Is AI Trading Legit? Every 'Beats the Market' Claim We've Tested — The Record*

Not an opinion page — a record. Every AI trading and 'beats-GPT' claim this court has ruled on, with the exact numbers, holds / doesn't hold / not proven.

The Charge

Every week an AI is said to beat the market, beat GPT, or never lose. This is not an opinion page. It is a record — every claim this court has tested, and exactly how each one held up. We tried ourselves first, and Case №002 through №004 are our own bots. We published the losses anyway.

Three verdicts exist: holds, doesn't hold, not proven. We never say scam. We show the subtraction.

The Record

Doesn't hold — the math is not a matter of proof, it is a matter of subtraction.

• [Case №002 — "always gets out in time"](/case/002): 453 trades, −€52.83. Our own bot.
• [Case №003 — "catches the burst first"](/case/003): 79 trades, −€47.84. Our own bot.
• [Case №004 — "one high-conviction pick a day"](/case/004): 135 trades, −€7.13. Our own bot.

Not proven — the claim may be true; the evidence shown does not establish it.

• [Case №005 — "six signals, net green"](/case/005): +€1.11 over 5 days. Below 90 days, a small green number is a coin that landed heads a few times.
• [Case №006 — "80.4% win rate, never a bad month"](/case/006): a win rate is not a profit. 80% with asymmetric sizing can still lose — Case №003 showed exactly that. And "verified" here means verified by the same hand that wrote the number.
• [Case №007 — "our open model beats GPT-5.5"](/case/007): true on two benchmarks, trailing another frontier model on the same set. Beat one opponent, then called it the race.
• [Case №008 — "AI agents beat the market"](/case/008): one agent, one season, above the line. A public forward-test of the category ran 22 model-seasons and found 5 profitable — 22.7%.
• [Case №009 — "+164% returns, one agent +252%"](/case/009): the headline is a backtest — a story about the past told after the ending is known.
• [Case №010 — "give an AI a wallet and it beats the market"](/case/010): the one benchmark that funded frontier models with real capital watched every one of them lose money.

How to Read This

A claim is not a lie because it failed our test. It is not proven — which is where most public AI-performance claims sit until a live, dated, third-party record says otherwise. The method is public: [how to spot a fake track record](/method/how-to-spot-a-fake-ai-trading-track-record), [how to tell if a model really beats GPT](/method/how-to-tell-if-an-ai-model-really-beats-gpt), and [can an AI agent actually make money trading](/method/can-an-ai-agent-actually-make-money-trading).

We have nothing to sell to the accused. We ran ourselves first, and we lost.

* the docket is open. bring receipts.