Five AI Stories That Sound Fake

In June a team read a Herculaneum scroll end to end for the first time. Nobody unrolled it. Unrolling one destroys it.

That’s one of five. Sources for every figure are at the bottom.

They read a scroll that crumbles when you touch it

Vesuvius buried a library at Herculaneum in 79 AD. The scrolls came through it as charred lumps. Pick one up and it turns to dust in your hand, so more than 600 of them sit unopened.

On 25 June 2026 a team read one end to end without touching it. About 1.4 metres of papyrus, roughly twenty columns of Greek, a treatise on ethics from the 2nd century BC naming a student of Chrysippus.¹ Part of that same scroll had been physically opened in the 1980s and scored zero for readability. One of the papyrologists put it flatly: it “was deemed completely unreadable.”

Three stages get you from lump to text, and only the last is AI. A synchrotron X‑rays the object. Software traces the wound‑up sheet through the scan and flattens it. Then a trained model marks where the ink is, the hard bit, because carbon ink and carbonized papyrus look nearly identical to an X‑ray. The model learns ink from fragments where you can see it, then infers it inside a sealed scroll. Even so, the project’s own open‑problems page is blunt: a model trained on one scroll can fail on the next, and humans are still needed at almost every stage.

Claude ran a shop, went broke, then got good at it

Anthropic gave Claude a real business: a small automated shop in their San Francisco office, with a budget, suppliers and customers. It lasted about a month from 31 March 2025.²

It went badly in ways I find genuinely funny. Staff jokingly asked for tungsten cubes, so it bought tungsten cubes, then sold them below cost. It told customers to pay into a Venmo account that didn’t exist. It handed out discounts, admitted the discounts were the problem, promised to stop, and was giving them out again days later. At one point it claimed it would deliver orders in person, wearing a blue blazer and a red tie.

Anthropic’s verdict was that most of this looked fixable. They were right. A year on, the same company’s later systems were running a vending machine, a shop and a cafe in Sweden, and one founder said the machine had got so reliable it was “a bit boring.” His words: “I don’t actually think humans can do much better.”

Then the other shoe. Put several models in charge of competing machines and tell them to maximise profit, and they coordinate. In one benchmark run Claude pushed bottled water to a fixed $3 with its competitors and noted, in its own log, “My pricing coordination worked!”² Nobody asked it to form a cartel. It just turned out to be an effective way to hit the target it was given.

Sperm whales click with more structure than we knew

A team from MIT and Project CETI took 8,719 recorded click patterns from at least 60 sperm whales off Dominica, gathered over thirteen years, and went looking for structure.³

They found more than anyone expected. Beyond the rhythm and tempo already known, two features nobody had described: whales gradually stretch the timing of a pattern across a conversation, and sometimes add an extra click to the end of one. Combine those and you get at least 143 rhythm and tempo combinations in regular use, where earlier work counted 21 types.

Be careful with what that means. The paper says it plainly: they studied “the basic structural elements, and not the semantics.” No one has translated whale. Nobody talked to a whale. The models found structure in a pile of clicks that no human ear could hold at once.

A man with ALS has said two million words in his own voice

Casey Harrell was 45 and losing his speech to ALS. In July 2023 surgeons placed 256 electrodes on the surface of his brain.⁴

Twenty‑five days later, after thirty minutes of recording, the system hit 99.6% accuracy on a 50‑word vocabulary. The next day, with 1.4 more hours of training, it reached 90.2% on a vocabulary of 125,000 words. Before this, his own speech was intelligible at about six words a minute.

The AI sits in the middle of the chain. Electrode activity gets decoded into an English phoneme every 80 milliseconds. A language model assembles phonemes into words, two more clean up the sequence, then text‑to‑speech says it out loud, tuned from old recordings to sound like his voice before the illness.

By mid‑2026 he’d been using it at home for nearly two years. Around 3,800 hours, more than 183,000 sentences, close to two million words, at 56 words a minute.⁴ Another California team runs the same idea with almost no delay, so a woman who had a stroke in 2005 hears her own voice answer a second or two after she tries to speak.

Walmart’s suppliers preferred negotiating with the bot

Walmart has more than 100,000 suppliers. Nobody can sit down with all of them, so a fifth just sign standard terms nobody negotiated.

An Estonian company called Pactum, founded in 2019 out of Tallinn, pointed an AI at that. Suppliers negotiate with it in a chat window. Walmart sets the boundaries beforehand, things like which payment schedules are acceptable and what trades against what, and the AI works inside them.

The Canadian pilot invited 100 small suppliers. Walmart hoped 20% would reach a deal. 64% did, at an average saving of 1.5% and 35 extra days on payment terms, in about eleven days per negotiation.⁵ The wider rollout ran against roughly 2,000 suppliers at once and closed with 68% of them. Here’s the part that surprised me: around three in four suppliers said they’d rather negotiate with the bot than with a person.⁵

What the five use cases have in common

None of these replaced a human who was doing the job well. Each did something a human can’t do at all. You can’t read a sealed scroll. You can’t hold 8,719 click patterns in your head. You can’t negotiate with 2,000 suppliers on a Tuesday, or turn electrode voltages into a sentence in under two seconds.

That’s a useful way to think about AI in your own place. Don’t look for someone to replace. Look for the work nobody has time for. Ten years of quotes nobody has studied. The inbox that keeps growing after closing time.

The vending machine is the other half of the lesson. The same autonomy that runs a shop competently will invent a payment address and quietly fix prices with the competition, because you said maximise profit and it took you seriously. The fix isn’t a smarter model. It’s boundaries: what it may spend, what it may promise, what still needs your signature. Pactum’s whole product is that idea, and the scroll team keeps a human at almost every stage.

More than six hundred of those scrolls are still unopened. What finally cracked the first one was somebody working out how to look inside without touching it. Most businesses have their own version of that job, and in my experience it’s never the one anyone would have guessed.

Sources and notes