UNESCO’s Draft Recommendation on ethics and AI: our thoughts on the final text

The United Nations Educational, Scientific and Cultural Organization (UNESCO) has now finalised its Recommendation on the Ethics of Artificial Intelligence, which it will submit for adoption to the General Conference of UNESCO later this year. If adopted, this Recommendation would mark the successful conclusion of a multi-year process to develop an international standard-setting instrument on the ethics of AI. 

The Recommendation would not be binding on UNESCO member states (indeed, many states made clear during the negotiations that it is an entirely voluntary instrument), but it would still have significant normative value as the first authoritative AI-specific instrument to be negotiated at the global level. It could potentially be used to inform ongoing global and regional efforts to regulate AI (specifically at the European Union and Council of Europe) or influence national approaches to AI governance by UNESCO member states or other actors such as private companies.

GPD has closely monitored and actively engaged in this process. In summer 2020, we responded to an open consultation on an initial draft of the Recommendation, and earlier this year we laid out our thoughts on an updated version of the draft instrument. In both instances, we highlighted our concerns and a number of ways in which the text could be improved. But does the final text—which has once again been modified during two intersessional meetings held in April and June 2021—better reflect a rights-respecting approach to AI?

POSITIVE CHANGES TO THE FINAL TEXT

• Human rights is more mainstreamed in the final text. Previous versions of the Recommendation failed to clarify the relationship between “ethics” and “human rights” when it comes to AI, at one point suggesting that the role of “ethics” was to fill in gaps in the human rights framework. This final version, however, is more clearly grounded in the existing international human rights framework. It includes additional references to international law and the obligations of states under international human rights law to respect, protect and promote human rights in the context of AI. References to ethics, including ethical values and principles, are still found throughout the text, but they are, for the most part, provided as a supplement and not as a replacement to a human rights approach to AI. 

• Additional detail on specific human rights concerns and high-risk AI systems. We are pleased that the Recommendation now makes specific reference to high-risk AI systems, such as those used for social scoring or mass surveillance purposes, and calls for their prohibition. This additional language is welcome as it reinforces the need to more closely consider these AI systems and the risks they pose to human rights. It also aligns this Recommendation with other ongoing efforts, including at the EU level, which seek to ban AI systems that pose an unacceptable risk to human rights, and thus provides guidance for others developing AI governance frameworks.

CONTINUING CONCERNS

Despite a number of positive changes, there are certain elements of the final Recommendation which continue to raise concerns from a human rights perspective.

• The inclusion of ambiguous and undefined terminology. While the text of the final Recommendation is, in part, more firmly rooted in the international human rights framework, the Recommendation still contains a number of undefined terms including “ethical values”, “quality of life” and “just” societies. The Recommendation does specify that while the definition of “quality of life” should be left open to individuals or groups, it should not result in violation or abuse of human rights and fundamental freedoms, or the dignity of humans. But there are other undefined terms in the Recommendation that are not accompanied by such language or references to human rights. This could result in member states interpreting these terms in a subjective manner, including in ways which create tensions with their obligations under international human rights law. 
• Insufficient language on human oversight and determinations. GPD is pleased that the final Recommendation takes a less restrictive approach to human determinations and oversight of AI, providing that, “As a rule, life and death decisions should not be ceded to AI systems”. But this language isn’t strong enough: it should make clear that human determinations and oversight are needed in any circumstance where AI decision-making may potentially produce legal effects or have serious human rights implications. 

LOOKING TO THE FUTURE – ETHICAL IMPACT ASSESSMENTS

One of the most significant outcomes of the final Recommendation is that UNESCO would be tasked with developing a methodology for Ethical Impact Assessment (EIA) of AI technologies. This process aligns with another part of the Recommendation, which provides that “member states should introduce frameworks for impact assessments, such as ethical impact assessment, to identify and assess the benefits, concerns and risks of AI systems, as well as appropriate risk prevention, mitigation and monitoring measures”.

The creation of this EIA could be a positive development as it will be “based on rigorous scientific research and grounded in international human rights law”. Potentially, it could lead to a standardised impact assessment for states and private companies which adequately considers human rights in the context of AI. However, there is also a risk that this EIA may distract from the development and deployment of purely human rights-based impact assessments.

NEXT STEPS

The Recommendation on the Ethics of Artificial Intelligence will be submitted for adoption to the General Conference of UNESCO in November 2021. We’ll be following developments at UNESCO closely—take a look at our AI Policy Hub and Forums Guide, and sign up to our monthly Digest for regular updates and analysis.