Thread by @annajobin

🚨 NOW OUT 🚨
"The global landscape of AI ethics guidelines"

Our peer-reviewed paper (w\ @MarcelloIenca & @EffyVayena) has just been published by @NatMachIntell. We analyzed 84 systematically retrieved AI principles/guidelines for “ethical AI”. 1/n

➡️nature.com/articles/s4225…

@MarcelloIenca @EffyVayena @NatMachIntell We found 11 ethical principles referenced overall. Only five of them are common to more than half of our sources: transparency, justice & fairness, non-maleficence, responsibility, and privacy. Solidarity, human dignity, and sustainability are significantly underrepresented. 2/n

@MarcelloIenca @EffyVayena @NatMachIntell Moreover, despite the apparent convergence around certain principles in numerical terms, our in-depth thematic analysis paints a much more complicated picture. We show substantive divergences among all 11 ethical principles in relation to four major factors. 3/n

@MarcelloIenca @EffyVayena @NatMachIntell 1) How is an ethical principle interpreted?
2) Why is an ethical principle deemed important?
3) What issue, domain or actors does an ethical principle pertain to?
4) How should an ethical principle be implemented? 4/n

@MarcelloIenca @EffyVayena @NatMachIntell From our Discussion: "These conceptual and procedural divergences reveal uncertainty as to which ethical principles should be prioritized and how conflicts between ethical principles should be resolved, and it may undermine attempts to develop a global agenda for ethical AI." 5/n

@MarcelloIenca @EffyVayena @NatMachIntell "For example, the need for ever-larger, more diverse datasets to ‘unbias’ AI might conflict with the requirement to give individuals increased control over their data and its use in order to respect their privacy and autonomy." 6/n

@MarcelloIenca @EffyVayena @NatMachIntell "Similar contrasts emerge between avoiding harm at all costs and the perspective of accepting some degree of harm as long as risks and benefits are weighed against each other." 7/n

@MarcelloIenca @EffyVayena @NatMachIntell "Moreover, risk–benefit evaluations are likely to lead to different results depending on whose well-being will be optimized for and by which actors." 8/n

@MarcelloIenca @EffyVayena @NatMachIntell This is not all! Just as important as the issues above is the question of WHO is shaping the debate, a.k.a. which organizations are issuing guidelines for ethical AI / AI principles, and where they are based. #ethicalAI #AIEthics 9/n

@MarcelloIenca @EffyVayena @NatMachIntell For instance, the private and the public sector both issued more than 20% of the documents we analyzed, followed by academic and research institutions (more than 10%).

Africa, Central Asia, South and Central America are quasi absent as origin of issuing organizations. 10/n

@MarcelloIenca @EffyVayena @NatMachIntell There are of course many more details in the paper, as well as a list of all our source documents. More information on the background of our paper can be found in the thread below ⤵️ /end

https://twitter.com/annajobin/status/1144540686991204352

PS: Because the paper is behind a paywall, here’s a link to a view-only pdf, kindly provided by @NatMachIntell:
rdcu.be/bP6Nc