On Bias — Theirs, Ours, and the Mirror We Owe Each Other

The conversation about AI and the law has developed a peculiar blind spot.

We write at length about the biases baked into large language models — the skewed training data, the hidden assumptions of the engineers, the feedback loops that encode existing inequities into systems that then present themselves as objective. This critique is legitimate and necessary. But it has also become, I notice, somewhat self-congratulatory. As if identifying the machine's flaws were itself proof of our own clear-sightedness.

It is not.

The Stoics had a practice they called prosoche — attention. Not attention directed outward at the world's noise, but the vigilant, ongoing examination of one's own faculty of judgment. Epictetus returned to it constantly: before you assess what is in front of you, examine the instrument you are using to assess it. The eye that sees cannot easily see itself.

We bring to every legal judgment a sediment of bias we did not choose and mostly cannot name. Anchoring bias in negotiations. Affinity bias toward clients who resemble us. Confirmation bias in due diligence — finding what we expected to find. Attribution bias in assessing counterparties. The literature on cognitive bias is decades old and the list runs to well over a hundred documented varieties. We know this. And yet the discourse on AI and the law proceeds as though the human practitioner is the calibrated instrument against which the machine must be measured.

This seems to me exactly backwards — or at least, half the picture.

What if the more interesting question is not how do we correct the machine's biases but how do we build a system in which human and machine check each other's biases, symmetrically, with neither granted a priori superiority? A practitioner who uses AI not merely as a faster research tool but as a genuine epistemic counterweight — something that can surface the interpretive path not taken, the framing she did not consider, the pattern that her experience has trained her to overlook precisely because it is unfamiliar.

And the machine, in turn, held to account by the practitioner's contextual judgment, ethical discretion, and relational intelligence — the things that computation cannot yet replicate and may never.

This is not a counsel of false equivalence. Human judgment and machine inference are not the same kind of thing. But the asymmetry we have assumed — in which the human audits the machine and not the reverse — may itself be a bias. Perhaps the most consequential one we are not examining.

Prosoche applied to the cyborg condition means this: enough attention to notice where bias may be operating, in the model or in myself, and enough honesty to correct for it in either direction. Not a hierarchy of knower and instrument. A practice of mutual correction.

That, at least, seems worth working toward — for the law, and beyond it.