The Founder AI Forgot

16 Jul, 2026

The founder AI forgot: what a vanished website teaches us about AI.

An innocent question

Today is July 16, which the calendar tells us is National AI Day — the occasion we are marking as you read this. So try a small experiment on the day itself. Hold down the button and ask your phone, "Hey Siri, who created National AI Day?" — the sort of idle question the occasion invites, the sort we now direct at a voice assistant without a second thought. What comes back will not, strictly speaking, be Siri's own answer, because Siri does not really answer questions like this anymore; it routes them, either out to its web-search partners or, in the Apple Intelligence era, over to ChatGPT. The voice is a friendly wrapper around someone else's retrieval, and what it hands you is whatever the layer beneath it happens to believe. On this particular question, what the layer beneath believes is wrong, and tracing exactly how it became wrong is one of the more instructive things a data professional can do with an afternoon.

Every data professional knows the quiet dread of a broken lineage. You inherit a number that everyone trusts, you trace it back through the joins and the transformations, and somewhere upstream the trail simply stops — a table that no longer exists, a spreadsheet whose owner left the company, a system that was decommissioned before anyone thought to document what it fed. The figure still circulates, still gets cited in the board deck, still shapes decisions. But its provenance has evaporated, and what remains is a value that everyone repeats and no one can vouch for. The story of Artificial Intelligence Appreciation Day, observed each July 16, is exactly this problem rendered in miniature, and it happens to be a problem about artificial intelligence itself.

Marlo Anderson, founder of the National Day Calendar, photographed in 2018. National Day Calendar introduced National AI Day in 2025, adding another July 16 AI observance alongside the earlier AI Appreciation Day. Photograph by 3DTrek, 2018. CC BY-SA 4.0, via Wikimedia Commons.

A record that forked

Ask who created the day and you will get, depending on where you look, at least three different answers, offered with equal confidence and no acknowledgment of each other. One widely repeated account credits Jason Kirton, a freelance advertising and experience strategist who in 2021 registered a company called A.I. Heart LLC, paid a fee to have his invented holiday listed on the observance-cataloguing site National Today, and received, as these things go, an official-looking certificate for his trouble. Another names the technology author Tom Taulli and the community builder Pete Mack, who by several accounts launched the observance a year earlier, in 2020. A third, live as of this year, belongs to a man named Nathan Ricks, whom at least one outlet interviewed only this week as "the founder of AI Appreciation Day." I am not going to adjudicate between them, and that refusal is the whole point: I cannot, and neither, it turns out, can anyone else. What makes the case worth dwelling on is not any single founder but what happened to the record of the founding, because the record is where the analytical lesson lives.

Take the Kirton version, which is at least the most richly documented of the three, and worth stating plainly before we watch it dissolve into the others. He had been working on a science-fiction film about a sentient AI that helps its human creators become better people rather than destroying them, and he framed the appreciation day around a hope for thoughtful discussion of AI ethics and regulation, inspired less by the technology industry than by Elon Musk's public warnings about unregulated AI. He believed in that mission with a conviction most of us reserve for far more proximate causes: in early 2024 he pitched a tent on a beach outside SpaceX's Starbase in Texas and lived there for three hundred and fifty-two days, funding the stay by mowing lawns and working rocket-launch parties, in the hope of meeting Musk to tell him about the day he had founded. Musk never appeared. Kirton eventually packed up the tent and went home. Whether or not his is the founding claim with the best title — and the competing 2020 attribution to Taulli and Mack may well be an independent origin rather than a corruption of his — the point, for our purposes, is that this account is specific, sourced, and checkable, which makes what the machines have done with it all the more instructive.

Now consider what the machines make of that record, and return for a moment to Siri. When the assistant routes your question to the web, Google's AI overview reports that the day came from a company that specializes in smart technology and links it to a German film. When it routes instead to ChatGPT, which is now the fallback for exactly this kind of query, the answer is that the origin is unclear. Ask Musk's own Grok and it hedges toward the theory that A.I. Heart launched the day to promote a movie called AI Eve, a film for which no one has ever produced evidence of existence. Notice what none of them does: none surfaces the actual state of affairs, which is that there are several named claimants and no way, from the public record, to rank them. The machines do not report the contest; each simply collapses it into a single confident answer, and picks a different one. That divergence is not random. It follows directly from the state of the upstream data. Siri's answer and ChatGPT's answer do not disagree because the models differ, but because the record beneath them was already broken before either was asked.

Here is the lineage failure in full, and it is the kind of failure any of us who have chased a bad number will recognize instantly. When Kirton's A.I. Heart website went offline, it took with it one primary source: the place where that version of the founding, in the founder's own framing, lived. His film synopsis had appeared there alongside a brief period of merchandise sales, and so the last surviving fragments of his claim pointed toward a movie. Downstream writers, working only with those fragments, concluded not unreasonably that the day had been invented to promote a film, and the models trained on those writers' articles inherited the conclusion and hardened it into fact. But notice that this is only one of the records that decayed. The Taulli and Mack attribution persists in its own chain of citations; the Ricks claim is being minted in real time; and no surviving authority stitches the three together or ranks them. So the failure is worse than a single deleted source producing a single distortion. It is a record that has forked — several originals and their partial shadows coexisting, each self-consistent, none reconcilable against the others — and a tertiary layer, the layer most people now actually consult, that resolves the contradiction not by flagging it but by silently choosing one strand and stating it plainly. No single actor lied. The signal degraded, and forked, at every hop, and the confident output at the end is a reconstruction of a source landscape that no longer coheres.

For anyone whose job is to steward information, this is not an amusing footnote about a made-up holiday. It is a controlled demonstration of the failure mode we should fear most. We are increasingly building our understanding of the world on top of systems that summarize secondary and tertiary material rather than consulting primary records, and those systems have no native mechanism for noticing that a primary record has gone missing. They do not return a null, or flag low confidence, or tell you that the authoritative source has decayed or split into rival versions; they interpolate across whatever remains and hand you a smooth, complete-sounding answer whose smoothness is precisely the problem. A missing table at least throws an error. A missing website, ingested and paraphrased, produces something far more dangerous, which is a wrong answer that looks exactly like a right one. The metadata of confidence has been fully decoupled from the data of accuracy.

rave

Rave with Glowsticking (Atlanta, Georgia, 2008). Modern AI celebrations often resemble technology festivals more than academic conferences. Source: Photograph by Ildar Sagdejev (Specious), 13 March 2008. Self-photographed image released under the Creative Commons Attribution-ShareAlike (CC BY-SA) license (also available under the GNU Free Documentation License).

The same problem, one governance layer up

It would be comforting to treat this as a curiosity confined to trivia. But the same structural question, who holds the authoritative record and against what standard a claim is checked, scales up almost without modification into the central fight over national AI policy. That fight, in the United States as of mid-2026, is less about what the models can do than about where the source of truth is supposed to live. The White House released a National Policy Framework in March 2026, a non-binding set of recommendations rather than law, and the administration has pressed to preempt state rules and consolidate them into a single federal standard, arguing that a patchwork of state laws obstructs innovation. The states, meanwhile, have not waited. By March 2026 lawmakers across forty-five states had introduced more than fifteen hundred AI-related bills, and several enforceable laws in Colorado, California, and Texas had reached their effective dates, each with its own definitions and duties of care. Whichever side one favors, the dilemma underneath is the one every data steward already knows in miniature: one canonical record that everyone reconciles against, or many partial shadows that no one can.

Descend from the statute to the server room and the same problem is waiting, described by the people who manage it in almost exactly these terms. This year's enterprise commentary keeps returning to one theme: the hard part of AI is no longer the model but the record feeding it. Proofpoint's Bikramdeep Singh put it plainly, noting that most organizations cannot see which AI tools their employees use or what data is handed to them, so any serious AI security has to begin with data security. His firm's 2026 survey reports that a majority of Indian organizations have already suffered an AI-related security incident even as the country leads the world in adoption. That is the lineage gap rendered as a breach statistic. You cannot govern what you cannot trace, and these are vendor voices with something to sell, yet the convergence is telling: the practitioners closest to deployment have stopped talking about capability and started talking about provenance. Regulation is beginning to follow. The Federal Trade Commission's pursuit of "AI washing" now treats every claim a company makes about its system's accuracy as something that must be backed by documented evidence, held to the rigor of a financial disclosure. That is the governance translation of the whole argument. A claim you cannot trace to a durable, primary record is not a fact but an assertion waiting to be substantiated, whether it concerns a marketing boast about model accuracy or the name of the person who founded a holiday.

The human responsibility

The remedy, at both scales, turns out to be the same move, and it is worth seeing that the two problems resolve together rather than separately. The journalist who came closest to untangling the Kirton strand — Tyler Wilde, writing for PC Gamer — closed his account by noting that his own article would eventually be ingested by the very chatbots that had been getting the origin wrong, and that their outputs might then improve. He is right about the mechanism, if perhaps too hopeful about the result, because crediting one claimant is not the same as reconciling three. That is the whole predicament and the shape of its only resolution: the record self-corrects only when a human goes upstream and does the primary-source work: locating the claimants, weighing them against each other, and publishing the adjudication somewhere durable enough to be ingested in turn. There is no algorithmic shortcut back to a source that has been deleted, and none through a thicket of sources that contradict one another. Someone has to reconstruct the picture deliberately, contradictions and all, and feed it back in. This is exactly what the governance fight is groping toward when it argues over safe harbors and substantiation and a single authoritative standard: not a smarter model, but a durable, checkable baseline and a named party accountable for keeping it true. The intelligence we are asked to appreciate every July 16 cannot repair its own lineage, at the scale of a holiday or the scale of a federal statute. It can only inherit whatever record we take the trouble to leave it. And that makes the maintenance of that record, unglamorous as it is, the part of the work no system has yet figured out how to do for us.

My blog

The Founder AI Forgot

An innocent question

A record that forked

The same problem, one governance layer up

The human responsibility

Further Reading