ABOUT ML Reference Document

Last Updated

1.1.1 About This Document and Version Numbering

About This Document and Version Numbering

Given the growing influence of AI transparency research at a global scale, leveraging PAI’s position as a multistakeholder organization to amalgamate key contributions across separate partner initiatives will serve to organize and advocate for the underlying themes of current ML system documentation proposals. Initiatives like ABOUT ML that offer guidance on ML documentation will give companies a head start on a path of overseeing, auditing, and monitoring ML technologies, contributing to the key business goal of earning and keeping trust from consumers and policymakers.

In order to make the ABOUT ML resource more digestible, ABOUT ML has added a section giving stakeholders guidance on how to use the document. Additionally, there is now a distinction between the ABOUT ML Reference Document and the forthcoming PLAYBOOK with downloadable templates, specifications, guidance, and recommendations. ABOUT ML still plans to seek public comment, input from the Steering Committee, and feedback from at least one Diverse Voices panel as the reference evolves.

The original Version 0 (v0) draft of the ABOUT ML Reference Document focused on extracting major themes from recent research on recommendations for transparency documentation. This new version of the Reference Document builds on major themes from recent research on documentation from v0 and adds initial feedback from public commenting, a working session at PAI’s All Partners Meeting on institutional challenges and enablers of success in implementing documentation, and the Diverse Voices process to create an initial resource. The primary target audience for the ABOUT ML Reference Document are individual champions at organizations who work in positions where they may be able to advocate for the adoption of documentation processes in their team and organization. Others may also find value in its contents.

Components of the forthcoming PLAYBOOK, and to a lesser extent future versions of this reference document, will merge existing practices with insights from research, formalize best practices through an investigation into attempts to implement recommendations, and set new industry norms for documentation in ML lifecycles. Future drafts may also include commentary on other enablers of ML transparency, such mechanisms to adjust team and institutional settings, model interpretability tools, test suites and modified evaluation procedures, and more detail on necessary feedback loops for transparency. See Section 1.1.3 ABOUT ML Project Process and Timeline Overview below for further detail.

For transparency, here are the Sections of ABOUT ML released as of mid-2021 and which feedback mechanisms each one has received thus far:

Section Latest version Public Comment Steering Committee Diverse Voices
1. Project Overview v1.0
2. Literature review v1.0
3. Preliminary Synthesized Documentation Suggestions v1.0
4. Challenges v1.0
5. Promising interventions to try v1.0
6. ML primer v1.0
7. Appendix v1.0


1.1.2 ABOUT ML Goals and Plan

ABOUT ML Goals and Plan

There are numerous goals and stakeholders for the ABOUT ML resources. In order to prioritize, the Steering Committee recommends that the ABOUT ML Reference Document is developed towards goals in the following order aimed at building momentum in the most practical sequence to achieve widespread adoption of a set of documentation questions that serve to enable both internal and external auditability.

ABOUT ML Reference Document

The ABOUT ML Reference Document will, going forward, continue to evolve. A few guides, specifications, and other useful artifacts contained within the Reference Document will also be accessible as standalone resources.

Further work to operationalize practices noted in the ABOUT ML Reference Document will be showcased on the ABOUT ML website in the form of a PLAYBOOK. This PLAYBOOK will serve as an evolving repository of resources for stakeholders within the ML documentation community to use.

Documentation is important to consider as both an institutional process and an artifact (this idea will be expanded upon later in this document) because many teams and individuals have to incorporate completing and updating such an artifact into their work in order for it to be useful with all of the necessary information. This means that ABOUT ML’s eventual end goal is not only to recommend what information should go into documentation for all ML systems but also recommend how organizations can reliably reshape their processes to enable the reliable completion and maintenance of documentation in an ongoing manner. Thus, when considering the subgoals of ABOUT ML and how to sequence them, it is important to think about which subgoals have outcomes that can enable subsequent subgoals.

Subgoal 1 is to create documentation for internal accountability because this motivates organizations to invest in and build the internal processes and infrastructure needed to implement and scale the creation of documentation artifacts. The call for internal accountability can come from top-down buy-in, public commitments to AI ethics principles, and a set of motivated individual champions who advocate for the need for robust internal oversight over ML systems that impact many communities and people. The role of the ABOUT ML Reference Document released in 2021 will be to empower individual champions at all levels and roles inside organizations that build ML systems who are interested in advocating for and implementing ML system documentation inside their organization. The internal changes needed to be able to create documentation for all ML systems will include building tooling that reduces friction for collecting and collating information from people at different parts of the ML system lifecycle (e.g., development, testing, deployment, and maintenance), change management to convince all of these people to incorporate these tools and steps into their workflow, alignment of executives to provide resources and mandates to complete all of this work, and coordination of all teams involved in this process, spanning product teams, legal, compliance, policy and more. All of this will be a significant investment of time and resources, but the end product will be infrastructure for internal oversight and review of ML systems which also yields artifacts that can disseminate documentation information of ML systems easily between internal teams.


We follow the lexicon of algorithms research by Kohli et. al. (2018) in defining accountability as “the answerability of actors for outcomes” and the tracing and verification of system action as well as those who take responsibility for those actions.

The infrastructure and documentation artifacts created by Subgoal 1 can then be modified to also enable Subgoal 2, external accountability. The main change would be to modify the sets of questions and information to be shared externally based on the constraints of what organizations are willing to share and what information external stakeholders need to consider the ML system sufficiently transparent. There should be a broad and public conversation between organizations that build ML systems and external stakeholders that should be consulted — including civil society organizations, policymakers, end users, and non-users impacted by ML systems — to determine what information would be necessary in documentation for external accountability. After some initial approximate agreement is reached, organizations can experiment with implementing and sharing this set of information and the external stakeholders can provide feedback on whether this information does, in practice, enable the level of accountability desired. Both sets of stakeholders can iterate together until they are satisfied with the results, with perhaps agreements to continue reevaluating at regular intervals. On the process side, because there will already be an internal process for review of the documentation, this can be extended to review what information will be released with external documentation.

With Subgoal 2 complete, there will be a broadly agreed upon set of questions that stakeholders consider sufficient for external accountability. At this stage, ABOUT ML resources will include templates for documentation for external and internal consumption as well as information on how to implement documentation as a process from pilot to scale. Subgoal 3 is to scale adoption of ABOUT ML recommendations across the AI industry. PAI can provide assistance here as there are many Partner companies that can lead general adoption. The end result would truly be new industry norms on documentation practices which could enable many of the responsible and ethical AI development goals that companies and consortiums have put forth in recent years.

None of these steps would be easy and each involves a lot of investment and coordination with numerous sets of stakeholders. However, these are worthy goals to work towards because the end result would yield a lot of benefits for the responsible development and deployment of ML systems.

1.1.3 ABOUT ML Project Process and Timeline Overview

ABOUT ML Project Process and Timeline Overview

PAI launched the ABOUT ML iterative multistakeholder process with the initial v0 draft in order to initiate a broader community discussion. The draft has been updated into the current release in an effort to move towards best practices by going through the following phases:


Phase 0

Set up and maintain infrastructure for public input, including:
Diverse Voices panel,
Steering Committee,
Public comment.

Phase 1

Understand the latest research

Phase 2

Understand current practice

Current Phase

Phase 3

Combine research theory and results of current practice into testable pilots

Phase 4

Run pilot tests with PAI Partners and organizations

Phase 5

Collect data from pilot tests for transparency practices

Phase 6

Iterate on pilots with the latest research and practice

Phase 7

When there is sufficient body of evidence for a certain practice, elevate it to a best practice

Phase 8

Promulgate effective practices to establish new industry norms for transparency


PAI recognizes that this effort can only succeed with input from as broad a set of stakeholders as possible, and will be seeking input not only from our many Partners, but also from stakeholders from academia, civil society organizations, companies designing and deploying ML technology, and the general public. We welcome your participation.

The process is modeled after iterative ongoing processes to design internet standards (such as W3CWorld Wide Web Consortium Process Document (W3C) process outlined here: https://www.w3.org/2019/Process-20190301/, IETFInternet Engineering Task Force (IETF) process outlined here: https://www.ietf.org/standards/process/, and WHATWGThe Web Hypertext Application Technology Working Group (WHATWG) process outlined here: https://whatwg.org/faq#process) and will include a revamped public forum for discussion and a place to submit any proposed changes. We will announce instructions for accessing this online community and will welcome you to join in the public discussion and to submit proposed changes as many times as desired.

Public comments were collected and batch evaluated by the ABOUT ML Steering Committee, which has included dozens of experts, researchers, and practitioners recruited from a diverse set of PAI Partner organizations. The Steering Committee guided the process of updating ABOUT ML drafts based on the public comments submitted and new developments in research and practice. The current Reference Document, which includes feedback from the Diverse Voices process, will be reviewed and voted on to approve new releases by “rough consensus”Oever, N., Moriarty, K. The Tao of IETF: A novice’s guide to the Internet Engineering Task Force. https://www.ietf.org/about/participate/tao/. which is commonly used by other multistakeholder working groups.

The Steering Committee reconvened on April 13th, 2021 to review and refine the ABOUT ML Reference Document in preparation for release.

To ensure that diverse perspectives — especially those from communities historically excluded from technology decision-making — contribute to any ABOUT ML recommendations, PAI engaged with the Tech Policy Lab at the University of Washington, a Partner organization, to conduct Diverse VoicesYoung, M., Magassa, L. and 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. panels. This method was designed to gather feedback from stakeholders whose perspectives might not otherwise be consulted and to ensure that those perspectives are reflected in the released text. Thus, for any ABOUT ML releases that go through the Diverse Voices process, the panel feedback will be the last edits incorporated before a new release. This also means that each round of Diverse Voices panels will cause public comment on the document to be closed for several months, although the public forum will remain open for discussion during that time. Public comment on the document itself will re-open with the new release of the draft. The first round of Diverse Voices panels for ABOUT ML was held in late 2019.

ABOUT ML Reference Document

Section 0: How to Use this Document

Recommended Reading Plan

Quick Guides

How We Define

Contact for Support

Section 1: Project Overview

1.1 Statement of Importance for ABOUT ML Project

1.1.0 Importance of Transparency: Why a Company Motivated by the Bottom Line Should Adopt ABOUT ML Recommendations

1.1.1 About This Document and Version Numbering

1.1.2 ABOUT ML Goals and Plan

1.1.3 ABOUT ML Project Process and Timeline Overview

1.1.4 Who Is This Project For? Audiences for the ABOUT ML Resources Stakeholders That Should Be Consulted While Putting Together ABOUT ML Resources Audiences for ABOUT ML Documentation Artifacts Whose Voices Are Currently Reflected in ABOUT ML? Origin Story

Section 2: Literature Review (Current Recommendations on Documentation for Transparency in the ML Lifecycle)

2.1 Demand for Transparency and AI Ethics in ML Systems 

2.2 Documentation to Operationalize AI Ethics Goals

2.2.1 Documentation as a Process in the ML Lifecycle

2.2.2 Key Process Considerations for Documentation

2.3 Research Themes on Documentation for Transparency 

2.3.1 System Design and Set Up

2.3.2 System Development

2.3.3 System Deployment

Section 3: Preliminary Synthesized Documentation Suggestions

3.4.1 Suggested Documentation Sections for Datasets Data Specification Motivation Data Curation Collection Processing Composition Types and Sources of Judgement Calls Data Integration Use Distribution Maintenance

3.4.2 Suggested Documentation Sections for Models Model Specifications Model Training Evaluation Model Integration Maintenance

Section 4: Current Challenges of Implementing Documentation

Section 5: Conclusions

Version 0

Version 1

Appendix A: Compiled List of Documentation Questions 

Fact Sheets (Arnold et al. 2018)

Data Sheets (Gebru et al. 2018)

Model Cards (Mitchell et al. 2018)

A “Nutrition Label” for Privacy (Kelley et al. 2009)

The Dataset Nutrition Label: A Framework To Drive Higher Data Quality Standards (Holland et al. 2019)

Data Statements for Natural Language Processing: Toward Mitigating System Bias and Enabling Better Science (Bender and Friedman 2018)

Appendix B: Diverse Voices Process and Artifacts

Procurement Recruitment Email

Procurement Confirmation Email 

Appendix C: Glossary

Sources Cited

  1. Holstein, K., Vaughan, J.W., Daumé, H., Dudík, M., & Wallach, H.M. (2018). Improving Fairness in Machine Learning Systems: What Do Industry Practitioners Need? CHI.
  2. Young, M., Magassa, L. and 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. World Wide Web Consortium Process Document (W3C) process outlined here: https://www.w3.org/2019/Process-20190301/
  4. Internet Engineering Task Force (IETF) process outlined here: https://www.ietf.org/standards/process/
  5. The Web Hypertext Application Technology Working Group (WHATWG) process outlined here: https://whatwg.org/faq#process
  6. Oever, N., Moriarty, K. The Tao of IETF: A novice's guide to the Internet Engineering Task Force. https://www.ietf.org/about/participate/tao/.
  7. Young, M., Magassa, L. and 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.
  8. Friedman, B, Kahn, Peter H., and Borning, A., (2008) Value sensitive design and information systems. In Kenneth Einar Himma and Herman T. Tavani (Eds.) The Handbook of Information and Computer Ethics., (pp. 70-100) John Wiley & Sons, Inc. http://jgustilo.pbworks.com/f/the-handbook-of-information-and-computer-ethics.pdf#page=104; Davis, J., and P. Nathan, L. (2015). Value sensitive design: applications, adaptations, and critiques. Handbook of Ethics, Values, and Technological Design: Sources, Theory, Values and Application Domains. (pp. 11-40) DOI: 10.1007/978-94-007-6970-0_3. https://www.researchgate.net/publication/283744306_Value_Sensitive_Design_Applications_Adaptations_and_Critiques; Borning, A. and Muller, M. (2012). Next steps for value sensitive design. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI '12). (pp 1125-1134) DOI: https://doi.org/10.1145/2207676.2208560 https://dl.acm.org/citation.cfm?id=2208560
  9. Pichai, S., (2018). AI at Google: our principles. The Keyword. https://www.blog.google/technology/ai/ai-principles/; IBM’s Principles for Trust and Transparency. IBM Policy. https://www.ibm.com/blogs/policy/trust-principles/; Microsoft AI principles. Microsoft. https://www.microsoft.com/en-us/ai/our-approach-to-ai; Ethically Aligned Design – Version II. IEEE. https://standards.ieee.org/content/dam/ieee-standards/standards/web/documents/other/ead_v2.pdf
  10. Zeng, Y., Lu, E., and Huangfu, C. (2018) Linking artificial intelligence principles. CoRR https://arxiv.org/abs/1812.04814.
  11. essica Fjeld, Hannah Hilligoss, Nele Achten, Maia Levy Daniel, Sally Kagay, and Joshua Feldman, (2018). Principled artificial intelligence - a map of ethical and rights based approaches, Berkman Center for Internet and Society, https://ai-hr.cyber.harvard.edu/primp-viz.html
  12. Jobin, A., Ienca, M., & Vayena, E. (2019). Artificial Intelligence: the global landscape of ethics guidelines. arXiv preprint arXiv:1906.11668. https://arxiv.org/pdf/1906.11668.pdf
  13. Jobin, A., Ienca, M., & Vayena, E. (2019). Artificial Intelligence: the global landscape of ethics guidelines. arXiv preprint arXiv:1906.11668. https://arxiv.org/pdf/1906.11668.pdf
  14. Ananny, M., and Kate Crawford (2018). Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability. New Media and Society 20 (3): 973-989.
  15. Whittlestone, J., Nyrup, R., Alexandrova, A., & Cave, S. (2019, January). The Role and Limits of Principles in AI Ethics: Towards a Focus on Tensions. In Proceedings of the AAAI/ACM Conference on AI Ethics and Society, Honolulu, HI, USA (pp. 27-28). http://www.aies-conference.com/wp-content/papers/main/AIES-19_paper_188.pdf; Mittelstadt, B. (2019). AI Ethics–Too Principled to Fail? https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3391293
  16. Greene, D., Hoffmann, A. L., & Stark, L. (2019, January). Better, nicer, clearer, fairer: A critical assessment of the movement for ethical artificial intelligence and machine learning. In Proceedings of the 52nd Hawaii International Conference on System Sciences. https://scholarspace.manoa.hawaii.edu/handle/10125/59651
  17. Raji, I. D., & Buolamwini, J. (2019). Actionable auditing: Investigating the impact of publicly naming biased performance results of commercial ai products. In AAAI/ACM Conf. on AI Ethics and Society (Vol. 1). https://www.media.mit.edu/publications/actionable-auditing-investigating-the-impact-of-publicly-naming-biased-performance-results-of-commercial-ai-products/
  18. Algorithmic Impact Assessment (2019) Government of Canada https://www.canada.ca/en/government/system/digital-government/modern-emerging-technologies/responsible-use-ai/algorithmic-impact-assessment.html
  19. Benjamin, M., Gagnon, P., Rostamzadeh, N., Pal, C., Bengio, Y., & Shee, A. (2019). Towards Standardization of Data Licenses: The Montreal Data License. arXiv preprint arXiv:1903.12262. https://arxiv.org/abs/1903.12262; Responsible AI Licenses v0.1. RAIL: Responsible AI Licenses. https://www.licenses.ai/ai-licenses
  20. See Citation 5
  21. Safe Face Pledge. https://www.safefacepledge.org/; Montreal Declaration on Responsible AI. Universite de Montreal. https://www.montrealdeclaration-responsibleai.com/; The Toronto Declaration: Protecting the right to equality and non-discrimination in machine learning systems. (2018). Amnesty International and Access Now. https://www.accessnow.org/cms/assets/uploads/2018/08/The-Toronto-Declaration_ENG_08-2018.pdf ; Dagsthul Declaration on the application of machine learning and artificial intelligence for social good. https://www.dagstuhl.de/fileadmin/redaktion/Programm/Seminar/19082/Declaration/Declaration.pdf
  22. Dobbe, R., Dean, S., Gilbert, T., & Kohli, N. (2018). A Broader View on Bias in Automated Decision-Making: Reflecting on Epistemology and Dynamics. https://arxiv.org/pdf/1807.00553.pdf
  23. Wagstaff, K. (2012). Machine learning that matters. https://arxiv.org/pdf/1206.4656.pdf ; Friedman, B., Kahn, P. H., Borning, A., & Huldtgren, A. (2013). Value sensitive design and information systems. In Early engagement and new technologies: Opening up the laboratory (pp. 55-95). Springer, Dordrecht. https://vsdesign.org/publications/pdf/non-scan-vsd-and-information-systems.pdf
  24. Dobbe, R., Dean, S., Gilbert, T., & Kohli, N. (2018). A Broader View on Bias in Automated Decision-Making: Reflecting on Epistemology and Dynamics. https://arxiv.org/pdf/1807.00553.pdf
  25. Safe Face Pledge. https://www.safefacepledge.org/
  26. Montreal Declaration on Responsible AI. Universite de Montreal. https://www.montrealdeclaration-responsibleai.com/
  27. Diverse Voices How To Guide. Tech Policy Lab, University of Washington. https://techpolicylab.uw.edu/project/diverse-voices/
  28. Bender, E. M., & Friedman, B. (2018). Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6, 587-604.
  29. Ethically Aligned Design – Version II. IEEE. https://standards.ieee.org/content/dam/ieee-standards/standards/web/documents/other/ead_v2.pdf
  30. Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., Daumeé III, H., & Crawford, K. (2018). Datasheets for datasets. https://arxiv.org/abs/1803.09010 https://arxiv.org/abs/1803.09010; Hazard Communication Standard: Safety Data Sheets. Occupational Safety and Health Administration, US Department of Labor. https://www.osha.gov/Publications/OSHA3514.html
  31. Holland, S., Hosny, A., Newman, S., Joseph, J., & Chmielinski, K. (2018). The dataset nutrition label: A framework to drive higher data quality standards. https://arxiv.org/abs/1805.03677; Kelley, P. G., Bresee, J., Cranor, L. F., & Reeder, R. W. (2009). A nutrition label for privacy. In Proceedings of the 5th Symposium on Usable Privacy and Security (p. 4). ACM. http://cups.cs.cmu.edu/soups/2009/proceedings/a4-kelley.pdf
  32. Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., ... & Gebru, T. (2019, January). Model cards for model reporting. In Proceedings of the Conference on Fairness, Accountability, and Transparency (pp. 220-229). ACM. https://arxiv.org/abs/1810.03993
  33. Hind, M., Mehta, S., Mojsilovic, A., Nair, R., Ramamurthy, K. N., Olteanu, A., & Varshney, K. R. (2018). Increasing Trust in AI Services through Supplier's Declarations of Conformity. https://arxiv.org/abs/1808.07261
  34. Veale M., Van Kleek M., & Binns R. (2018) ‘Fairness and Accountability Design Needs for Algorithmic Support in High-Stakes Public Sector Decision-Making’ in Proceedings of the ACM Conference on Human Factors in Computing Systems, CHI 2018. https://arxiv.org/abs/1802.01029.
  35. Benjamin, M., Gagnon, P., Rostamzadeh, N., Pal, C., Bengio, Y., & Shee, A. (2019). Towards Standardization of Data Licenses: The Montreal Data License. https://arxiv.org/abs/1903.12262
  36. Cooper, D. M. (2013, April). A Licensing Approach to Regulation of Open Robotics. In Paper for presentation for We Robot: Getting down to business conference, Stanford Law School.
  37. Responsible AI Practices. Google AI. https://ai.google/education/responsible-ai-practices
  38. Everyday Ethics for Artificial Intelligence. (2019). IBM. https://www.ibm.com/watson/assets/duo/pdf/everydayethics.pdf
  39. Federal Trade Commission. (2012). Best Practices for Common Uses of Facial Recognition Technologies (Staff Report). Federal Trade Commission, 30. https://www.ftc.gov/sites/default/files/documents/reports/facing-facts-best-practices-common-uses-facial-recognition-technologies/121022facialtechrpt.pdf
  40. Microsoft (2018). Responsible bots: 10 guidelines for developers of conversational AI. https://www.microsoft.com/en-us/research/uploads/prod/2018/11/Bot_Guidelines_Nov_2018.pdf
  41. Tramer, F., Atlidakis, V., Geambasu, R., Hsu, D., Hubaux, J. P., Humbert, M., ... & Lin, H. (2017, April). FairTest: Discovering unwarranted associations in data-driven applications. In 2017 IEEE European Symposium on Security and Privacy (EuroS&P) (pp. 401-416). IEEE. https://github.com/columbia/fairtest, https://www.mhumbert.com/publications/eurosp17.pdf
  42. Kishore Durg (2018). Testing AI: Teach and Test to raise responsible AI. Accenture Technology Blog. https://www.accenture.com/us-en/insights/technology/testing-AI
  43. Kush R. Varshney (2018). Introducing AI Fairness 360. IBM Research Blog. https://www.ibm.com/blogs/research/2018/09/ai-fairness-360/
  44. Dave Gershgorn (2018). Facebook says it has a tool to detect bias in its artificial intelligence. Quartz. https://qz.com/1268520/facebook-says-it-has-a-tool-to-detect-bias-in-its-artificial-intelligence/
  45. James Wexler. (2018) The What-If Tool: Code-Free Probing of Machine Learning Models. Google AI Blog. https://ai.googleblog.com/2018/09/the-what-if-tool-code-free-probing-of.html
  46. Miro Dudík, John Langford, Hanna Wallach, and Alekh Agarwal (2018). Machine Learning for fair decisions. Microsoft Research Blog. https://www.microsoft.com/en-us/research/blog/machine-learning-for-fair-decisions/
  47. Veale, M., Binns, R., & Edwards, L. (2018). Algorithms that Remember: Model Inversion Attacks and Data Protection Law. Phil. Trans. R. Soc. A, 376, 20180083. https://doi.org/10/gfc63m
  48. Floridi, L. (2010, February). Information: A Very Short Introduction.
  49. Data Information Specialists Committee UK, 2007. http://www.disc-uk.org/qanda.html.
  50. Harwell, Drew. “Federal Study Confirms Racial Bias of Many Facial-Recognition Systems, Casts Doubt on Their Expanding Use.” The Washington Post, WP Company, 21 Dec. 2019, www.washingtonpost.com/technology/2019/12/19/federal-study-confirms-racial-bias-many-facial-recognition-systems-casts-doubt-their-expanding-use/
  51. Hildebrandt, M. (2019) ‘Privacy as Protection of the Incomputable Self: From Agnostic to Agonistic Machine Learning’, Theoretical Inquiries in Law, 20(1) 83–121.
  52. D'Amour, A., Heller, K., Moldovan, D., Adlam, B., Alipanahi, B., Beutel, A., ... & Sculley, D. (2020). Underspecification presents challenges for credibility in modern machine learning. arXiv preprint arXiv:2011.03395.
  53. Selinger, E. (2019). ‘Why You Can’t Really Consent to Facebook’s Facial Recognition’, One Zero. https://onezero.medium.com/why-you-cant-really-consent-to-facebook-s-facial-recognition-6bb94ea1dc8f
  54. Lum, K., & Isaac, W. (2016). To predict and serve?. Significance, 13(5), 14-19. https://rss.onlinelibrary.wiley.com/doi/full/10.1111/j.1740-9713.2016.00960.x
  55. LabelInsight (2016). “Drive Long-Term Trust & Loyalty Through Transparency”. https://www.labelinsight.com/Transparency-ROI-Study
  56. Crawford and Paglen, https://www.excavating.ai/
  57. Geva, Mor & Goldberg, Yoav & Berant, Jonathan. (2019). Are We Modeling the Task or the Annotator? An Investigation of Annotator Bias in Natural Language Understanding Datasets. https://arxiv.org/pdf/1908.07898.pdf
  58. Bender, E. M., & Friedman, B. (2018). Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6, 587-604.
  59. Desmond U. Patton et al (2017).
  60. See Cynthia Dwork et al.,
  61. Katta Spiel, Oliver L. Haimson, and Danielle Lottridge. (2019). How to do better with gender on surveys: a guide for HCI researchers. Interactions. 26, 4 (June 2019), 62-65. DOI: https://doi.org/10.1145/3338283
  62. A. Doan, A. Y. Halevy, and Z. G. Ives. Principles of Data Integration. Morgan Kaufmann, 2012
  63. Momin M. Malik. (2019). Can algorithms themselves be biased? Medium. https://medium.com/berkman-klein-center/can-algorithms-themselves-be-biased-cffecbf2302c
  64. Fire, Michael, and Carlos Guestrin (2019). “Over-Optimization of Academic Publishing Metrics: Observing Goodhart’s Law in Action.” GigaScience 8 (giz053). https://doi.org/10.1093/gigascience/giz053.
  65. Vogelsang, A., & Borg, M. (2019, September). Requirements engineering for machine learning: Perspectives from data scientists. In 2019 IEEE 27th International Requirements Engineering Conference Workshops (REW) (pp. 245-251). IEEE
  66. Eckersley, P. (2018). Impossibility and Uncertainty Theorems in AI Value Alignment (or why your AGI should not have a utility function). arXiv preprint arXiv:1901.00064.
  67. Partnership on AI. Report on Algorithmic Risk Assessment Tools in the U.S. Criminal Justice System, Requirement 5.
  68. Eckersley, P. (2018). Impossibility and Uncertainty Theorems in AI Value Alignment (or why your AGI should not have a utility function). arXiv preprint arXiv:1901.00064.https://arxiv.org/abs/1901.00064
  69. If it is not, there is likely a bug in the code. Checking a predictive model's performance on the training set cannot distinguish irreducible error (which comes from intrinsic variance of the system) from error introduced by bias and variance in the estimator; this is universal, and has nothing to do with different settings or
  70. Selbst, Andrew D. and Boyd, Danah and Friedler, Sorelle and Venkatasubramanian, Suresh and Vertesi, Janet (2018). “Fairness and Abstraction in Sociotechnical Systems”, ACM Conference on Fairness, Accountability, and Transparency (FAT*). https://ssrn.com/abstract=3265913
  71. Tools that can be used to explore and audit the predictive model fairness include FairML, Lime, IBM AI Fairness 360, SHAP, Google What-If Tool, and many others
  72. Wagstaff, K. (2012). Machine learning that matters. arXiv preprint arXiv:1206.4656. https://arxiv.org/abs/1206.4656
Table of Contents