AI Forex Trading Claims: What "AI Signals" Usually Are, and What Regulators Found

Most “AI forex signals” are ordinary rule-based indicators, human-picked trades, or pure marketing with an AI label attached, and in several enforcement cases the AI did not exist at all. Machine learning is genuinely used in professional trading, but mostly for narrow tasks such as execution, risk and data processing, not as a machine that reliably predicts currency moves for retail subscribers.

In short: “AI” in a signal or bot advert is a claim, not a feature you can see. US regulators have brought “AI-washing” cases against firms that described AI they were not using, and the CFTC has warned that AI cannot predict the future or sudden market changes. Searching through many models and reporting the best one can make a system with no edge look brilliant, as the simulation below shows. Ask what the model is, what data it uses, how it was tested out of sample, and where the verified live record is.

The pitch you have probably seen

“Our AI analyses millions of data points per second.” “Neural-network signals with a 94% win rate.” “The first AI financial advisor.” The adverts rarely say what kind of model is used, what it was trained on, or how it performed on data it had never seen. They usually do say how much you can make, and how soon.

What “AI signals” usually turn out to be

What is sold as AI	What it often is
”AI-generated signals”	Standard technical indicators (moving averages, RSI, breakouts) with fixed rules
”AI-powered analysis”	A chatbot or template writing commentary around a human’s trade idea
”Self-learning bot”	A strategy whose settings were optimised on past data, then frozen
”AI trading fund”	In fraud cases, a pool where little or no trading happened

Rule-based systems are not dishonest by themselves. The issue is the label: calling a moving-average crossover “AI” implies an ability to learn and adapt that the product does not have. Rule-based and discretionary signals are compared in manual vs automated signals.

What regulators found: AI-washing cases

SEC: Delphia and Global Predictions (March 2024)

On March 18, 2024, the SEC announced settled charges against two investment advisers, Delphia (USA) Inc. and Global Predictions Inc., for false and misleading statements about their use of artificial intelligence. They agreed to pay $400,000 in total civil penalties: $225,000 from Delphia and $175,000 from Global Predictions, without admitting or denying the findings.

Delphia said it used client data with AI and machine learning in its investment process. The SEC found it did not actually use that data in any AI or machine-learning tool.
Global Predictions called itself the “first regulated AI financial advisor” and advertised “expert AI-driven forecasts”, which the SEC found false and misleading.

Neither firm was a forex signal seller, but these were the SEC’s first cases over misleading AI claims, and they set out the principle: describing AI you do not use is a false statement.

SEC: Rimar Capital (October 2024)

On October 10, 2024, the SEC announced settled charges against Rimar Capital entities, owner Itai Liptz and board member Clifford Boro over claims that the firm used AI to perform automated trading. According to the SEC, they raised about $3.725 million from 45 investors for a purportedly AI-based trading platform, but the firm had no trading application at the time and never had a trading platform. The parties agreed to pay $310,000 in total civil penalties.

SEC: American Bitcoin Academy (February 2024)

In February 2024 the SEC charged Brian Sewell, founder of an online crypto course, and his firm Rockwell Capital Management, alleging he persuaded students to invest about $1.2 million in a hedge fund that would use artificial intelligence and crypto trading strategies. The SEC said the fund never launched. The case settled without admission of the allegations.

CFTC: trading bots that barely traded

The CFTC’s largest bitcoin-related fraud case, against Mirror Trading International and Cornelius Johannes Steynberg, involved a forex trading “bot” that, according to the court’s findings, received only a small fraction of the at least 29,421 bitcoin collected from participants. In September 2026, the CFTC sued Cash FX Group and others, alleging a scheme that raised over $950 million with claims of expert traders, proprietary algorithms and artificial intelligence, while little trading occurred. The Cash FX allegations had not been decided as of November 2026. More detail on both cases is in forex trading bots, honestly.

What regulators tell consumers

CFTC (January 25, 2024): the advisory AI Won’t Turn Trading Bots into Money Machines warns that fraudsters exploit interest in AI to sell trading algorithms and trade signal strategies promising unreasonably high or near-certain returns, that AI “can’t predict the future or sudden market changes”, and that claims of huge returns or 100% win rates are the mark of a scam.
SEC, NASAA and FINRA (January 25, 2024): a joint investor alert warned that bad actors use the popularity and complexity of AI to lure victims, including firms claiming proprietary AI that boosts returns, and that AI tools can fake voices, images and video.
Canadian Securities Administrators (December 5, 2024): CSA Staff Notice and Consultation 11-348 warned against “AI washing” and said AI-related disclosure should avoid vague, unsubstantiated statements and jargon designed to attract investors.

In the US, a firm or person who sells trading advice for pay may also face registration requirements regardless of whether AI is involved; see are forex signal sellers regulated in the US. In June 2025 the SEC withdrew a 2023 proposal on predictive data analytics, but existing anti-fraud and advertising rules still apply to AI claims.

What genuine machine learning in trading looks like

Machine learning is real and widely used in finance. In professional settings it typically:

Processes data, such as classifying news, parsing central bank statements or cleaning price feeds.
Improves execution, for example deciding how to split a large order to reduce market impact.
Manages risk, by estimating volatility, correlations and exposures.
Detects fraud and anomalies in transactions and market activity.

Where academics have tested machine learning for return prediction, the gains are real but modest. Shihao Gu, Bryan Kelly and Dacheng Xiu, in “Empirical Asset Pricing via Machine Learning” (Review of Financial Studies, 2020), found that machine-learning methods improved forecasts of US stock returns compared with traditional regression approaches, but the predictable share of monthly returns remained small. Currencies are an even harder target: since Richard Meese and Kenneth Rogoff’s 1983 study in the Journal of International Economics, which found that standard exchange-rate models did not beat a random walk out of sample, beating simple benchmarks in currency forecasting has been a recognised challenge.

Genuine work also comes with dull details that marketing leaves out: a defined model, documented training and test periods, costs included, out-of-sample results, and performance that degrades over time and needs monitoring.

A worked example: how to make fake AI results from noise

One reason “AI” results can look spectacular is that model-building often means testing many variations and keeping the best. Here is a simulation, computed in Python, of what that does when no strategy has any edge.

Setup

Generate 1,000 strategies. Each has daily returns drawn at random with an average of zero and a standard deviation of 1% (about 15.9% a year). By construction, none can make money on average.
Measure each strategy’s annualised Sharpe ratio (return divided by volatility) over one year of 252 trading days.
Pick the best one, as an optimiser or a marketer would.
Then run the chosen strategy for a fresh year of random data.

Results from one run

The median strategy had a Sharpe ratio of about 0.02, as expected for zero edge.
168 of the 1,000 had a Sharpe ratio above 1, and 25 were above 2.
The best had a Sharpe ratio of 3.22 and a one-year return of about 55%.

Repeating the whole experiment 300 times

The best of each batch of 1,000 averaged a Sharpe ratio of 3.28 in its selection year, and never fell below 2.5. A Sharpe of 3.28 at 15.9% volatility implies an annual return of about 52%.
In the following year, those same “best” strategies averaged a Sharpe ratio of minus 0.01, and 51% of them lost money.

A Sharpe ratio above 3 would be extraordinary for a real fund. Here it appeared every single time, from pure noise, simply by choosing the best of 1,000. Any process that searches widely, whether it is called AI, optimisation or “our proprietary engine”, will produce impressive backtests unless the search is accounted for. The methods for doing that are covered in overfitting and data snooping, and the question of how long a record must be to mean anything in statistical significance of a track record.

Questions to ask about any AI claim

What kind of model is it? “AI” covers everything from a decision rule to a large neural network. A real developer can say which.
What data does it use, and when was it trained? If the answer is vague, so is the product.
What happened out of sample? Ask for results on data the model never saw during development, including costs.
How many versions were tested? The more variants searched, the less the best one’s backtest means.
Where is the live, verified record? A real-money account, verified through the broker or a linked third-party tracker, over a meaningful period.
Where does my money go? If you must send funds to the operator rather than keep them in your own regulated account, stop.

What can go wrong

The AI does not exist. Several enforcement cases involved no working system at all.
The AI is a relabelled indicator sold at a premium.
The results are selected, the best of many tries, and will not persist.
Deepfakes. Regulators warn that AI tools are used to fake endorsements, voices and video of well-known figures.
No accountability. “The algorithm decided” is not an explanation, and an unregistered seller offers no recourse.

The short version

An “AI” label on a forex signal or bot tells you about the marketing, not the method. US regulators have penalised firms for claiming AI they did not use, and the CFTC’s position is blunt: AI cannot predict the future or sudden market changes. Genuine machine learning in trading is mostly unglamorous work on data, execution and risk, and its forecasting gains are modest. Choosing the best of many models makes noise look like genius, so judge any AI claim the same way as any other: by a verified live record, honest out-of-sample testing, and money that stays in your own account.