Operationalizing AI Ethics Principles | Cansu Canca
Subtitle: A better ethics analysis guide for developers.
Premise:
However, the multiplication of these mostly vaguely formulated principles has not proven to be helpful in guiding practice. Only by operationalizing AI principles for ethical practice can we help computer scientists, developers, and designers to spot and think through ethical issues and recognize when a complex ethical issue requires in-depth expert analysis. (18)
This still does not amount to "complete systems for ethical decision-making" (18).
AI Ethics Lab developed Toolbox for displaying the dynamics of AI principles.
1) sort, locate and visualize sets of AI principles demonstrating their chronological, regional, and organizational development 2) compare key points of different sets of principles 3) show distribution of core principles 4) systematize the relation between principles.
Main point is to distinguish core versus instrumental principles:
By collecting, sorting, and comparing different sets of AI principles, we discovered a barrier for operationalization: many of the sets of AI principles mix together core and instrumental principles without regard for how they relate to each other. (18)
In any given set of AI principles, one finds a wide range of concepts like privacy, transparency, fairness, and autonomy. Such a list mixes core principles that have intrinsic values with instrumental principles whose function is to protect these intrinsic values. Human autonomy, for example, is an intrinsic value; it is valuable for its own sake. Consent, privacy, and transparency, on the other hand, are instrumental: we value them to the extent they protect autonomy and other intrinsic values. Understanding these categories and their relation to each other is the key to operationalizing AI principles that can inform both developers and organizations. (19)
The only guide that makes a proper distinction between core and instrumental principles is EI-4People. Ref: Floridi et al., “AI4People — An Ethical Framework for a Good AI Society: Opportunities, Risks, Principles, and Recommendations, Minds & Machines 28, (2018), 689–707; https://bit.ly/343poNt
Core vs. instrumental principles
Core (also appears in almost all principle sets): "autonomy, beneficence (avoiding harm and doing good), and justice." (19)
Transparency and accountability:
Take transparency for example. We do not think that something is valuable in and of itself solely because it is transparent. Rather transparency is valuable because it allows us to understand, engage with, and audit the systems that affect us. In other words, transparency is instrumental to uphold intrinsic values of human autonomy and justice. Similarly, accountability in itself is not an end but rather it is a means to safeguard justice by assigning responsibility and to avoid harm through deterrence. (20)
Operationalizing the AI Principles to Guide AI Practice
Why is the difference between core and instrumental principles so important?
"... because instrumental principles are interchangeable. Individuals and organizations can weigh instrumental principles to determine which to prioritize and how to use them to best achieve the core principles." (20)
Example of explainability:
It is an instrumental principle that can uphold human autonomy by allowing individuals to interact meaningfully with the system. It can also help minimize harm and safeguard justice by making it easier to detect errors and biases. Explainability is crucial for a risk analysis AI system that helps judges set bail and parole in the criminal justice system. (20)
BUT... (and I also pointed this out in my own work on assessing AI principles) there are also cases where explainability is not necessary. Example given here: "explainability is not necessary for an AI system t hat optimizes energy use in cars." (20) Why is it optional? Because it's not necessary for upholding autonomy. In fact, in some cases explainability has a trade-off with accuracy, and in these cases explainability can hurt accuracy and safety (which does affect human autonomy and justice).
Bottom line: "there is no deep ethical dilemma when an instrumental principle is not suitable for a given case."
Once we correctly categorize principles as core and instrumental, we can turn many vague AI principles into an operational checklist to guide practice. (20)
When properly operationalized, AI principles provide a helpful start for an ethics analysis and they can guide developers and organizations through many ethical questions, even though they are not sufficient for complex ethical problems. (21)
When core values are in conflict, in-depth ethical analysis is still required. In case of conflict, a statement can help an organization (or ethicists hired by them) to prioritize core values. (So they are still useful, even though most principle statements largely share the same principles).