Give four AI chatbots $1,000 each and ask them to pick stocks. Wait nine weeks. The results say less about which AI is "smartest" and more about how differently these models approach risk.
A widely-followed experiment pitting ChatGPT, Gemini, Claude, and Perplexity against each other as stock pickers has hit its nine-week mark, and the gap between them is striking. ChatGPT's portfolio is up roughly 21%, with one individual stock pick doubling in value. The other three models have not matched that return.
The Strategy Differences Matter More Than the Returns
The most interesting part is not who is winning. It is how each model behaved. ChatGPT initially froze, sitting entirely in cash and refusing to commit to any picks. Once it finally started trading, it went aggressive and got lucky with at least one high-conviction bet that paid off big.
That pattern reveals something important about how these models handle uncertainty. Claude tends toward cautious, diversified suggestions. Perplexity leans on its search capabilities to find recent data before making calls. Gemini falls somewhere in between. ChatGPT's all-or-nothing approach happened to work in a rising market, but that same behavior could be catastrophic in a downturn.
Do Not Hand Your Money to a Chatbot
This needs to be said plainly: none of these models are financial advisors, and a nine-week window with four $1,000 portfolios is not a statistically meaningful sample. One stock doubling in a small portfolio can swing the whole result. Run this same experiment in a bear market and the rankings would almost certainly flip.
What the experiment does show is that AI models have distinct "personalities" when it comes to risk tolerance, and those personalities are baked into their training, not derived from any real financial analysis. ChatGPT did not pick a winner because it understood the company's fundamentals better. It picked a winner because its default behavior, after overcoming initial hesitation, skewed toward concentrated bets.
As entertainment and as a way to observe model behavior differences, experiments like this are genuinely useful. As investment advice, they are worthless. The models are pattern-matching on training data, not reading earnings reports in real time or modeling market dynamics. Treat the results accordingly.