A writer asked GPT-5 to pick stocks that would make him "Lambo money" in six months, and the chatbot lost him $23 while teaching us exactly what AI agents can and can't do with your money.
The Summary
- In September 2025, a Fast Company writer asked ChatGPT's GPT-5 model to pick five aggressive stocks that would deliver Lamborghini-level returns in six months
- The bot spent 8 minutes researching 98 sources before making picks, then proceeded to lose the writer $23 over the test period
- The stunt got enough attention that Jason Alexander (yes, that one) invited him on his podcast to discuss AI investing on his birthday
The Signal
ChatGPT didn't phone it in. When asked for moonshot stock picks, the GPT-5 model churned through 98 different sources over 8 minutes. This wasn't the canned "invest in index funds" response you'd expect from a liability-averse AI. It actually did the homework, processed real market data, and delivered specific picks.
Then it lost money. Not catastrophically, $23 is coffee money, but it's the precision of the failure that matters here. This is what happens when you stress-test agent capabilities in the real world with actual stakes.
"The bot researched like a junior analyst and performed like one too."
The cultural reach of the experiment tells you something about where we are with AI agents. A Seinfeld actor doesn't usually cold-email writers about their ChatGPT stock picks. But Jason Alexander read the piece and wanted to discuss AI investing on his podcast. The writer initially thought the invitation was a scam (basement door at Warner Brothers, really?), which is its own signal about how weird the AI moment has made everything.
The experiment's timing matters. September 2025 was when GPT-5 was still fresh, when people were still genuinely curious whether these models could outthink markets. Six months later, we have an answer: they can research, they can synthesize, they can make calls. But making calls and making money are different games.
What this test actually measured:
- Information processing speed (8 minutes, 98 sources)
- Decision-making under explicit instructions for risk ("aggressive, somewhat crazy")
- Real-world financial outcome over a defined period (six months, -$23)
The negative return isn't the story. Plenty of human stock pickers lost more than $23 in that window. The story is that we now have public, documented examples of what frontier AI models do when you give them financial agency. They don't hallucinate fake tickers or refuse to engage. They do the work, make picks, and perform about like you'd expect from something with no skin in the game.
The Implication
If you're building an AI agent that handles money, this is your baseline. GPT-5 will research, decide, and lose your lunch money with equal competence. The models are good enough to look professional and bad enough to cost you. That gap is where the real product work lives.
For anyone actually considering AI-assisted investing: note that the writer asked for crazy picks and got exactly that. If you want an agent to manage your portfolio, the quality of your prompt matters more than the quality of the model. Garbage instructions, garbage returns.