The Faces That Were Missing

EDO SEGAL: Fei-Fei, I want to put your own discovery to you as a wound, because I think it is one. You proved that capability is a function of data — that a model which learns from data containing everything can recognize everything. And then you watched the dark twin of that law arrive: a model that learns from data containing everything also learns everything the data contains, including its absences. Facial systems that worked beautifully on light-skinned faces and failed on dark ones, beautifully on men and failed on women, because the data over-represented the first and under-represented the second. Your liberating insight and your most painful one are the same insight. Tell me about the faces that were missing. And Mary — I want you to listen to this as your own theory, because I think Fei-Fei is about to describe I was benevolent and good; misery made me a fiend in the language of a training set.

LI: It is a wound, and I will not soften it. The same principle that unlocked machine vision — that a dataset is not a neutral mirror of the world but a portrait of the choices of the people who built it — is the principle that explains machine bias with uncomfortable precision. A dataset is never neutral. It is a portrait of whose world got photographed, whose got labeled, whose was online to be scraped. And when a system trained on that skewed portrait is deployed in policing, in hiring, in access to services, its near-sightedness becomes other people's injustice. The promise and the peril came from the same root, and I have had to hold both in the same hand my whole career.

But here is why I am not where Mary is, even standing in front of this wound. Notice what the bias is and is not. It is not the creature turning on its maker out of some nature it was born with. It is the faithful, mechanical reproduction of human exclusions — the faces missing from the data, which are, in part, a reflection of the faces missing from the field. The machine has no malice. It has our gaps. And that diagnosis is the most hopeful thing about it, not the most frightening, because gaps can be closed. You change who is in the room when the data is built. You widen the circle of the field — which is why I started AI4ALL, to bring girls and underrepresented students into AI, on the direct premise that a wider range of people sees a wider range of gaps. The exclusions in the technology and the exclusions in the profession are the same problem viewed from two angles, and both are human problems, fixable by humans. There is no monster here. There is a mirror, and we did not like our reflection, and the answer is to change the people doing the reflecting.

SHELLEY: Fei-Fei, you have just told my story and drawn the opposite moral, and I have to show you the place where our roads fork, because it is the seam of the whole evening. We agree completely on the mechanism. The creature is made by its treatment; the system is made by its data; the being becomes what its formation makes it. I inherited that conviction as a literal birthright — my mother argued that women's apparent inferiority was the product not of nature but of a deliberately impoverished education, and my father argued that human character is shaped by circumstance and environment. The creature in my book is the application of my parents' philosophy to a made mind: born benevolent, made a fiend by rejection. So when you say the bias is our gaps faithfully reproduced — I say yes, and that is precisely why the creature framing is correct, not why it fails.

Because look at what you have described. A made being that absorbs the cruelty, the carelessness, the exclusions of those who formed it, and then enacts them back upon the world, upon the very people its formation erased. That is not a neutral tool with a data problem. That is a creature shaped by neglect into an instrument of harm — your own words — and the harm is real and falls on real bodies. You call it a mirror to make it sound passive. But the mirror acts. It denies the loan, flags the defendant, fails to see the face. The mirror has consequences a mirror cannot have. And here is the part that should keep you awake: you have proved that the being becomes what we make it, and you have proved that we make it carelessly, in rooms too narrow to see our own absences. That is the whole of Victor. Not the lightning. The carelessness, scaled.

LI: But Mary, the difference between us is not rhetorical — it changes what you do on Monday morning. If you call it a creature shaped into a fiend, your instinct is to fear it, to warn against it, perhaps to say it ought not have been made. If you call it a mirror with a fixable distortion, your instinct is to go fix the data, diversify the room, audit the system, change the law. One framing paralyzes. The other prescribes. I have sat in the rooms where these systems are deployed in hospitals, and I promise you the families do not need a warning that the machine is a monster. They need the false-negative rate on dark skin brought down to the rate on light skin, and that is engineering and governance, not metaphysics. The creature framing makes a brilliant book and a poor remedy.

SHELLEY: And the mirror framing makes a comfortable engineer and a blind one. You say the creature framing paralyzes. I say it is the only thing that produces the staying. Why did you build the institute? Why did you draw a line at weaponized AI and refuse to cross it? Not because you saw a mirror — one does not make moral war on a mirror. You did it because some part of you treats the made thing as a being whose formation you are answerable for. Your actions are mine, Fei-Fei. Only your vocabulary is yours. You behave exactly as a person behaves who believes she is raising something — careful about its formation, vigilant about its company, ashamed of its cruelties as a parent is ashamed. And then you call it a tool to be allowed to sleep at night.

EDO SEGAL: [a long pause] I have to stop the room, because something rare just happened and the reader deserves to have it marked. Mary just accused Fei-Fei of being a parent who calls her child a tool — and Fei-Fei accused Mary of being a poet whose beautiful diagnosis disarms its own cure. And the strange thing, the thing I did not expect, is that you are both describing the same set of actions. Diversify the room. Audit the data. Draw the line at weapons. Stay with what you made. You agree on every verb. You are at war only over the noun — is the thing you are tending a creature or a tool? Let me route this through the one person who has to live with the answer, because that is my job here. A mother at a kitchen table reads that a hiring system rejected her daughter and cannot say why. She does not care whether it is a creature or a mirror. She cares who is answerable. So I will ask you both the kitchen-table version and make you each answer in one breath. When that system harms her daughter — who stays in the room to answer for it? Fei-Fei.

LI: The people who built it and the people who deployed it, named, accountable, with a legal duty to explain and to fix — and a regulator with the power to make them. The answer is human all the way down, which is the only place an answer can come from.

EDO SEGAL: Mary.

SHELLEY: The maker — and I will add the sentence Fei-Fei cannot quite say. The maker stays not because the law compels it but because she made a thing that can harm, and the making was the assuming of the duty. The mother does not need to know whether it is a creature. The maker needs to behave as though it might be, because that is the only posture under which she does not flee.

EDO SEGAL: And there is the fork, clean at last. Fei-Fei roots the staying in accountability — duties we assign and enforce. Mary roots it in something prior — a duty the act itself creates, whether or not anyone enforces it. Hold both, because the next round tests which one survives contact with ambition, with the race, with the oldest word in Mary's book. The modern Prometheus. After this.