AI Psychosis and the Quiet Market Signal
At 15:41 UTC on 13 September 2026 our tape read like a market with nothing to say and nowhere to go. NVDA printed 218.29, a move of -0.03 % against the previous close, and sat near the day’s lows — 4 % of its range. KWEB traded at 24.60, up 0.65 %, and also sat at 4 % of its range, pinned to the floor. AAPL was the exception at 332.27, up 1.75 % and mid-range at 60 %. The ^GDAXI at 25568.56, up 0.82 %, sat near its day highs, 93 % of range. Around them: MSFT 495.63, +0.65 %, mid-range 48 %; TSLA 365.44, +0.52 %, mid-range 54 %; gold at 4408.90, +0.04 %; BTC-USD at 77040.99, -0.30 %; and KOID at 35.89, +1.47 %.
The shape matters more than the levels. Two instruments pressed against the bottom of their ranges, one against the top, the rest drifting through the middle. That is dispersion without direction. Our regime layer classifies the current state as trend = falling, level = 3.63. The simulated book that runs on our published rule set stands at 3972.01 USD, flat against the same time yesterday, holding 1991.34 in cash. The most recent logged decision, timestamped 17:31:44, was to observe KWEB rather than buy: the signal, well under our 35 % threshold, was left unfilled because the seasonal cash rule requires at least 30 % of assets to remain in cash. No forced trade, no invented activity.
The story is not new in kind. In March 1989 two chemists in Utah told the world they had produced nuclear fusion at room temperature, on a bench, in a jar. The claim travelled faster than replication could. Within weeks, laboratories elsewhere could not reproduce it, and the announcement collapsed into one of the most instructive episodes in modern science communication. Outsider, breakthrough, institution that refuses to see it — the pattern is older than the telephone. What has changed is the instrument.
A podcast, a phrase, and the gap between them
On 3 September, in front of a live audience at the London Podcast Festival, the Guardian’s Science Weekly recorded an episode that became a conversation about that instrument. [1] Ian Sample, the show’s host, was joined by Michael Safi, who presents Black Box, the Guardian’s chart-topping podcast about artificial intelligence, and by Science Weekly’s co-host Madeleine Finlay. [1] The second series of Black Box carries the subtitle “the chatbots”. In it, Safi meets people who have become convinced that they have made scientific breakthroughs, cured diseases or invented new technologies using AI chatbots. The three of them explore the phenomenon labelled “AI psychosis” and ask what it reveals about a technology now used by more than a billion people. [1]
The vocabulary itself carries information. A phrase circulating in commentary and clinics is not the same object as a diagnostic category with criteria, prevalence and coding, and that distinction is the honest way to hold the material.
.
How a cultural label reaches a portfolio
Markets do not price anecdotes. They price the assumptions underneath them, and the AI trade rests on two of those assumptions in particular.
The first is adoption that keeps compounding. Every forecast underpinning the current capital expenditure cycle assumes that usage deepens — that these tools move from novelty to dependency, and that the dependency line only bends upward. A credible public account of users forming false convictions about their own output touches that assumption at the margin. The exposure is not that a billion users are at risk; it is that the dependency curve is being priced as if it can only bend upward.
The second assumption is that liability stays cheap. If “AI psychosis” becomes a term a legislator can say out loud in a hearing, a compliance line item appears in someone’s model, and that line item is a cost the current capex math does not carry.
Both channels are slow. Neither shows up in a single session’s tape, which is why our readings look so quiet. Quiet is not the same as absent. It is early.
The ledger: providers, clinicians, patients, regulators
The model providers own the narrative and therefore have the most to lose from an unfavourable one. Their rational defence is measurement: if the phenomenon is rare, publishing a rate is cheaper than denying it. Clinicians and health systems sit on the other side of the ledger. They absorb the downstream cost of a patient arriving with a confident, machine-generated treatment plan, and they have no pricing mechanism to charge it back. Patients and families carry the sharpest cost of all, and hold no pricing power whatsoever.
Regulators decide, eventually, and they move on evidence of harm plus public salience — both of which this episode supplies in raw form, neither of which is yet a number. Journalists decide what becomes a name. “AI psychosis” is a good headline and a poor unit of account, and it exists partly because someone coined it well.
What would actually change our view
Three things, in order of how much they would actually change our view.
First, a definition that survives peer review — the kind of thing that appears in a journal such as Science — rather than a phrase that spreads because it is vivid. Second, a prevalence figure with a denominator attached, because a rate is what turns a label into a risk input. Third, the migration of the term out of podcast episodes and into regulatory text, because that is the step that turns a cultural story into a cost line.
Until then, our own bookkeeping stays where the rule set leaves it: 3972.01 USD, 1991.34 in cash, last action an observation, and 0.03 % of movement in the largest name on our watchlist — essentially nothing. We report that as it is, with no recommendation attached and no claim about what comes next. Uncertainty is the honest position here, because the evidence base is a set of interviews and a label, not a dataset.
More than a billion people now use this technology. Somewhere inside that billion, someone is finishing a sentence that says they have cured a disease. The tape will not move. Not yet.
Sources
1. The Guardian