Can firms develop and implement “ethical” artificial intelligence (AI) by taking a new approach to the acquisition and use of data used to develop machine learning, neural networks, and other AI-based tools? That is the question that Element AI, a Canadian firm developing AI products, hopes to answer through a new initiative to fund and support research on the use and development of so-called data trusts.

Data Trusts: Conceptually Interesting – But As Yet Unproven

Data trusts are a novel construct to manage access, consent, and use rights of data sets. As proposed by Element, data trusts reflect a move away from existing individual consent-based models of data governance towards a framework based on common law trust principles. As proposed, a data trust would direct and manage access, consent and use rights to data sets through a trustee empowered with certain authority and fiduciary obligations to persons whose data resides in the data trust.

This approach, as documented in an Element AI whitepaper, is intended to provide individuals greater control over their personal information, enhance privacy, and give the public the opportunity to “share in the value of data and artificial intelligence.” Element’s whitepaper builds on a recent workshop in which representatives from numerous fields, including data governance, machine learning, privacy, public policy, and property law, provided perspectives on the utility of data trusts to enable the development of ethical AI.

But despite its lofty ambitions, the proposal is only conceptual at this time. Indeed, whether this structure could actually yield the outcomes that Element contemplates is far from clear.

Ethical AI: An Emerging Foundational Concept Necessary to Achieve Trustworthy AI

Element is not the first organization to explore frameworks needed to establish ethical AI – far from it. In fact, the issue of whether and how to develop ethical principles governing the development, use and sale of AI is a foundational question that many public and private sector organizations are considering.

Indeed, over the past several years, dozens of proposals have emerged to implement ethical principles governing the development, use, and sale of AI. Notably, the U.S. Department of Defense, the EU, the OECD, the United Kingdom House of Lords, and governmental actors in Germany, France, Australia, Canada, Singapore, and Dubai have all released AI ethics proposals or frameworks.

At the same time, leading private sector AI developers, including Microsoft, IBM, Google, OpenAI, the Partnership on AI, and others, have proposed their own frameworks for ethical AI, as have other non-governmental actors including IEEE, the Future of Life Institute’s Asilomar Principles, Amnesty International to name a few.

While each of these proposals offer different perspectives, each begins from the view that society will benefit from the adoption and implementation of an ethical framework governing the development, use, and sale of AI. Notably, these proposals share the following common principles:

  1. Transparency: Developers and users of AI systems must provide explanations of how machine learning processes produce their results, including the disclosure of, where available and appropriate, the “key factors” that contribute to the outcome of an AI-based action.
  2. Justice and fairness: This principle requires organizations take affirmative steps to determine whether their systems have discriminatory effects and to take appropriate steps to mitigate unintended bias and discrimination.
  3. Responsibility: This principle holds that human beings should exercise appropriate levels of judgment and remain responsible for the development, use, and outcomes of AI systems.
  4. Privacy: Reflected in various ways across different frameworks, this principle holds that AI systems must comply with privacy laws that require transparency about the collection, use, and storage of data, and mandates that consumers have appropriate controls so that they can choose how their data is used.
  5. Non-maleficence: This principle imposes the well-accepted value of “do no harm” used in other contexts. In practice, AI developers and users would be required to implement safeguards within algorithmic platforms to help users protect personal data and mitigate harms or potential misuse of the AI system.

Whether Element’s proposed use of data trusts will bring us closer to reaching these core values (and others articulated under various ethical AI frameworks) is an open question. Building from the premise that current consent-based data collection models have significant flaws, the Element proposal rests upon the notion that a data trust construct will permit individuals or possibly a trustee acting on behalf of such persons to maintain direct control over data in a manner not available under current models.

While that proposition may be laudable, it does not answer the question of whether (or how) such a model will lead to the development of ethical AI. Further, several elements of the proposal could actually hinder the development of a robust, well-trained, transparent and ethical AI system.

For example, if implemented as proposed, data trusts could present operational challenges to AI developers, who require increasingly greater amounts of data to train their technology. To the extent this proposal would facilitate the removal of data from certain sets, such steps could undermine robust AI training procedures which are more likely to reduce potential risks of bias when the underlying AI is trained on larger, more diverse data sets. Further, selectively eliminating data from AI systems could potentially reduce transparency by limiting available data the system can rely upon to identify as the basis for certain outcomes.

While data trusts may be able to provide individuals with a greater measure of control over their personal data, the whitepaper offers no clear explanation of how that will help foster the development of ethical AI. Although we expect continued movement to establish formal governance frameworks around ethical AI in the months ahead, it is not at all clear that data trusts (at least as proposed here) will play a meaningful role in any emergent ethical AI frameworks.