Provenance: How to Establish Credibility When Every Signal Can Be Faked
In December 2025, NeurIPS — the world’s most prestigious machine learning conference — invited Zeynep Tufekci to speak. She is a sociologist at Princeton. She doesn’t build AI systems — she studies what they do to institutions, to trust, to the mechanisms societies use to establish what’s true. Her argument to the room full of people who build AI: you are afraid of the wrong things.
She was there because the AI research community had started to wonder whether the questions it knows how to ask are the right ones.
The nightmares dominating the cultural conversation about AI — superintelligence, AGI, mass unemployment — require AI to be extraordinary almost beyond our conception. Tufekci’s argument is that the real disruption requires no such thing.
It only requires AI to be good enough, cheap, fast, and deployable at scale. She calls this “Artificial Good-Enough Intelligence,” and her point is that this threshold has already been crossed, quietly, and without the fanfare we’d anticipated for the moment.
What Artificial Good-Enough Intelligence means in practice is this: since we don’t have direct access to the things we most need to evaluate — effort, sincerity, learning, authenticity, accountability — we infer them from observable signals. In other words, society runs on proxies.
A polished, personalized cover letter signals genuine interest because, historically, it was difficult to produce. An essay demonstrates learning because it required sustained effort to write. Named authorship carries accountability because the name can be traced.
None of these are perfect signals. But they have been functional. Institutions built hiring, admissions, credentialing, and trust on top of them. What AI has done is not merely make those signals unreliable. It has made them actively misleading, because the cost to produce a perfect facsimile has collapsed to near zero.
When any signal is cheap to fake, you can no longer use it to infer anything. The correlation breaks.
This is not an abstract problem. Anyone whose reputation was built on signals that can now be perfectly faked has had their credibility infrastructure hollowed out. The signal that conveyed their credibility no longer means what it used to. The honest and the fabricated now look identical at the point of evaluation. That is a different kind of problem than simply being harder to believe.
I discovered Tufekci's argument while I was researching what signals authenticity — in the sense that an artwork is authenticated — when everything can be perfectly replicated. The question I kept asking was: what remains, when the object can be faked, as proof that something real occurred? The answer I kept arriving at was context and provenance.
Two words you probably hear more associated with art.
Consider what happened recently, when a Rothko sold at Christie's for $98 million — a new record for the artist. The painting had belonged to Agnes Gund, who bought it directly from Rothko in 1967, three years before his death, and hung it in her living room, where it stayed until she died.
That chain of custody is not incidental to the price. You can replicate the pigment, the scale, the color relationships. You cannot replicate the fact that Agnes Gund stood in front of Mark Rothko and handed him money, and that the painting then lived inside her life for decades. What the auction house is pricing is not the visual experience — a high-resolution reproduction would do that adequately. What it is pricing is the provenance: the unbroken, witnessed, human chain of custody that makes this particular object traceable back to a specific moment between specific people.
Provenance is not a property unique to objects. It applies to any claim — any knowledge, any judgment, any insight — that can be traced back to a witnessed moment. Which means it applies to people.
I am a context window.
And so are you. But you’re a different one.
We’re not context windows in the literal sense, of course. But we are in the sense that matters for this discussion. What makes any human thinker, practitioner, or communicator valuable is not, in the end, the general knowledge they hold. It is the specific, accumulated, irreducible context that was built by living through particular experiences — particular conversations, particular decisions, particular moments of recognition and failure.
That context cannot be detached from the history that produced it. It has provenance. It is not portable. It cannot be extracted or transferred into a prompt. It exists because the person was there.
Artificial intelligence has something that looks structurally similar but is categorically different. It has mass context — patterns distilled from an enormous volume of human production. This is genuinely impressive and genuinely useful for what the philosopher Wilhelm Windelband, in 1894, called “nomothetic” work: finding the general principle, generating the common form, synthesizing across many. Windelband drew a line between nomothetic knowledge — law-seeking, pattern-seeking, concerned with the general — and idiographic knowledge, which concerns the particular: this person, this decision, this consequence, at this moment in time. He was defending history and the humanities against the assumption that they were just failed natural science. His argument was that idiographic knowledge is not a lesser kind. It is a different kind.
AI is constitutively nomothetic. It learned from the general. It works in the general. When you ask it to help with something that requires the particular — your specific voice, your specific client history, the specific texture of a conversation that happened last Thursday — it will, by default, substitute the general for the particular. It will give you something that resembles what you asked for, assembled from patterns. The resemblance can be remarkable. But the substance is not there.
The difference between a specific prompt and a general one is more significant than it appears. A constrained, particular prompt — one that hands the model only the precise situation you’re in — forces it to stay there. A vague prompt gives it room to find familiar patterns, to drift from your specific situation toward the average of everything it has seen that resembles it. Constraint is a workaround: a way of forcing the model to stay inside the particular rather than a solution to its gravitational pull toward the general.
The epistemological problem this creates — how do you establish what's real when you can't access the original, only representations of it — is not new. Historians have been working that cold case for centuries.
When historians study the past, they cannot directly access what happened. The event is gone. What remains are traces: documents, artifacts, testimonies, accounts. The historian’s job is to evaluate those traces — to figure out which ones are credible, which have been falsified or distorted, which corroborate each other, and what picture emerges when they are laid alongside each other. The methods they developed for this problem are not perfect, but they are rigorous in a particular way: they are designed to establish credibility for claims about particulars in the absence of direct verification.
The tools are worth naming. Provenance asks where a source came from and who handled it — what is its chain of custody? Corroboration asks whether independent sources, without access to each other, point to the same event. Testimony asks who witnessed this, what their relationship to it was, and what their interests were. And internal consistency asks whether the account contradicts itself, or contradicts other things the same source has produced. Each of these is a method for tracing a claim back to a moment — back to something that actually happened, that other people were present for, that leaves a residue in the world beyond the claim itself.
We are living in an epistemological condition that now requires these tools. Most of us are not trained to use them.
The question “how do we know what happened?” — which historians have always asked about the past — is now also the question we are forced to ask about the present.
How do we know this essay was written by this person? How do we know this image records what it claims to record? How do we know this account reflects what the customer actually said? The answer, increasingly, is: we trace it back. We ask for provenance. We look for corroboration. We identify the witnesses.
This points to a distinction that gets to the center of the problem. There is a difference between insight and testimony.
An insight is something you arrived at. It may be true, and it may be valuable, but it is in principle unverifiable on its own. “I’ve noticed that organizations with clear narrative architecture outperform those without.” This might be correct. But it is a claim floating free of any specific moment that could be examined. No one else was there when you noticed it. There is no record of the instance that generated the observation.
Testimony is different. Testimony is: this happened between me and another person, and that person exists. It has a witness built into the structure of the moment itself. The customer said this. The CSR was on the call. The CEO responded this way in that room. These moments are not just true; they are, in principle, corroborable. The other person remembers. The call log exists. The email thread is there. The witnesses can confirm.
This is the distinction that matters for anyone trying to establish credibility in a world where the proxies have broken. The question to ask of any communication is not “is this well-written?” or even “is this true?” but “can this be traced?” — traced back through the record to a moment that had human witnesses, a moment whose existence does not depend solely on the person making the claim.
This will not happen uniformly. Most audiences, most of the time, will not do the work of tracing. They will read the cover letter, accept the case study, move on. The demand for provenance concentrates where the cost of being wrong is high enough to feel. Enterprise buyers evaluating vendors. Editors deciding what to publish under their masthead. Institutions making decisions they will be held accountable for. These are the audiences for whom the broken proxy already has a price — who have started to discover, slowly, that the polished account that checked every box was assembled rather than earned. Provenance becomes a competitive advantage not because everyone will ask for it, but because the people whose judgment matters most will start to.
The best credibility, in other words, is not manufactured. It is mined. It comes from excavating what actually happened — the calls that were made, the conversations that took place, the decisions that were argued about in real rooms — and rendering that particular with enough texture that it remains recognizable as particular rather than general.
There is a version of this that will eventually also be gameable. As AI gets better at mimicking the texture of the particular — the awkward aside, the inconvenient detail — the simple presence of grime in a story will no longer be sufficient proof that the story is real.
What remains harder to fake, for now, is the network of corroboration: the other people who were in the room, the client who remembers the conversation differently, the colleague who was cc’d on the email. You can fabricate a story with a single act of synthesis. Fabricating a distributed network of adjacent people — each with their own account, their own memory, their own paper trail — requires coordination that multiplies both the effort and the exposure. It is not impossible. But the cost asymmetry is real, and cost asymmetry is what credibility has always run on. The question is not whether this threshold is permanent. It is whether it is high enough, right now, to matter.
This is why the human-to-human moment — not just “something happened” but “something happened between people” — becomes the credibility bedrock. It is the moment that carries witnesses structurally, not incidentally.
Most organizations are actively destroying these moments as fast as they are created.
The customer service call happens. It becomes a CSAT score — a number that records whether the interaction occurred and whether the customer was satisfied, but strips away everything human about the exchange. What did the customer actually say? What specific frustration did they articulate? What analogy did they reach for when trying to explain their problem? What word did they use three times without realizing it? Gone. The score captures the outcome and discards the content.
The sales conversation becomes a CRM entry. “Discussed pricing, customer expressed concern about implementation timeline, advancing to proposal stage.” This is not wrong. It is the process of converting particular into general — taking the living texture of a forty-five-minute conversation between two people and compressing it into a category. By the time it reaches the quarterly review, it is a data point. By the time it is on a slide, it proves nothing, because it came from everywhere and nowhere simultaneously.
What is left, after this process of abstraction, is the kind of material that AI is very good at working with and very bad at improving. The patterns it can find in a dataset of CRM notes are real patterns. But they are patterns in what organizations chose to record, which is already a filtered and compressed version of what actually happened. The particular — the specific customer in the specific conversation — has already been extracted before the model ever sees the data.
This is the structural argument for recovering the primary material before it disappears. The particular is where credibility lives, and the particular has a very short half-life in most organizational systems.
But organizations are only the first place where it gets destroyed. The second is in the act of communication itself.
The natural impulse, when you have a strong story, is to shape it. To find the arc, clean up the rough edges, remove the parts that do not serve the lesson. This impulse is not wrong — structure helps people receive things. But there is a version of this shaping that destroys the very quality that made the story worth telling. When you smooth it into a three-part case study with a clear lesson and a tidy resolution, you have done to your own material exactly what the CRM did to the sales call. The grime is gone. And with it, the proof.
The grime — the awkward detail. The moment that doesn’t quite resolve, or the thing that happened that the person telling the story isn’t entirely sure what to do with — that’s what signals that something is real. It is the residue of contact with the particular. Fabricated stories tend toward the clean because there is no reason to include anything that doesn’t serve the point. Actual stories have texture that the teller can’t fully account for, because it came from something that happened, not from something designed.
This is not a license for incoherence. The point is not to leave everything raw. It’s to preserve enough texture that the story remains traceable or that someone reading it could say, “that could only have come from somewhere specific,” and be right.
When the general is free and the particular is scarce, that texture is the proof of work. Most organizations are generating it constantly — in every customer call, every sales conversation, every decision argued about in a real room — and discarding it just as fast. What remains is pattern-ready and credibility-poor. The particular was there. Someone just decided it wasn’t worth keeping.