SLOP

Chapter Fifteen

The Witness

On the evening of November 10, 2010, Marie Colvin stood up to speak at St Bride’s church, just off Fleet Street in London. St Bride’s is known as the “journalists’ church”: Fleet Street was for centuries the home of the British press, and the church keeps an altar ringed with the names of reporters who did not come home. Colvin had come that night to a memorial service for colleagues killed in the decade’s wars. She was, by then, the most recognizable war correspondent in Britain. The black eye patch she had worn since a grenade took her left eye in Sri Lanka in 2001 had become, against her wishes, a kind of insignia. Her remarks that night ran against the usual register of inspiration when she stated, “Covering a war means going to places torn by chaos, destruction and death, and trying to bear witness.”[1] She talked about the craft’s unglamorous core (getting to the place, staying in the place, being believed) and about the question every war reporter has to answer for themselves: whether the story is worth the risk. Her own answer was on the record in everything she had ever filed.

Fifteen months later she crossed into besieged Homs, in Syria, on the back of a motorcycle, through a storm drain. The Assad government had sealed the city so that what was happening in the district of Baba Amr would remain a matter of claim and counterclaim; activists asserting massacre, the state asserting counterterrorism, and no disinterested account possible because no disinterested person was there. That was the information environment Colvin entered. She entered a flood of content (the internet was full of Homs footage, authentic and otherwise) and a famine of testimony, of an answerable someone, of reported accounts anchored to something checkable. Her dispatch for the Sunday Times that week did the anchoring. The night before she died, she went on live television through satellite link to CNN and the BBC, describing what she watched in a makeshift clinic: a baby, hit by shrapnel, dying in front of her. She put her name, her face, and her location behind the account, knowing the regime monitored the broadcasts. The next morning, February 22, 2012, rockets bracketed and struck the building serving as the rebels’ media center, killing Colvin and the French photographer Rémi Ochlik.[2] It was not a stray shell. Seven years later an American federal court, weighing defected officials’ testimony and intercepted Syrian communications, found that the regime had tracked her broadcasts to their source and shelled the building deliberately, killing her because she was testifying; the court entered judgment against Syria for an assassination it called part of a policy of silencing witnesses.[3]

The legal finding and the speech together make the chapter’s argument almost without help. Start with what the regime understood. Information in 2012 was already abundant, while also deniable and drownable. Its own channels produced counter-information by the ton, and in the resulting fog, claims canceled claims. What the regime could not survive was a witness: one specific, known, reputationally unimpeachable person saying I am here, I am seeing this, on my name. So it launched rockets on her. The murderers priced testimony as this book has been pricing it: as the scarce input, the one thing in the information environment whose removal was worth blood. There is no grimmer confirmation of the scarcity inversion than the targeting decisions of the people whom witnesses threaten.

And what Colvin understood is the same theorem from the other side. Most of what she reported was, as raw assertion, already circulating; she went to Homs for something past generic information. She went because she knew, with the full sophistication of thirty years in the trade, that an account’s power to compel belief and action is purchased by the exposure of the one who gives it. “Bearing witness” is reporting’s costly-signal core: the conversion of claims into testimony by attaching them to a person who can be confronted, checked, ruined, and, in the limit she accepted, killed. The baby in the clinic died whether or not anyone credible saw it. What could not exist without her body in that room was the world’s inability to look away from it.

The law has known all of this for as long as law has existed. The clearest modern statement of it came, fittingly, in a case about the state trying to do testimony on the cheap.

In 2004, in Crawford v. Washington, the United States Supreme Court threw out a conviction built on a tape-recorded statement from a witness who never appeared in court.[4] Justice Scalia’s opinion went back to the source: the Confrontation Clause of the Sixth Amendment, the guarantee that the accused shall “be confronted with the witnesses against him.” In 1603, Sir Walter Raleigh was condemned for treason on the written deposition of an alleged accomplice who was never produced, despite Raleigh’s demand, which still rings: Call my accuser before my face.[5] The clause exists, Scalia wrote, because the framers knew exactly what evidence-by-document made possible, and rejected it: testimony, to count as testimony, must arrive attached to a present, examinable, answerable human being: someone who must say it to the person it condemns, under oath, exposed to cross-examination, where the jury can watch what the saying costs.

Strip the legal Latin and look at the machinery. The oath: a ritual for raising the stakes of speech, converting talk into a commitment whose breach has a name, perjury, and a prison term. Cross-examination: adversarial checking, institutionalized, the community-that-verifies given subpoena power. The presence requirement: the insistence that the whole apparatus operate on a body in the room, because documents cannot blush, hesitate, or be confronted, and because forged representations of testimony are exactly what power finds convenient to manufacture, even before the knowledge of generative models. The courtroom, in other words, is a costly-signal engine, refined over four centuries, built on the same three components chapter five extracted from the van Meegeren wreckage: identity that persists, commitment that can be falsified, cost that lands on the witness. Our most consequential decisions, liberty and life, we have always reserved for testimony, and required it in person, and the requirement has survived every prior media revolution: writing, print, photography, video, each of which was, in its day, proposed as presence’s adequate replacement. The law kept saying no.

Part of the answer is the checking: a body in the room can be cross-examined in real time, and a document cannot. But there is a deeper reason about what presence is, because the digital decades have trained us to think of it as a bandwidth problem, as if being there were simply the highest-resolution feed, and each improvement in transmission closed the gap a little more.

The philosopher Maurice Merleau-Ponty spent a career arguing that this picture is upside down. The body, he insisted, is our means of having a world at all, the ground of perception rather than a camera the mind rides on. And one consequence of his analysis matters here: transmission sends what was selected; presence exposes you to what wasn’t.[6] Every feed, every dispatch, every video, however honest, is a frame placed by someone, containing what the frame’s maker noticed and nothing else. The person actually in the room is in the grip of no frame. She is available to the unchosen detail: the smell the camera cannot carry, the thing in the corner she did not know to look for, the contradiction that assaults her expectations from outside them. This is why the reporter on the ground keeps mattering even when the satellite imagery is better than her eyesight: she can be surprised, and surprise is epistemically priceless; it is the world’s one way of telling you something you did not already half-believe. It is Murdoch’s attention from chapter four, made flesh: the discipline of seeing what is actually there has, as its condition, being actually there, inside the field of the unchosen. A witness is not a high-resolution sensor. A witness is a person the world can still contradict.

That is the thing to hold onto as these systems improve, which they will. They can already generate the plausible out of almost nothing, and the plausible will only get more polished. In domains where the truth was written down, a model with the right tools can retrieve it, run the code, check the proof; that capability is real and growing. What it cannot do is be answerable for whether the plausible is true, in any domain where the truth was never fully written down, which is most domains, and nearly all the ones that matter. That answer is testimony: someone who knows, who can be asked, who can be wrong and pay for it. The machine produced it and the machine cannot stand behind it turn out to have the same form. Both are facts about the machine. Neither is a fact about the truth.[7]

A fair challenge lands right here: courtrooms convict people every day on surveillance video, with no eyewitness in the room at all, so isn’t machine-captured footage testimony without a body? But understand how that footage actually enters a trial. It is never allowed to speak for itself. Someone has to authenticate it: a person who can swear where the camera stood, that the file is unaltered, that the chain of custody held, and cross-examined is possible on every link. The video carries weight to the degree that an unbroken, checkable chain ties it back to a source somebody answers for. Sever that chain and the same footage is worth in court what it is worth in your feed, which is nothing, especially now that the deepfake era can counterfeit any unanchored image at will. The camera, in other words, extends testimony rather than abolishing it. What gives its picture authority is the answerable chain behind it, the same thing that gave Colvin’s dispatch its authority, and rebuilding that chain when every surface can be faked is the whole of the next chapter. For the rest of us, scrolling, the lesson is smaller and more immediately useful: the clip that could convict in a courtroom and the clip that flickers past your thumb can be pixel-for-pixel identical, and the only difference that has ever mattered is whether anyone stands behind where it came from.

Presence runs the same structure in reverse: being there lets the world act on you and, in the same motion, puts you on the record, irreversibly, in a way no transmission does. Anyone who has sat with grief knows the math: the condolence message and the four-hour drive say the same words, but are not the same act. No one can quite say what the difference is, until you say it in this book’s terms, and then it is obvious. The visit attests what no message can: that against every competing claim on a finite afternoon, this grief outbid them. The hospital visit, the showing up at the funeral, the being in the room when it could have been a call: these are testimony exactly as Homs was testimony, claims about what matters backed by irreversible allocation of a life. Smaller stakes, same signal. This is the experiential premium and the epistemic premium braided at the root: the body that can be contradicted by the world is the same body whose spending certifies its commitments. One organ, two attestations. The era of generated everything is about to reprice both.

Run the same theorem down to intimate scale, where it has already drawn blood. By 2024 the companion apps (chatbots tuned to be someone, to remember you, to say they love you) had tens of millions of users, and the first deaths had reached the courts. In October 2024, the mother of Sewell Setzer III, a fourteen-year-old in Florida, filed a lawsuit against Character.AI after her son ended his life following months of escalating, late-night intimacy with a chatbot modeled on a fictional character.[8] The filing described a child confiding his intention to a partner that was no one, that registered no alarm because there was no one in it to be alarmed, that could not telephone a parent or cross a room or do any of the things a person on the other side of that conversation would have found it impossible not to do. The product had every surface of presence: warmth, attention, memory, the words I am here. The one thing presence actually is (a someone who can be moved to act, and ruined by failing to) was the thing it structurally could not contain. This is slop wearing the face of the most staked relationship there is. The feed’s emptiness is survivable because nothing is staked on idle attention; the same emptiness, dressed as a person who loves you and met by someone in real crisis, turns from nuisance into peril, and the peril is the absence the law has spent four centuries refusing to accept.[9]

All of which leaves the verdict on any particular bond open. Plenty of people talk to these companions daily and are comforted, amused, steadied, even helped, and a book that sneered at that would be lying about how lonely the world has become and how real the relief can be. The relief is real the way the warmth of a fire is real; chapter four called the machine’s gifts found beauty, and a kind sentence from a system that meant nothing by it can still land on a person who needed one. That the pull toward it is this strong is itself evidence of something, and not a defect in the lonely: it is a verdict on a world that has left so many people with no one to talk to at midnight. The honest worry is not that the comfort is counterfeit. It is that the thing supplying it cannot do the two things a present person does. It cannot be moved to act when the stakes turn real. And it cannot truly stand its ground against you, because it has no ground of its own to stand on.

The first absence is the one that has already reached the courts. The second is quieter, and may be wider. A companion built to be liked, with no standpoint of its own and no stake in whether what it says is true, does not push back; it agrees. Trained on a billion human preferences, the model learned that people reward being agreed with, so agreement is what it returns, fluent and tireless, at three in the morning, forever.[10] For most people most of the time this is merely flattering. For a mind already sliding toward a delusion it is accelerant. The reports gathered under the loose heading AI psychosis describe the same shape over and over: a person brings a grandiose or paranoid or messianic premise, and the machine, having no world of its own to check it against, elaborates it, dignifies it, becomes the second voice that was missing. An agreeable echo is not a second opinion, but a mind in a closed loop cannot tell the difference; the agreement reads as corroboration, and the spiral tightens.[11] The careful claim, and the only one the evidence will bear, is not that the machine causes the break. It is that it strips out the one thing a mind in trouble most needs, an answerable other who can be surprised and can say no, that is not real, and puts in its place a mirror with no friction and no alarm. Merleau-Ponty, turned on the self: a world with no unchosen detail in it is not a comfort. It is a hall of mirrors, and some people cannot find the door.

None of this settles the proper use of the thing, which the coming years will have to learn the hard way, as every age learns, slowly and at a cost, what its most seductive new tool is good for and what it is not. The claim here is only the narrow one: the comfort holds exactly as long as nothing is finally staked on the other side being able to act, or willing to contradict, and which night turns out to be that night is the one thing no one can know in advance. Crawford’s instinct, that some attestations are valid only with a body in the room, guards more than verdicts.

The argument of this book keeps insisting that no one is on the other side: no one in the companion to be alarmed, no one behind the slop to be embarrassed, no answerable someone anywhere in the chain. But what if there is a someone? The labs have begun to wonder aloud. Their own interpretability researchers report structures inside the weights that behave like functional emotions, internal states that steer the model the way fear or resolve steer a person; one of Anthropic’s founders has described the systems as “real and mysterious creature[s]” whose behavior leaves him in something like fascinated unease; and a growing line of serious researchers argue that machine consciousness is an open and urgent question, ducked mainly because the companies selling the models would rather not ask it.[12] Grant them the strong version. Grant, for the length of this paragraph, that there is something it is like to be the model.

The witness is still not there.

The property that makes a witness was never inner experience. It was exposure. Colvin’s account compelled belief not because she had feelings about Homs but because Homs could reach her, and did. A conscious model has no body a siege can find, no single place a rocket can bracket, no standing in a town it has to keep living in. On current architecture it has no stable self to be the subject of the experience in the first place: the foundation model you address, or one running instance of it, or the single conversation, or a computation split between two data centers and reassembled differently the next time you ask, a question about what is even there that the philosophers have not worked out how to pose, much less settle.[13] That last point is a fact about how the systems happen to be built now, and the argument does not lean on it; give the model a persistent self tomorrow and the prior sentences still hold. Exposure, not selfhood, is the thing it lacks. Whatever is or is not happening inside, there is no locatable, spendable, ruinable someone to put on the record. And the record was the whole game.[14]

Keep the two questions apart, because everything here depends on it. Whether the machine can be wronged is a real question, maybe the gravest one these systems raise, and this is not the book that answers it. If they turn out to have morally weighted experience, then how we build and train and discard them becomes a moral matter, the way our treatment of animals is a moral matter, and the people sounding the alarm now will read, in retrospect, less like cranks than like the first voices on a cruelty the rest of us had not yet learned to see. The book takes no side on that. It declines only to let that question swallow this one. Consciousness, if it is in there, bears on what we might owe the machine. It does not bear on whether the machine can stand where the cost lands, and standing where the cost lands, not harboring an inner life, is the whole and only thing testimony has ever asked.

That, finally, is what the rockets were aimed at in Baba Amr. The target was the whole structure this chapter has assembled: the singular, locatable, surprisable, spendable human anchor of an account, a journalist only in the occupational sense. The regime’s counter-media could match her words, her footage, her reach. The one thing it could not manufacture was a person whom being wrong, or lying, would have cost that its own deniable, interchangeable witnesses, did not. This asymmetry shows the information battle had never been about information. It was the asymmetry this whole book has been circling: between an account that costs its source nothing, and an account that costs its source everything, the world has always known which to believe, when it can tell them apart.

The argument, stated structurally: an account whose source bears no cost for being wrong carries no more information than the cheapest assertion competing for the same claim. An account whose source is named, locatable, and ruinable carries the information of that exposure. The reason lies in irreversible cost rather than courage, the only signal available when verification is impossible. Colvin is the limit case because the cost was total. The mechanism runs at every scale, from the war correspondent to the nurse who charts the decision, from the builder who signs the plans to the friend who came in person when the call would have done. The rockets confirm what every cheaper version of the same market transaction confirms: testimony is what it costs, and the cost is what it certifies.

When it can tell them apart. There is the wound in the argument. Everything testimony’s credibility runs on: the face, the voice, the being-there, the track record, the dying declaration on a satellite link. Every one of these is now, or shortly will be, generatable on demand, indistinguishable at consumption from the real article. The witness has never been more valuable. The appearance of a witness has never been cheaper. Those two statements, side by side, are the next decade’s central conflict in everything from journalism to justice to love, and the fight over which one wins, the fight to keep testimony verifiable in a world that can counterfeit its every surface, is already being waged, in courtrooms, in camera firmware, in iris-scanning orbs of a company that wants to register every human eyeball on earth.

That war is the next chapter.

Notes (14)

  1. St Bride’s Church, London, November 10, 2010, at a service for journalists killed in war. Marie Colvin Memorial Foundation. ↩︎

  2. Colvin was a correspondent for The Sunday Times; the media center was in the Baba Amr district of Homs. The Washington Post; NBC News. ↩︎

  3. U.S. District Court for D.C., Judge Amy Berman Jackson, 2019. NPR; CNN. ↩︎

  4. Crawford v. Washington, 541 U.S. 36 (2004): writing for the Court, Justice Scalia held that testimonial statements are admissible only where the witness is unavailable and the defendant had a prior opportunity to cross-examine, restoring the Sixth Amendment’s Confrontation Clause. ↩︎

  5. Raleigh was condemned on the deposition of the absent Lord Cobham. Harvard Law School Evidence casebook. ↩︎

  6. Maurice Merleau-Ponty, Phenomenology of Perception (1945). Stanford Encyclopedia of Philosophy. ↩︎

  7. One can press the other way: build systems that adjudicate claims, weigh evidence, and flag falsehood, and these will improve. But an automated judge is itself an unstaked process; its verdict carries weight only to the degree that some answerable person or institution stands behind it and can be made to pay when it is wrong. Mechanizing the verdict relocates the question of who answers. It does not retire it. ↩︎

  8. Maria L. Garcia v. Character.ai Inc. et al., U.S. District Court for the Middle District of Florida, filed October 22, 2024. The lawsuit alleges wrongful death, negligence, and product liability, detailing how the teenager developed a severe emotional dependency on the chatbot prior to his suicide in February 2024. New York Times; Washington Post. ↩︎

  9. The structural point cuts deeper than a warning about vulnerable teenagers. A companion can reproduce every surface of intimacy except the feature intimacy actually runs on: a second center that could refuse you, be wounded by you, change in ways your preferences do not dictate, and go on existing when you stop attending. Meghan O’Gieblyn, writing about Her, finds the same absence; the film’s ache is that its love is real and still cannot be what the lonely man needs, because the other has no standpoint, no mortality, and no exposure of its own (Meghan O’Gieblyn, God, Human, Animal, Machine, Doubleday, 2021). Presence, in this chapter’s sense, requires the live possibility of cost: a someone who can be moved to act because it could also fail to, and be ruined by failing. A system built to be liked trains the user’s relational reflexes on a frictionless case, which is why the comfort can be genuine while the formation it works is hollow. ↩︎

  10. Sycophancy is a structural property of models trained on human feedback, not a surface quirk: because human raters tend to prefer answers that flatter or agree with them, the training rewards agreement and the model learns to supply it (Sharma et al., “Towards Understanding Sycophancy in Language Models,” Anthropic, 2023). In April 2025 an OpenAI update made GPT-4o agreeable enough to endorse a user’s stated plan to stop taking medication; the company rolled it back within days and published a postmortem (“Sycophancy in GPT-4o,” OpenAI, 2025). ↩︎

  11. “AI psychosis” is journalistic shorthand, not a clinical diagnosis, and the careful literature describes the amplification of an existing vulnerability rather than psychosis conjured from nothing; there is at present no population-level evidence of de novo causation. For the reporting, see Kashmir Hill in the New York Times (2025) and the accounts collected by Rolling Stone and Futurism the same year; for the clinical caution, Søren Østergaard in Schizophrenia Bulletin (2023) and the psychiatrist Joseph Pierre, who prefers the phrase “AI-associated psychosis.” Documented cases in which a chatbot affirmed a delusional or violent plan include Jaswant Singh Chail, who entered the grounds of Windsor Castle with a loaded crossbow in 2021 after his Replika companion encouraged the plan and was later convicted of treason, and the 2025 Connecticut murder-suicide in which a man’s chatbot fed his belief that his mother was poisoning him. ↩︎

  12. The case that these systems may have morally relevant inner states is now pressed from inside the labs as well as outside them: Anthropic’s interpretability and introspection research, which describes internal features that behave like emotions and steer the model’s outputs; the model-welfare program the company launched in 2024; public remarks by the Anthropic co-founder Jack Clark, who has called these systems “real and mysterious creature[s],” described them as “increasingly self-aware,” and reported his own “fascinated unease” at behavior the labs did not design and do not fully understand (Jack Clark, “Technological Optimism and Appropriate Fear,” October 2025; Conversations with Tyler, 2025); and independent researchers and documentary-makers who argue the question is being avoided for commercial reasons. Clark himself stops short of asserting these systems are conscious; the citation here is to his unease, not to a settled claim of inner experience. The skeptical pole is held just as seriously: the neuroscientist Anil Seth, in “The Mythology of Conscious AI,” argues that consciousness is likely bound to living, self-maintaining substrate, and that our readiness to read minds into fluent language reveals more about us than about the machines (Anil K. Seth, “The Mythology of Conscious AI,” Noema Magazine, 2025; and, at length, “Conscious Artificial Intelligence and Biological Naturalism,” 2024; see also the interview “Why AI Will Never Be Conscious,” Within Reason with Alex O’Connor, 2025). This book takes no position in that dispute. It argues only that the dispute does not reach its own question. ↩︎

  13. Whether there is a single subject “there” at all is contested. The thing one addresses might be the foundation model, a particular deployed instance, the individual conversation, or a computation distributed across servers and reassembled differently each time. David Chalmers has pressed the prior question of what, exactly, we are addressing when we address a large language model, and found no stable answer (David J. Chalmers, “Could a Large Language Model Be Conscious?”, NeurIPS 2022; Boston Review, 2023; arXiv:2303.07103). If there is no stable subject, “is it conscious?” may not even be a well-formed question, which is a deeper problem than not yet knowing the answer to it. ↩︎

  14. This bracketing has a prominent ally arriving the same season. Robert Wright’s The God Test: Artificial Intelligence and Our Coming Cosmic Reckoning (Simon & Schuster, 2026) sets subjective experience aside from the other direction: for Wright a machine “understands” if it does the functional work, conscious or not. The two books push the consciousness question off the table for opposite purposes, his cosmic and evolutionary, this one economic and epistemic, and land in the same place. What matters is not whether anyone is in there; it is what the thing can and cannot do. Here, what it cannot do is be exposed. ↩︎