Table 3.
Companies practicing and developing AI in order to apply/sell/advise about AI that explicitly have committed to principles or guidelines for AI and/or to the tenets for AI of the Partnership on AI—tabulated according to their internal governance for responsible AI; training and educational materials about responsible AI; new tools for fair/non-biased AI and explainable AI and secure/privacy-proof AI and accountability of AI; and their external collaboration and funding for research into responsible AI
| AI companies committed to AI principles1 |
Governance for responsible AI inside the firm2 |
Training and educational materials produced about responsible AI | New tools for fair/non-biased AI | New tools for explainable AI | New tools for secure/privacy-proof AI3 | New tools for accountability of AI | External collaboration and funding for responsible AI research |
|---|---|---|---|---|---|---|---|
| Amazon | SHAP values and feature importance tools (proprietary) | Co-funding of NSF project ‘Fairness in AI’ | |||||
| Advanced Technology External Advisory Council (now defunct); Ethics & Society team; responsible innovation teams review new projects for conformity to AI principles | Employee training about ethical AI, educational materials (see ‘People + AI Guidebook’) | Facets, What-If tool, Fairness Indicators (all open source) | What-If tool (open source) | CleverHans (open source); Private Aggregation of Teacher Ensembles (open source), Tensor Flow Privacy (open source); Federated Learning, RAPPOR, Cobalt (open source) | Model cards | ||
| Microsoft | AI and Ethics in Engineering Research Committee, Office of Responsible AI | Internal guidelines and checklists (e.g., ‘In Pursuit of Inclusive AI’, ‘Inclusive Design’) |
FairLearn (open source) |
InterpretML (open source) |
WhiteNoise package (open source) |
Data sheets for datasets | |
| IBM | AI Ethics Board (chaired by AI Ethics Global Leader and Chief Privacy Officer) | Guidelines for AI developers (‘Everyday Ethics for AI’) |
AI Fairness 360 Toolkit (open source) |
AI Explainability 360 Toolkit (open source) | Adversarial Robustness 360 Toolbox (open source) | Fact sheets | Joint research with Institute for Human-Centred AI (Stanford University), funding of Tech Ethics Lab (University of Notre Dame) |
| Intel | AI Ethics and Human Rights Team | ||||||
| AI Ethics Team | Fairness Flow (proprietary) | Captum (for deep neural networks) (open source) | Funding of Institute of Ethics in AI (TU Munich) | ||||
| Telefónica | AI Ethics course and AI Ethics self-assessment (for employees) | ||||||
| Accenture | Advocates ‘responsible, explainable, citizen AI’ to clients, educational materials | AI Fairness Tool, part of AI Launchpad (proprietary) | |||||
| SAP | AI Ethics Advisory Panel, AI Ethics Steering Committee; diverse and interdisciplinary teams | Course about ‘trustworthy AI’ (for employees and other stakeholders) | |||||
| Philips | |||||||
| Salesforce | Ethical Use Advisory Council, Office of Ethical and Humane Use of Technology, data science review board; inclusive teams | Teaching module about bias in AI (for employees and clients) | Einstein discovery tools (proprietary) | Einstein discovery tools (proprietary) | |||
| McKinsey (Quantum Black) | Advocates ‘responsible AI’ approach to clients, educational materials |
CausalNex (open source) |
|||||
| Sage | Team diversity | ||||||
| Tieto | In-company ethics certification, special AI ethics engineers, and trainers appointed | ||||||
| Health Catalyst | |||||||
| Deep Mind (Google subsidiary) | External ‘fellows’, Ethics Board, Ethics and Society Team | ||||||
| Element AI | Team diversity | Internal blogposts about responsible AI | Fairness tools (proprietary) | Explainability tools (proprietary) |
1Companies are ordered by revenue. Apple, Samsung, Deutsche Telekom, Sony, Kakao, Unity Technologies, and Affectiva have been omitted from the table since my searches yielded no results for them.
2 ‘Ethics Team’ denotes what is variously referred to as ethical/ethics/review committee/board (highest corporate body concerning affairs of ethical AI).
3Techniques such as privacy by design (de-identification of data) and security by design (encryption) are not mentioned in the table since they are well-known and do not refer specifically to problems of ML.
Note: The sources that I drew my information in the table from are for the most part given in footnotes in the text of the article; otherwise, the sources are available on request.