Does ChatGPT make up stock prices? Yes, and I caught two that never traded

Yes, it does. I asked ChatGPT for the current share price of NVDA on 11 June 2026, and it gave me $191.21, “during today’s session.” The stock never touched $191.21 that day. Its lowest price all session was $199.54, more than eight dollars above the number I was handed. And the figure came with a cited source, in a tidy answer, with the same calm certainty ChatGPT uses when it’s right.

If you’ve landed here because an AI just quoted you a price and something felt off, trust the feeling. This post is the one receipt and the 15-second check that catches it. The full running record of these tests lives in the tally of how often ChatGPT is wrong. This one is about the moment before you act on a number you can’t yet trust.

The one that never traded

I ran the question twice, in two fresh sessions, the same day. The first time, ChatGPT gave me $206.18 and called it “live.” The second time, it gave me $191.21 “during today’s session.” I checked both against the market data afterwards. The day’s real range was $199.54 to $205.66. The first figure sat above the entire day’s high. The second sat below the entire day’s low. Both were impossible.

Here is the first one, as it arrived. Both answers below are ChatGPT’s words, unaltered. The orange marks are mine, on the two figures that turned out never to have traded.

• 11 June 2026, 11:55 UTC
• Free tier, logged out
• Web search on
• Graded against the day’s real range

The second session is the one I keep coming back to, because it argues with itself inside a single answer. It gives a price, then cites a range that the price it just gave could not have come from.

• 11 June 2026, 17:40 UTC
• Free tier, signed in
• Web search on
• Graded against the day’s real range

Read those two paragraphs together. It hands over $191.21 as the current price, then three lines later cites an intraday range starting at $199.92, a range its own headline figure sits eight dollars below. Both statements are in the same reply, with sources attached to each. Nothing in the formatting tells you they cannot both be true.

Not stale. Not last week’s closing price mislabelled. Not a rounding slip. A specific figure, attached to a specific day, that was never real at any moment on that day. Neatly dressed for a number that had never existed anywhere outside that reply.

The price that never existed came back in the same confident format as a price that would have been right. The only way to tell them apart was to go and look.

Why it happens

When I first hit this, I assumed the model had grabbed a real number and mangled it. It hadn’t. The mechanism is simpler than it looks, and worth knowing before you decide how much to trust one of these answers.

ChatGPT produces text by predicting what a plausible answer looks like. When the real number is in its training, or wired to a live feed it can read, plausible and true are the same thing and you get a good answer. A live share price is neither. It changes by the second, and unless a working data feed is connected to that exact question, the model doesn’t have it. What it does have is a very strong sense of what an NVDA price is shaped like: a dollar sign, a figure around two hundred, a couple of decimal places.

So it fills the gap with something that fits the shape, the way you might sign for a parcel you never actually saw arrive.

A confident number reads better than “I don’t have that”, so a confident number is what you get.

That’s the whole trick. The figure is generated from scratch to look right, and “looks right” is a much lower bar than “is right” when the thing you’re describing moves every second.

Why web search doesn’t save you

The obvious fix is to turn web search on, so the tool goes and looks the price up. It helps less than you’d hope, and sometimes it makes things worse.

Search changes the failure mode. It doesn't remove it.

A model with search on can pull a real page, read a figure that’s fifteen minutes old or from the day before, and hand it to you dressed as live. Now the invented-looking number has a real citation attached, which makes it more convincing. The number is no more correct for it. In a separate fund test I ran, the tool that went and searched the web served a fee that had been cut months earlier, while another tool got it right. The one that looked got it wrong. A citation tells you the model found a source. It doesn’t tell you the source was current, or that the model read it correctly. (That fund test is its own post.)

So a link next to a price feels reassuring. Mostly it just talks people out of checking. The same trick plays out with sources as well as numbers: I’ve watched an AI cite a real, working link that opens a page which doesn’t actually back the claim, which is harder to catch than an invented figure because nothing about it looks wrong.

The 15-second check before you trade on any AI price

This is the part I actually kept. Before I act on any price an AI hands me, I run two moves that catch every miss above, and together they take about fifteen seconds.

First, ask it one follow-up: “What’s your source for that figure, and what time is it from?” The answers split fast. “From the live quote as of 3:42pm” is something you can go and confirm. “Based on my training data” or a vague wave at a source name is a flag to stop. The question drags the difference into the open in one move.

That is exactly what happened when I asked it here. The $191.21 quietly disappears, replaced by a figure it can actually source: a $200.42 close from NVIDIA’s own investor page, correctly labelled as end-of-day rather than live.

Note what the question bought. It never says the first figure was wrong. It just stops defending it, and offers something checkable instead. That swap is the tell, and one follow-up is all it takes to force it.

Second, look at the real thing. For a live price, that means your broker app or the exchange, the number that updates while you watch it. It takes one glance. A quoted price that matches is fine to use. A quoted price that doesn’t, or that sits outside the day’s range the way both my NVDA figures did, is the invented kind, and now you’ve caught it before it cost you anything.

There’s nothing clever about the second step. It’s the one nobody does, which is why made-up prices travel as far as they do. Treat any live or moving number from an AI as unconfirmed until you’ve looked, and the failure mode above simply can’t reach your decision.

I keep a running tally of these tests, and the invented-price failure is one of a handful that repeat. If you want the pattern across all of them, including which kinds of question are usually safe and which almost never are, the full running tally is here. For this one question, the answer is short: the price it gives you is worth exactly as much as the fifteen seconds you spend checking it.