Is AI Trading Legit? Here's What the Bots Actually Delivered

The question people started typing when the market got loud#

Google's advertising data shows US searches for the exact phrase "is ai trading legit" running at roughly 140 a month in August 2025. By May 2026 the same phrase was typed 3,600 times in a month, and searches for "ai trading" peaked at 33,100 (Google Ads search-volume data retrieved via DataForSEO, 10 September 2026). Volumes have since fallen back, which usually means the advertising wave moved on rather than that anyone answered the question.

It is a fair question and it deserves better than a yes or a no. The label "AI trading" covers a bank's order-routing software, a $120 million exchange-traded fund, a $9-a-month crypto bot and a man in Texas whose bots did not exist. Those things share an acronym and very little else.

So this piece stays narrow. It looks at what the published record shows: one fund with a nine-year audited history, the mathematics of why backtests flatter, and the fraud numbers regulators put in writing.

What an "AI trading bot" is, once you strip the marketing#

Three different products get sold under the same label.

The simplest is rules-based automation, which executes instructions you wrote yourself. Buy if the price drops 3%, sell at a 5% gain. There is no intelligence in it. It is a fast, tireless clerk.

The second is machine learning: software that finds patterns in historical data and forecasts from them. Nobody writes the rules; the model infers them. A backtest is the model's report card on the past, showing what it would have earned on old data. Out-of-sample results are what it earns on data it has never seen, which is a far harder test and the only one worth much.

The third is reinforcement learning, where a model learns by trial and error against a simulated market and is rewarded for profit. That is the frontier, and it sometimes behaves in ways its designers did not plan.

One number follows you around all three: the Sharpe ratio, meaning return above cash divided by volatility. Higher means more reward per unit of stomach-churn. A long-run Sharpe of 1 is genuinely good. Anyone advertising 5 is either running money at a fee you cannot access or making it up.

It is worth knowing what the industry actually uses AI for. The International Organization of Securities Commissions, the global body of market regulators, surveyed the field for its March 2025 report on AI in capital markets. Among broker-dealers, 67% used AI for client communications and 63% for algorithmic trading, mostly execution, meaning slicing a big order into small pieces so it does not move the price against you. Half used it for anti-money-laundering checks. The report also notes something no sales page mentions: heavy AI models add latency, which "may make them inappropriate in many algorithmic trading contexts where speed of execution is especially important."

Finance is using AI to trade more cheaply. That is not the same as using it to know what will go up.

The one long track record you can check yourself#

Bot vendors publish backtests. Regulated funds have to publish audited returns in a standard format, which makes them the only honest scoreboard available to an outsider.

The Amplify AI Powered Equity ETF, ticker AIEQ, launched on 17 October 2017. It picks US shares using an EquBot model running on IBM Watson, which the fund says reads "news, social media, industry and analyst reports, and financial statements on thousands of U.S. companies." It charges 0.75% a year and held $120.4 million on 8 September 2026. Its published returns to 31 August 2026 sit below, next to same-date figures for the SPDR S&P 500 ETF Trust, the largest plain index tracker.

Sources: Amplify ETFs and State Street Global Advisors, standardised month-end performance. Past performance is not a guide to future returns.

The five-year row is where it bites. Compounded, 12.65% a year turns £10,000 into about £18,140. At 4.18%, the same £10,000 becomes roughly £12,270. Nine years in, AIEQ's since-inception return is 10.08% a year, a respectable number in isolation and a poor one beside what the index handed out for almost nothing over the same stretch.

Interpretation, offered as opinion: one fund does not prove that machines cannot pick shares. It does show that an AI badge buys you no entitlement to a better outcome, and that a fee gap of roughly 0.66 percentage points a year works against you the whole time you are waiting to find out.

Why the backtest always looks better than the bank statement#

There is a mathematical reason bot marketing works, and it was published in a maths journal rather than a trading magazine.

In Notices of the American Mathematical Society in May 2014, David Bailey, Jonathan Borwein, Marcos López de Prado and Qiji Zhu showed how easily a strategy with no genuine edge produces a dazzling backtest. Test ten variations of a strategy whose true expected return is zero, keep the winner, and you should expect an in-sample Sharpe ratio of about 1.57. Test 128 variations and the best will show above 2.6. Every one of them is expected to earn nothing in future.

Their practical rule is uncomfortable. With five years of data you can afford roughly 45 independent trials before a Sharpe ratio of 1 becomes meaningless noise. Modern software runs 45 trials before lunch. A performance chart is therefore worth very little unless you know how many configurations were binned to produce it, and the binned ones are never shown.

That does not make machine learning useless in markets. In the Review of Financial Studies in 2020, Shihao Gu, Bryan Kelly and Dacheng Xiu tested machine-learning methods on roughly 30,000 US stocks and 920 predictors from 1957 to 2016. Neural networks produced a value-weighted long-short portfolio with an annualised Sharpe of 1.35, against 0.61 for the conventional regression benchmark. Serious researchers, real result.

Read the conditions, though. Sixty years of data, thousands of stocks, a portfolio that sells short as well as buying, and returns measured before trading costs and tax. A retail app with a monthly subscription and one account is not running that experiment.

The part that is not investing at all#

Some of what markets itself as AI trading is simply theft, and the sums are large enough to matter.

The FBI's Internet Crime Complaint Center logged 1,008,597 complaints and $20.877 billion of losses in 2025. Investment fraud was the largest single category at $8,648,617,756. Crypto-related complaints accounted for $11.366 billion. For the first time the report carried a section on artificial intelligence: more than 22,000 complaints, with adjusted losses above $893 million.

Regulators have been blunt. On 25 January 2024 the Commodity Futures Trading Commission published a customer advisory titled "AI Won't Turn Trading Bots into Money Machines", noting that promoters advertise returns of "tens of thousands of percent" and 100 percent "win" rates. Its flattest line: AI "can't predict the future or sudden market changes".

There is a softer version of the problem. On 18 March 2024 the SEC fined two investment advisers for overstating their use of AI, Delphia $225,000 and Global Predictions $175,000. Global Predictions had called itself the "first regulated AI financial advisor". The agency named the practice AI washing.

And a harder one. On 29 May 2026 the SEC sued Nathan Fuller and Privvy Investments, alleging he raised about $12.3 million from roughly 150 investors by claiming to run AI bots doing high-frequency crypto arbitrage. The complaint says at least $6.2 million went on personal spending and about $5.5 million on Ponzi-style payments to earlier investors, who were sent fabricated account statements. The allegations are untested in court.

What decides your result, and it is rarely the algorithm#

Two things usually settle the outcome before the software gets any credit or blame: cost and leverage.

When the European Securities and Markets Authority restricted contracts for difference in March 2018, it found that 74% to 89% of retail accounts lost money, with average losses between €1,600 and €29,000. Those leveraged products are exactly what many retail trading apps plug into.

Frequency does the rest. Fernando Chague, Rodrigo De-Losso and Bruno Giovannetti tracked everyone who started day trading Brazilian equity futures between 2013 and 2015. Among those who kept at it for more than 300 days, 97% lost money and only 1.1% earned more than the Brazilian minimum wage. Automating a strategy does not change its arithmetic. It removes the friction that used to slow you down, which cuts both ways.

One newer risk deserves a mention, flagged as hypothesis rather than observed fact. Winston Dou, Itay Goldstein and Yan Ji, in NBER working paper 34054 of July 2025, simulated markets full of reinforcement-learning traders and found they "autonomously sustain collusive supra-competitive profits without agreement, communication, or intent". That is a simulation, not a real exchange. It does suggest the more interesting question is shifting from whether your bot can win to what happens once everybody runs one.

Key takeaways#

1. Over the five years to 31 August 2026 the longest-running AI-selected equity ETF returned 4.18% a year, against 12.65% for an S&P 500 tracker charging about one-eighth of the fee.
2. Ten variations of a worthless strategy are expected to yield a backtested Sharpe ratio near 1.57. Without knowing how many versions were tried and discarded, a performance chart tells you nothing.
3. Machine learning does produce real out-of-sample returns in peer-reviewed research, under conditions a retail subscription cannot reproduce.
4. The industry's own heaviest AI use is execution and compliance, not prediction, and regulators note complex models can be too slow where speed matters.
5. Investment fraud cost Americans $8.65 billion in 2025, and both the CFTC and the SEC have named AI-branded trading claims as an active channel for fraud and mis-selling.

Frequently asked questions#

Is AI trading legal? Yes. Automated and algorithmic trading is legal and used daily by regulated firms. What is illegal is misrepresenting what your software does, or taking money for a bot that does not exist.

Do AI trading bots work? It depends on the job. For executing orders and watching positions, they work well. For predicting prices reliably enough to beat a cheap index fund after fees, the published evidence is thin and the one long public track record has lagged.

Why do the backtests look so good? Because they are selected. Run enough variations over the same history and one will look brilliant by luck alone. The published mathematics on backtest overfitting shows how easily this happens.

How do I check whether an AI trading platform is legitimate? Start with your national regulator's register and confirm the firm is authorised. The CFTC advisory also suggests checking how recently the website's domain was registered. Guaranteed returns and advertised win rates are the two clearest warnings.

Are robo-advisers the same thing? No. A robo-adviser usually allocates you to low-cost index funds and rebalances automatically rather than trying to time markets. Different product, much duller risk profile.

Is a free AI trading bot better than a paid one? Free tools generally earn from order flow, spreads or exchange referrals, so the cost sits in your trading rather than in a subscription. Neither pricing model tells you whether the strategy works.

Glossary#

Backtest: a simulation of how a strategy would have performed on historical data. Easy to flatter, and no guarantee of anything ahead.

Out-of-sample: testing on data the model has never seen. The meaningful check on whether a pattern is real.

Overfitting: tuning a model so tightly to past data that it captures noise instead of a durable relationship, then fails in live use.

Sharpe ratio: return above cash divided by volatility, a measure of reward per unit of risk taken.

Execution algorithm: software that breaks a large order into smaller pieces to limit market impact. It decides how to trade, not what to trade.

AI washing: overstating the role of artificial intelligence in a product or investment process. The SEC has brought enforcement actions over it.

Contract for difference (CFD): a leveraged contract on price movement without owning the asset, restricted for retail investors in the EU and UK.

This is reporting, not investment advice. Figures attributed to named sources are facts as those sources published them. Sentences that weigh those figures are interpretation, and are flagged as such.

References#

1. Commodity Futures Trading Commission, Customer Advisory: AI Won't Turn Trading Bots into Money Machines, 25 January 2024 (claimed returns, "can't predict the future" statement, verification advice).
2. Commodity Futures Trading Commission, Release Number 8854-24, 25 January 2024 (advisory title and date).
3. US Securities and Exchange Commission, SEC Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence, 18 March 2024 (Delphia and Global Predictions penalties, AI washing).
4. US Securities and Exchange Commission, Litigation Release No. 26558: Nathan Fuller, 29 May 2026 (amount raised, investor count, alleged misappropriation, charges).
5. Federal Bureau of Investigation, 2025 Internet Crime Report, published April 2026 (complaint volumes, total losses, investment fraud losses, crypto losses, AI section).
6. International Organization of Securities Commissions, Artificial Intelligence in Capital Markets: Use Cases, Risks, and Challenges, CR/01/2025, March 2025 (adoption rates by use case, latency constraint, risk categories).
7. Amplify ETFs, Amplify AI Powered Equity ETF (AIEQ), standardised performance to 31 August 2026, net assets to 8 September 2026 (returns, fee, inception date, methodology).
8. State Street Global Advisors, SPDR S&P 500 ETF Trust (SPY), standardised performance to 31 August 2026 (returns, gross expense ratio, inception date).
9. David H. Bailey, Jonathan M. Borwein, Marcos López de Prado and Qiji Jim Zhu, Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance, Notices of the American Mathematical Society, May 2014 (expected Sharpe ratios from repeated trials, minimum backtest length).
10. Shihao Gu, Bryan Kelly and Dacheng Xiu, Empirical Asset Pricing via Machine Learning, Review of Financial Studies, volume 33, issue 5, 2020 (out-of-sample Sharpe ratios, sample period, predictor set).
11. European Securities and Markets Authority, ESMA agrees to prohibit binary options and restrict CFDs to protect retail investors, 27 March 2018 (74% to 89% loss rate, average losses per client).
12. Fernando Chague, Rodrigo De-Losso and Bruno Giovannetti, Day Trading for a Living?, SSRN working paper, revised June 2020 (share of persistent day traders losing money, earnings comparison).
13. Winston Wei Dou, Itay Goldstein and Yan Ji, AI-Powered Trading, Algorithmic Collusion, and Price Efficiency, NBER Working Paper 34054, July 2025 (simulated collusion among reinforcement-learning traders).
14. Google Ads search-volume data for the United States, retrieved via DataForSEO on 10 September 2026 (monthly search volumes for "is ai trading legit" and "ai trading").