Normative Philosophy of Computing

This handout is for my presentation at FAccT 2023. The full paper is here. I wrote it with Jake Stone.

1. Introduction

Maximalists about algorithmic fairness say: forget about abstract and idealised criteria. Build just societies, not just algorithms.

Minimalists say: leave social costs to policy-makers. Just focus on model performance.

We argue: predictive models not subject only to epistemic standards, but also to predictive justice, distinct from outcome or procedural justice that apply to decisions made with the model.

Norms of predictive justice can be situated and contextual, don't have to be abstract and idealised. And they are the proper responsibility of ML engineers, who should be leery of technocratic implementation of 'just societies'.

2. Predictions, Decisions, and Epistemic Ethics

Prediction = inference from past experience to out-of-sample cases, identifying probability target variable obtains (not just about future).

Interested in all predictions, not just ML.

Decisions = what you do with the model, how you use it to change the world.

Moral norms constrain decisions.

Epistemic norms constrain epistemic practices.

Our thesis: Predictive models are constrained by moral norms of predictive justice, which are explanatorily distinct from the norms governing the decisions to which they lead.

Existing theories of epistemic ethics can light the way to predictive justice (for more on how, see the paper).

3. A Principle of Predictive Justice

PPP: A model is predictively just only if its performance for systematically disadvantaged groups cannot be improved without a disproportionate decline in its performance for systematically advantaged groups.

Not settling what makes a salient group, or intersectionality. PPP is situated: relies on context of background injustice. It is political, not welfarist.

Predictive and other representational models deployed within structurally unjust societies are both caused by that background injustice, and help sustain it.

Representations routinely centre the experiences of advantaged populations as 'normal', and optimise for that 'normal' setting. As a result perform worse for disadvantaged groups.

From sensors to the census, from medical research to computer vision, our resources to understand and represent the world work better for the better-off.

This in turn helps sustain that structural injustice.

If this empirical claim is true, then we have reason to counter it when developing predictive models, by prioritising their performance for systematically disadvantaged groups.

Justified on backward-looking remedial grounds, and on forward-looking grounds not to contribute to or participate in the cultural schemas that sustain background injustice.

4. Degrees and Kinds of Predictive Injustice

Degree of avoidable underperformance. Key point here is avoidable. If relative underperformance is due to genuinely innocent statistical artefact (e.g. inframarginality) then that mitigates predictive injustice.

Corollary of this: the more objectionable the reason for relative underperformance, the more serious the predictive injustice. If due to being trained on socially unjust past, that exacerbates wrong. This means that understanding underperformance is really important (consider e.g. invigilation software that works worse for darker skin in part due to how cameras treat lighter skin as normal).

Social power of the agent endorsing the model also matter: it’s more unjust, other things equal, for a powerful organisation to endorse a predictively unjust model. E.g. state vs individual, big tech company vs small contractor. Acquiescence to injustice matters more the greater one's social power.

5. Cosmic vs Situated Predictive Justice?

Predictive justice doesn't have to be idealised and abstract, 'cosmic' approach, like often found in ML literature, exemplified in philosophical literature by Brian Hedden's work.

We discuss it in the paper—no time for details now. Basically we think in Hedden's idealised case the algorithm's differential performance fully explained by an innocent statistical artefact; no systematic background injustice; no stakes, so no predictive injustice.

If we add in repeat games under same conditions, with benefits and burdens attached, then starts to seem more unfair. Next, assume systematic background injustice, etc. Highly abstract hypotheticals sometimes remove the features that cause the injustices they are intended to analyse, then assume that simple compositionality applies. This is false.

Necessary norms of justice are not the only ones. Situated, contextual norms of justice matter too, arguably more.

6. The Place of Predictive Justice

Predictive justice is not the most important thing. But it does matter, and there are good reasons to keep it separate from outcome (or procedural) justice.

1. You want to justify the decision because it was supported by the prediction. But evaluation of the prediction presupposes the decisions it will make. That's a worrying circularity.

2. Failing to clearly distinguish sites of justice means slippery slope towards just developing predictive models to support political ends.

3. Technocratically manipulating predictive models to build your idea of a just society may bypass the democratic process.

So: identify range of reasonable thresholds to test performance. Mitigates circularity worry, leaves decisions to those authorised to make them (not authorised to make unreasonable decisions), and keeps disagreements separate.

Outcome and procedural justice matter more, true. But predictive justice is the particular responsibility of those who design and deploy ML models. For outcome and procedural justice, important to give those with proper authority influence over the decisions.