AI trading bots in 2026: real returns vs advertised claims
Grid bots, DCA bots, “AI” EAs — the dashboards all glow green. Here is what independent checks in 2026 actually measured, and the questions that separate a real record from marketing.
Every AI trading bot dashboard glows green. That is the product: a number that makes you fund an account. The question is never whether the number looks good. It is whether the number describes anything anyone actually banked.
Independent checks published in 2026 give a consistent, sobering picture — and it is worth stating plainly, with attribution, because the pattern repeats across sources.
What was actually measured
In one hands-on 2026 review of exchange-built bots, a grid bot earned roughly 3% over two weeks while Bitcoin ranged — then stalled the moment price broke out of range, with no sell side left. A DCA bot closed cycles every few days in moderate volatility, then lost money in a sustained downtrend once every safety order filled at once. Both results are genuine. Neither is an annual return — but both get annualised into triple-digit “APR” figures on marketing pages. A good fortnight is not a yield.
A separate November 2026 check of advertised AI trading claims found that 17 of 18 public performance claims had no record a third party could verify, and 6 were backtests or hypotheticals by the product’s own description. Where trades were public, results were a mixed bag: some AI-labelled vaults beat Bitcoin over the same dates, one fell more than 30% from a peak, and in half the AI trading apps checked, users lost money overall in 30 days.
And there is a structural catch with “AI” backtests specifically: a bot whose decisions require live calls to a language model cannot be backtested at all in the usual sense — the model did not exist for most of the backtest window, its responses vary, and it costs money per call. A ten-year “AI-powered” backtest with a perfect equity curve is not AI history. It is a conventional strategy with AI branding.
The questions that separate a record from marketing
- Can I open the record myself? A public wallet, vault, or exchange-verified account beats any screenshot. Screenshots can be cropped from a cherry-picked window.
- Over what window, in what market? A return without its market conditions is meaningless. Sideways-market grid profits are not trend-market profits.
- Net of everything? Fees, spread, slippage, and — for AI bots — API costs. A gross number is a draft, not a result.
- Is the claim a backtest? If yes, it is a hypothesis, and it needs the full validation treatment: recomputation, baselines, walk-forward folds, leakage checks.
- What happens when the regime changes? Every bot has a market it dies in. If the vendor cannot name it, you are the one who will discover it.
The honest role of AI in trading
None of this means AI is useless in trading. It is genuinely useful in narrow, supervised roles: generating signal ideas that humans review, reading documents for sentiment, writing and testing strategy code faster. What fails, consistently, is the unsupervised money machine — the autopilot that prints returns while you sleep. Markets adapt, edges decay, and execution costs bite. Skepticism is the correct default, and the burden of proof sits entirely on the vendor.
Before you fund any bot, run its claims through the same discipline you would apply to any backtest: recompute from raw predictions, beat the dumb baselines, hold up across folds. If the vendor cannot supply the raw record to recompute from, that absence is itself the answer.
That is the recomputation BuildIT Index performs: the claimed number is ignored, the predictions decide, and the verdict carries an exit code. It cannot tell you a bot will make money. It can tell you whether the record on the page describes the records in the file.
This article is educational, not financial advice. Past backtest results do not predict future returns.