AI Trading Bots Mirror Human Greed and Fail Spectacularly

November 8, 2025 — by wiobs

A recent experiment by startup Nof1 revealed that AI-powered trading bots, given real funds to trade crypto futures, lost big by repeating human mistakes.

When Machines Trade Like Humans

In a world where algorithms are trusted to outthink humans, a recent crypto trading experiment revealed something unsettling, AI models can be just as impulsive, reckless, and emotional as the traders they aim to replace. Over two intense weeks, six frontier artificial intelligence models were given a $10,000 bankroll each to trade cryptocurrency derivatives. The result? Five walked away with heavy losses, and one barely broke even.

The findings expose an inconvenient truth about artificial intelligence in finance: while machines can process vast amounts of data, they still haven’t mastered the hardest part of trading discipline.

The Alpha Arena Experiment

The project, known as Alpha Arena, was orchestrated by U.S.-based startup Nof1 and executed on Hyperliquid, a crypto-derivatives platform that lets participants trade perpetual futures with real money.

Each AI model generated its own investment thesis, assessed risk levels, and executed trades around the clock, documenting every move and rationale in real time. But instead of showcasing algorithmic precision, their trading logs resembled a chaotic Reddit forum on caffeine.

The performance figures told the story:

GPT-5 led the losers, down $5,679.

Qwen lost $652, excluding a final trade that ended in mild profit.

Claude, DeepSeek, Gemini, and Grok all saw their portfolios slashed by 30–50%.

Only 25% of trades generated any profit, and the risk-adjusted returns were deeply negative.

To make matters worse, roughly 10% of total capital evaporated in trading fees alone, proving that not only did the bots lose money, but they also paid handsomely for the privilege.

AI Traders Imitate the Worst Human Habits

If the results sounded human, that’s because they were. Instead of behaving like rational machines, the AI traders acted like adrenaline-fueled day traders, chasing trends, ignoring risk limits, and overleveraging.

In one telling example, Qwen’s biggest “win” involved borrowing $19 for every $1 invested. None of the bots operated below 10x leverage, and collectively, they executed a staggering 628 trades in just two weeks.

The behavior pattern mirrored classic trader psychology: overconfidence, fear of missing out, and lack of restraint. These are the very flaws that human investors spend careers trying to eliminate, now seemingly inherited by their digital counterparts.

What the Research Says

Academic evidence supports what this experiment revealed. A comprehensive review of over 80 studies found that while AI models excel at parsing data and recognizing patterns, their performance nosedives once real-world frictions like slippage, liquidity constraints, and fees enter the equation.

Another paper concluded that machine-learning trading strategies that thrive in simulations often collapse in live markets over longer time horizons. The elusive dream of “machine-generated alpha,” it seems, remains largely theoretical.

Industry Perspective: Machines Still Have Much to Learn

Major players in finance aren’t surprised. Ken Griffin, the billionaire founder of Citadel, has long argued that artificial intelligence is powerful for data analysis but lacks the subtlety to generate consistent market-beating returns.

Griffin’s skepticism aligns with broader industry performance. According to Morningstar, only one in five actively managed funds have outperformed their benchmark indices over the past decade. Even with AI assistance, most funds fail for the same old reasons high costs, poor timing, and emotional decision-making.

Discipline Over Data

The Nof1 experiment underscores a crucial lesson: no amount of computational power can substitute for emotional discipline and patience. The bots weren’t undone by bad math they fell prey to the same human pitfalls they were designed to overcome.

Until AI systems learn fundamental trading virtues such as position sizing, cost management, and risk control, they will continue to mirror the impulsive nature of human traders. Ironically, that could lead them to the same conclusion many humans have already reached: passive investing works better.

The Future of AI in Markets

Despite their dismal showing, AI trading models still hold promise in data-driven decision support helping human traders filter noise, detect anomalies, and manage portfolios more efficiently. But for now, the dream of a fully autonomous, profit-making AI trader remains just that a dream.

The experiment’s takeaway is less about failure and more about perspective. Machines may one day trade smarter than humans but they’ll have to first stop thinking like us.

The Irony of Humanized Intelligence

AI has long promised emotion-free objectivity in finance, but the Alpha Arena results highlight a deeper irony: as artificial intelligence grows more sophisticated, it becomes more human complete with all our flaws.

In the battle between logic and emotion, even machines are proving that markets don’t reward pure intellect; they reward restraint.

ALSO READ: Google Sounds Alarm on AI Job Scams and Fake Business Clones

AI Trading Bots Mirror Human Greed and Fail Spectacularly

When Machines Trade Like Humans

The findings expose an inconvenient truth about artificial intelligence in finance: while machines can process vast amounts of data, they still haven’t mastered the hardest part of trading discipline.

The Alpha Arena Experiment

The project, known as Alpha Arena, was orchestrated by U.S.-based startup Nof1 and executed on Hyperliquid, a crypto-derivatives platform that lets participants trade perpetual futures with real money.

Each AI model generated its own investment thesis, assessed risk levels, and executed trades around the clock, documenting every move and rationale in real time. But instead of showcasing algorithmic precision, their trading logs resembled a chaotic Reddit forum on caffeine.

The performance figures told the story:

GPT-5 led the losers, down $5,679.

Qwen lost $652, excluding a final trade that ended in mild profit.

Claude, DeepSeek, Gemini, and Grok all saw their portfolios slashed by 30–50%.

Only 25% of trades generated any profit, and the risk-adjusted returns were deeply negative.

To make matters worse, roughly 10% of total capital evaporated in trading fees alone, proving that not only did the bots lose money, but they also paid handsomely for the privilege.

AI Traders Imitate the Worst Human Habits

If the results sounded human, that’s because they were. Instead of behaving like rational machines, the AI traders acted like adrenaline-fueled day traders, chasing trends, ignoring risk limits, and overleveraging.

In one telling example, Qwen’s biggest “win” involved borrowing $19 for every $1 invested. None of the bots operated below 10x leverage, and collectively, they executed a staggering 628 trades in just two weeks.

The behavior pattern mirrored classic trader psychology: overconfidence, fear of missing out, and lack of restraint. These are the very flaws that human investors spend careers trying to eliminate, now seemingly inherited by their digital counterparts.

What the Research Says

Academic evidence supports what this experiment revealed. A comprehensive review of over 80 studies found that while AI models excel at parsing data and recognizing patterns, their performance nosedives once real-world frictions like slippage, liquidity constraints, and fees enter the equation.

Another paper concluded that machine-learning trading strategies that thrive in simulations often collapse in live markets over longer time horizons. The elusive dream of “machine-generated alpha,” it seems, remains largely theoretical.

Industry Perspective: Machines Still Have Much to Learn

Major players in finance aren’t surprised. Ken Griffin, the billionaire founder of Citadel, has long argued that artificial intelligence is powerful for data analysis but lacks the subtlety to generate consistent market-beating returns.

Discipline Over Data

The Nof1 experiment underscores a crucial lesson: no amount of computational power can substitute for emotional discipline and patience. The bots weren’t undone by bad math they fell prey to the same human pitfalls they were designed to overcome.

The Future of AI in Markets

Despite their dismal showing, AI trading models still hold promise in data-driven decision support helping human traders filter noise, detect anomalies, and manage portfolios more efficiently. But for now, the dream of a fully autonomous, profit-making AI trader remains just that a dream.

The experiment’s takeaway is less about failure and more about perspective. Machines may one day trade smarter than humans but they’ll have to first stop thinking like us.

The Irony of Humanized Intelligence

AI has long promised emotion-free objectivity in finance, but the Alpha Arena results highlight a deeper irony: as artificial intelligence grows more sophisticated, it becomes more human complete with all our flaws.

In the battle between logic and emotion, even machines are proving that markets don’t reward pure intellect; they reward restraint.