population. The bill allows the Attorney General to use currently-existing risk and needs assessment tools, as appropriate, in the development of this system.
  4. In addition, many of our civil society partners have taken a clear public stance to this effect, and some go further in suggesting that only individual-level decision-making will be adequate for this application regardless of the robustness and validity of risk assessment instruments. See The Use of Pretrial ‘Risk Assessment’ Instruments: A Shared Statement of Civil Rights Concerns, http://civilrightsdocs.info/pdf/criminal-justice/Pretrial-Risk-Assessment-Full.pdf (shared statement of 115 civil rights and technology policy organizations, arguing that all pretrial detention should follow from evidentiary hearings rather than machine learning determinations, on both procedural and accuracy grounds); see also Comments of Upturn; The Leadership Conference on Civil and Human Rights; The Leadership Conference Education Fund; NYU Law’s Center on Race, Inequality, and the Law; The AI Now Institute; Color Of Change; and Media Mobilizing Project on Proposed California Rules of Court 4.10 and 4.40, https://www.upturn.org/static/files/2018-12-14_Final-Coalition-Comment-on-SB10-Proposed-Rules.pdf (“Finding that the defendant shares characteristics with a collectively higher risk group is the most specific observation that risk assessment instruments can make about any person. Such a finding does not answer, or even address, the question of whether detention is the only way to reasonably assure that person’s reappearance or the preservation of public safety. That question must be asked specifically about the individual whose liberty is at stake — and it must be answered in the affirmative in order for detention to be constitutionally justifiable.”) PAI notes that the requirement for an individualized hearing before detention implicitly includes a need for timeliness. Many jurisdictions across the US have detention limits at 24 or 48 hours without hearings. Aspects of this stance are shared by some risk assessment tool makers; see, Arnold Ventures’ Statement of Principles on Pretrial Justice and Use of Pretrial Risk Assessment, https://craftmediabucket.s3.amazonaws.com/uploads/AV-Statement-of-Principles-on-Pretrial-Justice.pdf.
  5. See Ecological Fallacy section and Baseline D for further discussion of this topic.
  6. Quantitatively, accuracy is usually defined as the fraction of correct answers the model produces among all the answers it gives. So a model that answers correctly in 4 out of 5 cases would have an accuracy of 80%. Interestingly, models which predict rare phenomena (like violent criminality) can be incredibly accurate without being useful for their prediction tasks. For example, if only 1% of individuals will commit a violent crime, a model that predicts that no one will commit a violent crime will have 99% accuracy even though it does not correctly identify any of the cases where someone actually commits a violent crime. For this reason and others, evaluation of machine learning models is a complicated and subtle topic which is the subject of active research. In particular, note that inaccuracy can and should be subdivided into errors of “Type I” (false positive) and “Type II” (false negative) – one of which may be more acceptable than the other, depending on the context.
  7. Calibration is a property of models such that among the group they predict a 50% risk for, 50% of cases recidivate. Note that this says nothing about the accuracy of the prediction, because a coin toss would be calibrated in that sense. All risk assessment tools should be calibrated, butthere are more specific desirable properties such as calibration within groups (discussed in Requirement 2 below) that not all tools will or should satisfy completely.
  8. Sarah L. Desmarais, Evan M. Lowder, Pretrial Risk Assessment Tools: A Primer for Judges, Prosecutors, and Defense Attorneys, MacArthur Safety and Justice Challenge (Feb 2019). The issue of cross-comparison applies not only to geography but to time. It may be valuable to use comparisons over time to assist in measuring the validity of tools, though such evaluations must be corrected for the fact that crime in the United States is presently a rapidly changing (and still on the whole rapidly declining) phenomenon.
  9. As a technical matter, a model can be biased for subpopulations while being unbiased on average for the population as a whole.
  10. Note here that the phenomenon of societal bias—the existence of beliefs, expectations, institutions, or even self-propagating patterns of behavior that lead to unjust outcomes for some groups—is not always the same as, or reflected in statistical bias, and vice versa. One can instead think of these as an overlapping Venn diagram with a large intersection. Most of the concerns about risk assessment tools are about biases that are simultaneously statistical and societal, though there are some that are about purely societal bias. For instance, if non-uniform access to transportation (which is a societal bias) causes higher rates of failure to appear for court dates in some communities, the problem is a societal bias, but not a statistical one. The inclusion of demographic parity measurements as part of model bias measurement (see Requirement 2) may be a way to measure this, though really the best solutions involve distinct policy responses (for instance, providing transportation assistance for court dates or finding ways to improve transit to underserved communities).
  11. For instance, Eckhouse et al. propose a 3-level taxonomy of biases. Laurel Eckhouse, Kristian Lum, Cynthia Conti-Cook, and Julie Ciccolini, Layers of Bias: A Unified Approach for Understanding Problems with Risk Assessment, Criminal Justice and Behavior, (Nov 2018).
  12. Some of the experts within the Partnership oppose the use of risk assessment tools specifically because of their pessimism that sufficient data exists or could practically be collected to meet purposes (a) and (b).
  13. Moreover, defining recidivism is difficult in the pretrial context. Usually, recidivism variables are defined using a set time period, e.g., whether someone is arrested within 1 year of their initial arrest or whether someone is arrested within 3 years of their release from prison. In the pretrial context, recidivism is defined as whether the individual is arrested during the time after their arrest (or pretrial detention) and before the individual’s trial. That period of time, however, can vary significantly from case to case, so it is necessary to ensure that each risk assessment tool predicts an appropriately defined measure of recidivism or public safety risk.
  14. See, e.g., Report: The War on Marijuana in Black and White, ACLU (2013), https://www.aclu.org/report/report-war-marijuana-black-and-white; ACLU submission to Inter-American Commission on Human Rights, Hearing on Reports of Racism in the Justice System of the United States, https://www.aclu.org/sites/default/files/assets/141027_iachr_racial_disparities_aclu_submission_0.pdf, (Oct 2017); Samuel Gross, Maurice Possley, Klara Stephens, Race and Wrongful Convictions in the United States, National Registry of Exonerations, https://www.law.umich.edu/special/exoneration/Documents/Race_and_Wrongful_Convictions.pdf; but see Jennifer L. Skeem and Christopher Lowenkamp, Risk, Race & Recidivism: Predictive Bias and Disparate Impact, Criminology 54 (2016), 690, https://risk-resilience.berkeley.edu/sites/default/files/journal-articles/files/criminology_proofs_archive.pdf (For some categories of crime in some jurisdictions, victimization and self-reporting surveys imply crime rates are comparable to arrest rates across demographic groups; an explicit and transparent reweighting process is procedurally appropriate even in cases where the correction it results in is small).
  15. See David Robinson and John Logan Koepke, Stuck in a Pattern: Early evidence on ‘predictive policing’ and civil rights, (Aug. 2016). https://www.upturn.org/reports/2016/stuck-in-a-pattern/ (“Criminologists have long emphasized that crime reports, and other statistics gathered by the police, are not an accurate record of the crime that happens in a community. In short, the numbers are greatly influenced by what crimes citizens choose to report, the places police are sent on patrol, and how police decide to respond to the situations they encounter. The National Crime Victimization Survey (conducted by the Department of Justice) found that from 2006-2010, 52 percent of violent crime victimizations went unreported to police and 60 percent of household property crime victimizations went unreported. Historically, the National Crime Victimization Survey ‘has shown that police are not notified of about half of all rapes, robberies and aggravated assaults.’”) See also Kristian Lum and William Isaac, To predict and serve? (2016): 14-19.
  16. Carl B. Klockars, Some Really Cheap Ways of Measuring What Really Matters, in Measuring What Matters: Proceedings From the Policing Research Meetings, 195, 195-201 (1999), https://www.ncjrs.gov/pdffiles1/nij/170610.pdf. [https://perma.cc/BRP3-6Z79] (“If I had to select a single type of crime for which its true level—the level at which it is reported—and the police statistics that record it were virtually identical, it would be bank robbery. Those figures are likely to be identical because banks are geared in all sorts of ways…to aid in the reporting and recording of robberies and the identification of robbers. And, because mostly everyone takes bank robbery seriously, both Federal and local police are highly motivated to record such events.”)
  17. ACLU, The War on Marijuana in Black and White: Billions of Dollars Wasted on Racially Biased Arrests, (2013), available at https://www.aclu.org/files/assets/aclu-thewaronmarijuana-rel2.pdf.
  18. Lisa Stoltenberg & Stewart J. D’Alessio, Sex Differences in the Likelihood of Arrest, J. Crim. Justice 32 (5), 2004, 443-454; Lisa Stoltenberg, David Eitle & Stewart J. D’Alessio, Race and the Probability of Arrest, Social Forces 81(4) 2003 1381-1387; Tia Stevens & Merry Morash, Racial/Ethnic Disparities in Boys’ Probability of Arrest and Court Actions in 1980 and 2000: The Disproportionate Impact of ‘‘Getting Tough’’ on Crime, Youth and Juvenile Justice 13(1), (2014).
  19. Delbert S. Elliott, Lies, Damn Lies, and Arrest Statistics, (1995), http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.182.9427&rep=rep1&type=pdf, 11.
  20. Simply reminding people to appear improves appearance rates. Pretrial Justice Center for Courts, Use of Court Date Reminder Notices to Improve Court Appearance Rates, (Sept. 2017).
  21. There are a number of obstacles that risk assessment toolmakers have identified towards better predictions on this front. Firstly, there is a lack of consistent data and definitions to help disentangle willful flight from justice from failures to appear for reasons that are either unintentional or not indicative of public safety risk. Policymakers may need to take the lead in defining and collecting data on these reasons, as well as identifying interventions besides incarceration that may be most appropriate for responding to them.
  22. This is known in the algorithmic fairness literature as “fairness through unawareness”; see Moritz Hardt, Eric Price, & Nathan Srebro, Equality of Opportunity in Supervised Learning, Proc. NeurIPS 2016, https://arxiv.org/pdf/1610.02413.pdf, first publishing the term and citing earlier literature for proofs of its ineffectiveness, particularly Pedreshi, Ruggieri, & Turini, Discrimination-aware data mining, Knowledge Discovery & Data Mining, Proc. SIGKDD (2008), http://eprints.adm.unipi.it/2192/1/TR-07-19.pdf.gz. In other fields, blindness is the more common term for the idea of achieving fairness by ignoring protected class variables (e.g., “race-blind admissions” or “gender-blind hiring”).
  23. Another way of conceiving omitted variable bias is as follows: data-related biases as discussed in Requirement 1 are problems with the rows in a database or spreadsheet: the rows may contain asymmetrical errors, or not be a representative sample of events as they occur in the world. Omitted variable bias, in contrast, is a problem with not having enough or the right columns in a dataset.
  24. These specific examples are from the Equivant/Northpoint COMPAS risk assessment; see sample questionnaire at https://assets.documentcloud.org/documents/2702103/Sample-Risk-Assessment-COMPAS-CORE.pdf
  25. This list is by no means exhaustive. Another approach involves attempting to de-bias datasets by removing all information regarding the protected class variables. See, e.g., James E. Johndrow & Kristian Lum, An algorithm for removing sensitive information: application to race-independent recidivism prediction, (Mar. 15, 2017), https://arxiv.org/pdf/1703.04957.pdf. Not only would the protected class variable itself be removed but also variation in other variables that is correlated with the protected class variable. This would yield predictions that are independent of the protected class variables, but could have negative implications for accuracy. This method formalizes the notion of fairness known as “demographic parity,” and has the advantage of minimizing disparate impact, such that outcomes should be proportional across demographics. Similar to affirmative action, however, this approach would raise additional fairness questions given different baselines across demographics.
  26. See Moritz Hardt, Eric Price, & Nathan Srebro, Equality of Opportunity in Supervised Learning, Proc. NeurIPS 2016, https://arxiv.org/pdf/1610.02413.pdf.
  27. This is due to different baseline rates of recidivism for different demographic groups in U.S. criminal justice data. See J. Kleinberg, S. Mullainathan, M. Raghavan. Inherent Trade-Offs in the Fair Determination of Risk Scores. Proc. ITCS, (2017), https://arxiv.org/abs/1609.05807 and A. Chouldechova, Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Proc. FAT/ML 2016, https://arxiv.org/abs/1610.07524. Another caveat is that such a correction can reduce overall utility, as measured as a function of the number of individuals improperly detained or released. See, e.g., Sam Corbett-Davies et al., Algorithmic Decision-Making and the Cost of Fairness, (2017), https://arxiv.org/pdf/1701.08230.pdf.
  28. As long as the training data show higher arrest rates among minorities, statistically accurate scores must of mathematical necessity have a higher false positive rate for minorities. For a paper that outlines how equalizing FPRs (a measure of unfair treatment) requires creating some disparity in predictive accuracy across protected categories, see J. Kleinberg, S. Mullainathan, M. Raghavan. Inherent Trade-Offs in the Fair Determination of Risk Scores. Proc. ITCS, (2017), https://arxiv.org/abs/1609.05807; for arguments about the limitations of FPRs as a sole and sufficient metric, see e.g. Sam Corbett-Davies and Sharad Goel, The Measure and Mismeasure of Fairness: A Critical Review of Fair Machine Learning, working paper, https://arxiv.org/abs/1808.00023.
  29. Geoff Pleiss et al. On Fairness and Calibration (describing the challenges of using this approach when baselines are different), https://arxiv.org/pdf/1709.02012.pdf.
  30. The stance that unequal false positive rates represents material unfairness was popularized in a study by Julia Angwin et al. Machine Bias, ProPublica, https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing, (2016), and confirmed in further detail in e.g, Julia Dressel and Hany Farid, The accuracy, fairness and limits of predicting recidivism, Science Advances, 4(1), (2018), http://advances.sciencemag.org/content/advances/4/1/eaao5580.full.pdf. Whether or not FPRs are the right measure of fairness is disputed within the statistics literature.
  31. See, e.g., Alexandra Chouldechova, Fair prediction with disparate impact: A study of bias in recidivism prediction instruments, Big Data 5(2), https://www.liebertpub.com/doi/full/10.1089/big.2016.0047, (2017).
  32. See, e.g., Niki Kilbertus et al., Avoiding Discrimination Through Causal Reasoning, (2018), https://arxiv.org/pdf/1706.02744.pdf.
  33. Formally, the toolmaker must distinguish “resolved” and “unresolved” discrimination. Unresolved discrimination results from a direct causal path between the protected class and predictor that is not blocked by a “resolving variable.” A resolving variable is one that is influenced by the protected class variable in a manner that we accept as nondiscriminatory. For example, if women are more likely to apply for graduate school in the humanities and men are more likely to apply for graduate school in STEM fields, and if humanities departments have lower acceptance rates, then women might exhibit lower acceptance rates overall even if conditional on department they have higher acceptance rates. In this case, the department variable can be considered a resolving variable if our main concern is discriminatory admissions practices. See, e.g., Niki Kilbertus et al., Avoiding Discrimination Through Causal Reasoning, (2018), https://arxiv.org/pdf/1706.02744.pdf.
  34. In addition to the trade-offs highlighted in this section, it should be noted that these methods require a precise taxonomy of protected classes. Although it is common in the United States to use simple taxonomies defined by the Office of Management and Budget (OMB) and the US Census Bureau, such taxonomies cannot capture the complex reality of race and ethnicity. See Revisions to the Standards for the Classification of Federal Data on Race and Ethnicity, 62 Fed. Reg. 210 (Oct 1997), https://www.govinfo.gov/content/pkg/FR-1997-10-30/pdf/97-28653.pdf. Nonetheless, algorithms for bias correction have been proposed that detect groups of decision subjects with similar circumstances automatically. For an example of such an algorithm, see Tatsunori Hashimoto et al., Fairness Without Demographics in Repeated Loss Minimization, Proc. ICML 2018, http://proceedings.mlr.press/v80/hashimoto18a/hashimoto18a.pdf. Algorithms have also been developed to detect groups of people that are spatially or socially segregated. See, e.g., Sebastian Benthall & Bruce D. Haynes, Racial categories in machine learning, Proc. FAT* 2019, https://dl.acm.org/authorize.cfm?key=N675470. Further experimentation with these methods is warranted. For one evaluation, see Jon Kleinberg, An Impossibility Theorem for Clustering, Advances in Neural Information Processing Systems 15, NeurIPS 2002.
  35. The best way to do this deserves further research on human-computer interaction. For instance, if judges are shown multiple predictions labelled “zero disparate impact for those who will not reoffend”, “most accurate prediction,” “demographic parity,” etc, will they understand and respond appropriately? If not, decisions about what bias corrections to use might be better made at the level of policymakers or technical government experts evaluating these tools.
  36. Cost benefit models require explicit tradeoff choices to be made between different objectives including liberty, safety, and fair treatment of different categories of defendants. These choices should be explicit, and must be made transparently and accountably by policymakers. For a macroscopic example of such a calculation see David Roodman, The Impacts of Incarceration on Crime, Open Philanthropy Project report, September 2017, p p131, at https://www.openphilanthropy.org/files/Focus_Areas/Criminal_Justice_Reform/The_impacts_of_incarceration_on_crime_10.pdf.
  37. Sandra G. Mayson, Dangerous Defendants, 127 Yale L.J. 490, 509-510 (2018).
  38. Id., at 510. (“The two risks are different in kind, are best predicted by different variables, and are most effectively managed in different ways.”)
  39. For instance, needing childcare increases the risk of failure to appear (see Brian H. Bornsein, Alan J. Thomkins & Elizabeth N. Neely, Reducing Courts’ Failure to Appear Rate: A Procedural Justice Approach, U.S. DOJ report 234370, available at https://www.ncjrs.gov/pdffiles1/nij/grants/234370.pdf ) but is less likely to increase the risk of recidivism.
  40. For example, if the goal of a risk assessment tool is to advance the twin public policy goals of reducing incarceration and ensuring defendants appear for their court dates, then the tool should not conflate a defendant’s risk of knowingly fleeing justice with their risk of unintentionally failing to appear, since the latter can be mitigated by interventions besides incarceration (e.g. giving the defendant the opportunity to sign up for phone calls or SMS-based reminders about their court date, or ensuring the defendant has transportation to court on the day they are to appear).
  41. Notably, part of the holding in Loomis, mandated a disclosure in any Presentence Investigation Report that COMPAS risk assessment information “was not developed for use at sentencing, but was intended for use by the Department of Corrections in making determinations regarding treatment, supervision, and parole,” Wisconsin v. Loomis (881 N.W.2d 749).
  42. M.L. Cummings, Automation Bias in Intelligent Time Critical Decision Support Systems, http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.91.2634&rep=rep1&type=pdf.
  43. It is important to note, however, that there is also evidence of the opposite phenomenon, whereby users might simply ignore the risk assessment tools’ predictions. In Christin’s ethnography of risk assessment users, she notes that professionals often “buffer” their professional judgment from the influence of automated tools. She quotes a former prosecutor as saying of risk assessment, “When I was a prosecutor I didn’t put much stock in it, I’d prefer to look at actual behaviors. I just didn’t know how these tests were administered, in which circumstances, with what kind of data.” From Christin, A., 2017, Algorithms in practice: Comparing web journalism and criminal justice, Big Data & Society, 4(2).
  44. See Wisconsin v. Loomis (881 N.W.2d 749).
  45. “Specifically, any PSI containing a COMPAS risk assessment must inform the sentencing court about the following cautions regarding a COMPAS risk assessment’s accuracy: (1) the proprietary nature of COMPAS has been invoked to prevent disclosure of information relating to how factors are weighed or how risk scores are to be determined; (2) risk assessment compares defendants to a national sample, but no cross- validation study for a Wisconsin population has yet been completed; (3) some studies of COMPAS risk assessment scores have raised questions about whether they disproportionately classify minority offenders as having a higher risk of recidivism; and (4) risk assessment tools must be constantly monitored and re-normed for accuracy due to changing populations and subpopulations.” Wisconsin v. Loomis (881 N.W.2d 749).
  46. Computer interfaces, even for simple tasks, can be highly confusing to users. For example, one study found that users failed to notice anomalies on a screen designed to show them choices they had previously selected for confirmation over 50% of the time, even after carefully redesigning the confirmation screen to maximize the visibility of anomalies. See Campbell, B. A., & Byrne, M. D. (2009). Now do voters notice review screen anomalies? A look at voting system usability, Proceedings of the 2009 Electronic Voting Technology Workshop/Workshop on Trustworthy Elections (EVT/WOTE ’09).
  47. This point depends on the number of input variables used for prediction. With a model that has a large number of features (such as COMPAS), it might be appropriate to use a method like gradient-boosted decision trees or random forests, and then provide the interpretation using an approximation. See Zach Lipton, The Mythos of Model Interpretability, Proc. ICML 2016, available at https://arxiv.org/pdf/1606.03490.pdf, §4.1. For examples of methods for providing explanations of complex models, see, e.g., Gilles Louppe et al., Understanding the variable importances in forests of randomized trees, Proc. NIPS 2013, available at https://papers.nips.cc/paper/4928-understanding-variable-importances-in-forests-of-randomized-trees.pdf; Marco Ribeiro, LIME – Local Interpretable
  48. Laurel Eckhouse et al., Layers of Bias: A Unified Approach for Understanding Problems With Risk Assessment, 46(2) Criminal Justice and Behavior 185–209 (2018), https://doi.org/10.1177/0093854818811379
  49. See id.
  50. See id.
  51. The lowest risk category for the Colorado Pretrial Assessment Tool (CPAT) included scores 0-17, while the highest risk category included a much broader range of scores: 51-82. In addition, the highest risk category corresponded to a Public Safety Rate of 58% and a Court Appearance Rate of 51%. Pretrial Justice Institute, (2013). Colorado Pretrial Assessment Tool (CPAT): Administration, scoring, and reporting manual, Version 1. Pretrial Justice Institute. Retrieved from http://capscolorado.org/yahoo_site_admin/assets/docs/CPAT_Manual_v1_-_PJI_2013.279135658.pdf
  52. User and usability studies such as those from the human-computer interaction field can be employed to study the question of how much deference judges give to pretrial or pre-sentencing investigations. For example, a study could examine how error bands affect judges’ inclination to follow predictions or (when they have other instincts) overrule them.
  53. As noted in Requirement 4, these mappings of probabilities to scores or risk categories are not necessarily intuitive, i.e. they are often not linear or might differ for different groups.
  54. In a simple machine learning prediction model, the tool might simply produce an output like “35% chance of recidivism.” A bootstrapped tool uses many resampled versions of the training datasets to make different predictions, allowing an output like, “It is 80% likely that this individual’s chance of recidivating is in the 20% – 50% range.” Of course these error bars are still relative to the training data, including any sampling or omitted variable biases it may reflect.
  55. The specific definition of fairness would depend on the fairness correction used.
  56. Humans are not naturally good at understanding probabilities or confidence estimates, though some training materials and games exist that can teach these skills; see eg: https://acritch.com/credence-game/
  57. To inform this future research, DeMichele et al.’s study conducting interviews with judges using the PSA tool can provide useful context for how judges understand and interpret these tools. DeMichele, Matthew and Comfort, Megan and Misra, Shilpi and Barrick, Kelle and Baumgartner, Peter, The Intuitive-Override Model: Nudging Judges Toward Pretrial Risk Assessment Instruments, (April 25, 2018). Available at SSRN: https://ssrn.com/abstract=3168500 or http://dx.doi.org/10.2139/ssrn.3168500;
  58. See the University of Washington’s Tech Policy Lab’s Diverse Voices methodology for a structured approach to inclusive requirements gathering. Magassa, Lassana, Meg Young, and Batya Friedman, Diverse Voices, (2017), http://techpolicylab.org/diversevoicesguide/.
  59. Such disclosures support public trust by revealing the existence and scope of a system, and by enabling challenges to the system’s role in government. See Pasquale, Frank. The black box society: The secret algorithms that control money and information. Harvard University Press, (2015). Certain legal requirements on government use of computers demand such disclosures. At the federal level, the Privacy Act of 1974 requires agencies to publish notices of the existence of any “system of records” and provides individuals access to their records. Similar data protection rules exist in many states and in Europe under the General Data Protection Regulation (GDPR).
  60. Reisman, Dillon, Jason Schultz, Kate Crawford, Meredith Whittaker, Algorithmic Impact Assessments: A Practical Framework for Public Agency Accountability, AI Now Institute, (2018).
  61. See Cal. Crim. Code §§ 1320.24 (e) (7), 1320.25 (a), effective Oct 2020.
  62. First Step Act, H.R.5682 — 115th Congress (2017-2018).
  63. For further discussion on the social justice concerns related to using trade secret law to prevent the disclosure of the data and algorithms behind risk assessment tools, see Taylor R. Moore,Trade Secrets and Algorithms as Barriers to Social Justice, Center for Democracy and Technology (August 2017), https://cdt.org/files/2017/08/2017-07-31-Trade-Secret-Algorithms-as-Barriers-to-Social-Justice.pdf.
  64. Several countries already publish the details of their risk assessment models. See, e.g., Tollenaar, Nikolaj, et al. StatRec-Performance, validation and preservability of a static risk prediction instrument, Bulletin of Sociological Methodology/Bulletin de Méthodologie Sociologique 129.1 (2016): 25-44 (in relation to the Netherlands); A Compendium of Research and Analysis on the Offender Assessment System (OaSys) (Robin Moore ed., Ministry of Justice Analytical Series, 2015) (in relation to the United Kingdom). Recent legislation also attempts to mandate transparency safeguards, see Idaho Legislature, House Bill No.118 (2019).
  65. See, e.g., Jeff Larson et al. How We Analyzed the COMPAS Recidivism Algorithm, ProPublica (May 23, 2016), https://www.propublica.org/article/how-we-analyzed-the-compas-recidivism-algorithm. For a sample of the research that became possible as a result of ProPublica’s data, see https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=propublica+fairness+broward. Data provided by Kentucky’s Administrative Office of the Courts has also enabled scholar’s to examine the impact of the implementation of the PSA tool in that state. Stevenson, Megan, Assessing Risk Assessment in Action (June 14, 2018). Minn. L. Rev, 103, Forthcoming; available at https://ssrn.com/abstract=3016088
  66. For an example of how a data analysis competition dealt with privacy concerns when releasing a dataset with highly sensitive information about individuals, see Ian Lundberg et al., Privacy, ethics, and data access: A case study of the Fragile Families Challenge (Sept. 1, 2018), https://arxiv.org/pdf/1809.00103.pdf.
  67. See Arvind Narayanan et al., A Precautionary Approach to Big Data Privacy (Mar. 19, 2015), http://randomwalker.info/publications/precautionary.pdf.
  68. See id. at p. 20 and 21 (describing how some sensitive datasets are only shared after the recipient completes a data use course, provides information about the recipient, and physically signs a data use agreement).
  69. For a discussion of the due process concerns that arise when information is withheld in the context of automated decision-making, see Danielle Keats Citron, Technological Due Process, 85 Wash. U. L. Rev. 1249 (2007), https://ssrn.com/abstract=1012360. See also, Paul Schwartz, Data Processing and Government Administration: The Failure of the American Legal Response to the Computer, 43 Hastings L. J. 1321 (1992).
  70. Additionally, the ability to reconstitute decisions evidences procedural regularity in critical decision processes and allows individuals to trust the integrity of automated systems even when they remain partially non-disclosed. See Joshua A. Kroll et al., Accountable algorithms, 165 U. Pa. L. Rev. 633 (2016).
  71. The ability to contest scores is not only important for defendant’s rights to adversarially challenge adverse information, but also for the ability of judges and other professionals to engage with the validity of the risk assessment outputs and develop trust in the technology. See Daniel Kluttz et al., Contestability and Professionals: From Explanations to Engagement with Algorithmic Systems (January 2019), https://dx.doi.org/10.2139/ssrn.3311894
  72. “Criteria tinkering” occurs when court clerks manipulate input values to obtain the score they think is correct for a particular defendant. See Hannah-Moffat, Kelly, Paula Maurutto, and Sarah Turnbull, Negotiated risk: Actuarial illusions and discretion in probation, 24.3 Canada J. of L. & Society/La Revue Canadienne Droit et Société 391 (2009). See also Angele Christin, Comparing Web Journalism and Criminal Justice, 4.2 Big Data & Society 1.
  73. For further guidance on how such audits and evaluations might be structured, see, AI Now Institute, Algorithmic Impact Assessments: A Practical Framework for Public Agency Accountability, https://ainowinstitute.org/aiareport2018.pdf; Christian Sandvig et al., Auditing algorithms: Research methods for detecting discrimination on internet platform (2014).
  74. See John Logan Koepke and David G. Robinson, Danger Ahead: Risk Assessment and the Future of Bail Reform, 93 Wash. L. Rev. 1725 (2018).
  75. For a discussion Latanya Sweeney & Ji Su Yoo, De-anonymizing South Korean Resident Registration Numbers Shared in Prescription Data, Technology Science, (Sept. 29, 2015), https://techscience.org/a/2015092901. Techniques exist that can guarantee that re-identification is impossible. See the literature on methods for provable privacy, notably differential privacy. A good introduction is in Kobbi Nissim, Thomas Steinke, Alexandra Wood, Mark Bun, Marco Gaboardi, David R. O’Brien, and Salil Vadhan, Differential Privacy: A Primer for a Non-technical Audience, http://privacytools.seas.harvard.edu/files/privacytools/files/pedagogical-document-dp_0.pdf.
  76. Brandon Buskey and Andrea Woods, Making Sense of Pretrial Risk Assessments, National Association of Criminal Defense Lawyers, (June 2018), https://www.nacdl.org/PretrialRiskAssessment. Human Rights Watch proposes a clear alternative: “The best way to reduce pretrial incarceration is to respect the presumption of innocence and stop jailing people who have not been convicted of a crime absent concrete evidence that they pose a serious and specific threat to others if they are released. Human Rights Watch recommends having strict rules requiring police to issue citations with orders to appear in court to people accused of misdemeanor and low-level, non-violent felonies, instead of arresting and jailing them. For people accused of more serious crimes, Human Rights Watch recommends that the release, detain, or bail decision be made following an adversarial hearing, with right to counsel, rules of evidence, an opportunity for both sides to present mitigating and aggravating evidence, a requirement that the prosecutor show sufficient evidence that the accused actually committed the crime, and high standards for showing specific, known danger if the accused is released, as opposed to relying on a statistical likelihood.” Human Rights Watch, Q & A: Profile Based Risk Assessment for US Pretrial Incarceration, Release Decisions, (June 1, 2018), https://www.hrw.org/news/2018/06/01/q-profile-based-risk-assessment-us-pretrial-incarceration-release-decisions.
Table of Contents
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16