Conclusion

Conclusion

One of the primary motivations for undertaking this collaboration and sharing these findings was to lower the barriers for companies to adopt responsible sourcing practices outlined in PAI’s responsible sourcing white paper. Though there is more work to be done, we believe that putting the recommendations to the test was an important first step towards developing a deeper understanding of how companies can incorporate ethical sourcing practices.

Prior to the changes introduced during this collaboration, research teams at DeepMind would need to think through project design individually and seek out information on their own from the various sources. Introducing a centralized set of guidelines and a review process saves them time and allows them to benefit from the shared learnings of their colleagues. By asking teams to submit an application that documents their approach to setting up data enrichment projects based on a centralized guidance document, the Human Data Review Group and review process served as a starting point to centralize the gathering of best practices and learnings from across the organization so teams could learn from each other.

We are sharing this case study and accompanying resources in the hopes that this will serve as a guide for other AI practitioners to adopt similar types of guidelines.

We are sharing this case study and accompanying resources in the hopes that this will serve as a guide for other AI practitioners to adopt similar types of guidelines and so that we can push the industry towards better practices. While PAI is not positioned to audit AI practitioners, we hope that sharing resources and our documentation of how these resources were developed will provide practitioners with confidence in making similar changes to their data enrichment practices. Recognizing that different organizations may have different resources and constraints, the resources developed over the course of this collaboration are meant to make it easier for organizations with less infrastructure to incorporate these guidelines into their own workflows without having to replicate the rigorous process we have undertaken here.

In addition to helping us identify what AI companies are positioned to do to positively impact worker experience, this has also helped us understand the limits of what individual companies can do to impact worker experience and what action is needed from platforms/vendors and policymakers. These insights will help shape our future work in this area as we continue to push for more ethical data supply chains.

We hope that this level of transparency creates an opportunity to discuss additional avenues to improve conditions for data enrichment workers and helps recenter considerations of labor at the heart of the industry’s data enrichment decisions.