AI will not hand you an edge. It will hand you more ideas, more code and more backtests per hour than ever before. Most of them will be wrong, and AI makes the wrong ones look convincing.

That is the common thread in what hedge funds, asset managers, regulators and researchers have published over the last two years. Professional firms use AI almost everywhere, yet the evidence that it produces lasting alpha (returns beyond what the market or known risk factors pay) is thin. Individuals now have chatbots, AI scanners and agents that can place orders. But the base rates for retail trading have not changed, and new failure modes have arrived: invented numbers, persuasive bad advice and backtests that "know" the future.

This post is a practical guide for both audiences. It summarizes what the professionals actually do with AI, what the evidence says, where it breaks, and how an individual can use the same tools without fooling themselves.

Educational content only. Nothing here is investment advice or a recommendation to buy, sell or hold anything. See the risk disclaimer at the end.

1. What professionals actually do with AI

Adoption is near universal. An AIMA survey in September 2025 found 95% of hedge fund managers use generative AI, up from 86% in 2023, and 58% expect to use more of it in the investment process itself. A 2024 Bank of England and FCA survey found 75% of UK financial firms already use AI. But only 2% of use cases make fully autonomous decisions, and 46% of firms admitted only a "partial understanding" of the AI they use.

The pattern across firms: AI does the reading, coding and first drafts, and humans keep the decisions and the risk limits.

Firm (as published or reported) What they do with AI Note
Man Group (AlphaGPT) Agentic system that proposes signal ideas, writes the code and backtests them Several dozen signals approved for live trading as of mid-2025, behind human checkpoints. Man itself warns it can find "patterns that appear significant but represent statistical artefacts".
Two Sigma LLMs widen the top of the research funnel Their framing: the bottleneck moved from "we need more ideas" to "we need to evaluate ideas faster".
Bridgewater (AIA Labs fund) Machine learning generates the investment logic itself, launched in 2024 Humans stay in risk management, data and execution. Reported to have gained about 11% in 2025.
AQR Machine learning signals, NLP on text Bloomberg reported in 2025 that ML powers about a fifth of the signals in its flagship multi-strategy fund. Cliff Asness had earlier called AI "still just statistics".
BlackRock Systematic LLMs turn earnings calls, filings and news into measurable signals Tests thousands of datasets a year, with "human judgment" kept central.
Citadel Research productivity Ken Griffin said in October 2025 that GenAI "falls short" for finding alpha. By May 2026 he said agents do in hours or days work that took finance PhDs weeks or months.

Two lessons stand out. First, the biggest gain is research throughput: reading, summarizing, coding, replicating papers. Second, the firms that use AI most aggressively also talk most about guardrails. More ideas per day means more false discoveries per day, unless validation keeps up.

2. Where AI helps and where it fails

Stage Where AI helps Where it fails
Idea generation Brainstorms hypotheses, reads papers, finds analogies across markets Generates plausible stories for random patterns. Ideas are cheap; true ones are not.
Data Turns text (filings, calls, news, social posts) into numbers; cleans and joins data Hallucinates figures when it answers from memory instead of a data source
Coding and backtesting Writes and debugs research code in minutes The code can quietly implement a different idea than the one described; look-ahead bugs; thousands of variants tested until one looks good
Forecasting returns Reacts fast to news and sentiment Returns are mostly noise and the market adapts. Edges decay as everyone uses the same models.
Execution Smarter order routing and timing (a long-standing ML use) A bug or hallucinated order executes in milliseconds, without review
Risk and monitoring Summarizes exposures, flags anomalies, explains a portfolio in plain language Correlated behavior: if many firms use similar models, they sell together in a crisis
Personal use Learning, summarizing, journaling, checking your own reasoning Persuasion, overconfidence, and confirmation of what you already believe

The underlying problem was stated years before LLMs. Marcos Lopez de Prado, in The 10 Reasons Most Machine Learning Funds Fail: "When misused, ML algorithms will confuse statistical flukes with patterns. This fact, combined with the low signal-to-noise ratio that characterizes finance, all but ensures that careless users will produce false discoveries at an ever-greater speed." AI agents make "ever-greater speed" literal.

3. What the evidence says about AI performance

AI hedge funds. A 2026 NBER paper by Chen, Sialm and Xu found that funds using AI beat other hedge funds by about 0.5% a month before 2017. After that, the difference became statistically indistinguishable from zero. The Eurekahedge AI Hedge Fund Index returned 9.8% a year from December 2009 to July 2024, against 13.7% for the S&P 500. One useful finding from the NBER paper: AI funds were much less correlated with each other (average 0.02) than other funds (0.12). So "AI" is not one strategy.

LLMs reading news. The best-known paper, by Lopez-Lira and Tang, found that GPT models predict next-day returns from headlines. The headline "90% hit rate" refers to the initial reaction, which cannot be traded. The authors also found that strategy returns decline as LLM adoption rises. Once everyone can do it, the edge gets priced in.

Complex models. Kelly, Malamud and Zhou argued in the Journal of Finance (2024) that large models beat simple ones in return prediction. Stefan Nagel replied in 2025 that the strong result came mostly from a simple volatility-timed momentum strategy that happened to work in that historical sample. Even top academics disagree on whether the complexity is the edge or the disguise.

LLMs trading live. In nof1's Alpha Arena (October and November 2025), six frontier LLMs each got $10,000 to trade crypto for about two weeks. Four of the six lost 42% to 59%, as reported by ForkLog. The organizer's founder offered an explanation: LLMs do not handle numerical time series well, and that was all the context they were given. The LiveTradeBench study (2025) ran 21 LLMs live for 50 days and found that "high LMArena scores do not imply superior trading outcomes". A model that writes great code is not, for that reason, a good trader.

Individuals. Before any AI, the base rates were already harsh. In a study of Taiwanese day traders, Barber, Odean and co-authors estimated that 97% are likely to lose money. In Brazil, 97% of people who day traded futures for at least 300 days lost money and only 0.4% earned more than a bank teller (Chague and co-authors, 2020). AI tools do not change these numbers by default. They may make overtrading easier.

4. Three AI-specific traps

Trap 1: The backtest that knows the future

An LLM trained on data up to 2025 has read the news, the earnings and often the price moves of 2018-2024. Ask it to "trade" 2020 headlines and it may simply remember what happened. This is look-ahead bias built into the model. Researchers have proposed fixes: anonymize company names in the input (Glasserman and Lin), use point-in-time models trained only on data available at each date (He, Manela and co-authors), or measure the bias directly. One 2026 paper found that correcting for it cut in-sample returns by up to 67%.

Practical rule: any LLM backtest over a period inside the model's training data is suspect until proven otherwise. The only clean test is forward: paper trade after the model's knowledge cutoff.

Trap 2: The best of a thousand backtests

If you test enough random strategies, one of them will look brilliant. This is backtest overfitting (selection bias under multiple testing), and Bailey and Lopez de Prado call it a leading reason for investment failures. Their fix is to count every trial and raise the bar accordingly (the Deflated Sharpe Ratio). An AI agent that runs 1,000 variants overnight needs a much higher bar than a human who tested three:

from math import e, sqrt
from statistics import NormalDist

EULER_GAMMA = 0.5772156649
z = NormalDist().inv_cdf


def expected_max_sharpe(n_trials: int, years: float) -> float:
    """Best annual Sharpe you should expect from n_trials zero-skill strategies."""
    if n_trials == 1:
        return 0.0
    sd = 1 / sqrt(years)  # approx. std. error of an annual Sharpe ratio
    return sd * ((1 - EULER_GAMMA) * z(1 - 1 / n_trials)
                 + EULER_GAMMA * z(1 - 1 / (n_trials * e)))


for n in (1, 10, 100, 1_000, 10_000):
    print(f"{n:>6} backtests, 5 years of data -> best Sharpe by luck ~ {expected_max_sharpe(n, 5):.2f}")
#      1 backtests, 5 years of data -> best Sharpe by luck ~ 0.00
#     10 backtests, 5 years of data -> best Sharpe by luck ~ 0.70
#    100 backtests, 5 years of data -> best Sharpe by luck ~ 1.13
#   1000 backtests, 5 years of data -> best Sharpe by luck ~ 1.46
#  10000 backtests, 5 years of data -> best Sharpe by luck ~ 1.73

With five years of data and 1,000 tries, a Sharpe ratio of about 1.5 is what pure luck produces. Log every variant the AI tests, not only the winner.

Trap 3: Confident numbers, persuasive advice

ESMA, the EU markets regulator, warned in 2024: "AI tools for natural language generation are known to "hallucinate" and produce outputs that are factually incorrect even if they sound realistic and accurate." One 2026 test on French CAC 40 companies, reported by UC Strategies, found chatbots answering from memory got almost no financial figures right. With retrieval from official documents, accuracy rose to as much as 80%.

The behavioral side is worse. In a 2026 CESifo experiment, a chatbot raised the share of good fund choices by over 20 percentage points when it promoted the better fund, and cut them by almost 30 points when it promoted a worse one. A SAFE study the same year found that access to LLMs helped people confirm and strengthen beliefs they already held. A chatbot is an excellent assistant for building the case you wanted to make.

5. Systemic risk: when everyone uses the same brain

Regulators are less worried about one bad AI trader than about many similar ones.

  • Herding. The IMF (October 2024) found that some AI-driven ETFs traded much more during the March 2020 turmoil, a sign of possible herd-like selling under stress. The Bank of England warned in 2025 of firms "taking increasingly correlated positions and acting in a similar way during a stress, thereby amplifying shocks".
  • Concentration. A few model and cloud providers sit underneath much of the industry. If one fails or changes behavior, many firms are affected at once.
  • Tacit collusion. A 2025 NBER paper by Dou, Goldstein and Ji showed that reinforcement-learning traders in simulated markets learned to keep prices away from competitive levels "without agreement, communication, or intent".
  • Scale. AI in algorithmic trading is growing fast. The IMF found AI content in algo-trading patent filings rose from 19% in 2017 to over 50% each year since 2020.

For an individual, the practical lesson is simple: a crowded AI trade can reverse sharply, and liquidity disappears when everyone runs the same exit.

6. Personal trading: a practical playbook

The new retail toolset is real. Robinhood Cortex scans the market from a plain-language prompt. Interactive Brokers' Ask IBKR answers questions about your portfolio. Public.com turns a prompt into a custom index. Alpaca's MCP server and eToro's agent sub-accounts let AI agents place orders through an API. Usage is growing: in an eToro survey, 30% of US retail investors said they used AI tools to choose or adjust investments in 2025, up 75% in a year. And 38% gave as a reason that the AI "will make better decisions than me".

That last number is the problem. Use AI where it is strong and keep it away from where it is weak:

Good uses Risky uses
Explaining concepts, products and jargon Asking "what should I buy?" and acting on the answer
Summarizing filings and earnings calls, with the source document attached Asking for prices, ratios or results from the model's memory
Writing backtest code you can read, run and check on real data Letting the chatbot "run" the backtest and report results
Stress-testing your thesis: "argue the opposite case" Asking it to confirm your thesis
Keeping a trading journal and reviewing your own mistakes High-frequency or leveraged trading driven by AI signals
Checking position size, concentration and worst-case loss Giving an agent unrestricted API keys to a funded account

Rules that cover most situations:

  1. Data from sources, reasoning from the model. Paste in the filing or connect a real data feed. Never trust a number the model "remembers".
  2. Forward test, do not backtest only. Paper trade any AI-assisted strategy for months, after the model's knowledge cutoff, before risking money.
  3. Count your attempts. If you or an agent tried 200 versions, the best one is mostly luck (section 4).
  4. Limit the agent. Use a separate sub-account, small funds, scoped API keys, hard loss limits and order confirmation. Alpaca's own guidance: review and confirm orders on the dashboard.
  5. Ask for the opposite case every time. Make the model argue against your trade before you place it.
  6. Benchmark against doing nothing. A low-cost index fund is the bar. Most active strategies, human or AI, do not clear it after costs.
  7. Know your protection. In the UK, FCA research reported in 2026 found 44% of investors surveyed wrongly believe AI financial information is regulated. A general chatbot is not a regulated adviser, and there is usually no compensation scheme if it is wrong.

Red flags: AI-powered scams

The SEC, NASAA and FINRA issued a joint investor alert in January 2024 on AI-themed fraud, and the CFTC warned the same month that "AI technology can't predict the future or sudden market changes". The SEC has also fined advisers for AI washing, claiming AI capabilities they did not have (Delphia and Global Predictions settled for a combined $400,000 in March 2024). Walk away from:

  • "Guaranteed" or "can't lose" AI trading returns, or fixed monthly returns. The CFTC cites a bot scheme that promised at least 10% a month and took over $1.7 billion in bitcoin.
  • AI bots sold by influencers, strangers in chat groups, or "mentors" who contact you first
  • Deepfake videos of famous investors or executives endorsing a platform
  • Platforms that are not registered with your national regulator
  • Pressure to deposit more to "unlock" withdrawals

Same word, different meaning

Term Meaning A Meaning B
AI-powered Machine learning models genuinely drive decisions Marketing label on a rules-based tool, or plain AI washing
Agent An AI system that plans steps and calls tools, including placing orders A broker or a person acting on your behalf
Alpha Return beyond market and known factor exposure, after costs Informal "returns" or "a good idea"
Backtest A simulation on past data with realistic costs and no look-ahead A curve-fit chart that only shows the winning variant
Signal A tested, quantified predictor used in a portfolio Any tip, alert or chatbot opinion
Accuracy Hit rate of a prediction (direction right or wrong) Profitability, which also depends on the size of wins and losses, costs and risk
Sharpe ratio Excess return per unit of volatility, measured out of sample The best number from many in-sample backtests (see section 4)

Why this matters

AI changes the cost of research, not the rules of markets. A few practical takeaways:

  • Speed is not edge. AI makes ideas, code and backtests nearly free. Validation, risk control and judgment are now the scarce part, for funds and individuals alike.
  • Professionals keep humans in the loop. The firms with the most advanced AI still use checkpoints, limits and human sign-off. An individual giving an agent free rein is taking more risk than they are.
  • Evidence beats demos. AI fund outperformance has faded in the data, and live LLM trading contests have mostly lost money. Treat any spectacular AI trading claim as a red flag until it is independently verified.
  • The model's memory is not data. Use real sources for numbers, and forward tests for strategies.
  • Your behavior is still the main risk. Overtrading, leverage and confirmation bias predate AI. A persuasive assistant can make them worse.

Final Thoughts

The real change of the AI era is not that machines trade better than humans. It is that anyone can now produce a convincing strategy, a polished thesis and an impressive backtest in an afternoon. Most of those will be noise, and AI makes noise look like signal.

The professionals who do well with AI treat it as a fast, tireless research assistant whose every output must be checked. Individuals can do the same: let it read, explain, code and argue the other side, keep the decisions and the risk limits for yourself, and judge every result by how it does in the future, not the past.

Risk disclaimer

This article is for educational and informational purposes only. It is not investment, financial, legal or tax advice, and it does not recommend any security, strategy, product, platform or provider. Trading and investing involve risk, including the loss of all capital invested. Leveraged products and day trading carry especially high risk. Past performance, backtests and simulated results do not guarantee future results. Figures cited come from third-party sources as of their stated dates and may change. Consider seeking advice from a licensed professional who knows your circumstances before making financial decisions.

Sources