Every analytical method works within limits. The expert’s edge is knowing where those limits are. AI hands the novice the method without the limits.
This is no longer a thought experiment. In a March 2026 Investing.com survey of 938 US investors, 51% said they use AI tools regularly or occasionally, and generating trading ideas was among the uses they cited (34%).
The gap between method and limits appears at every horizon:
- Discounted cash flow (DCF), months to years. The arithmetic is easy. The judgment sits in the growth, margin, capital spending and discount-rate assumptions, and in knowing which of them do not fit this business.
- Multiples, weeks to months. A chatbot will readily compare four companies on price-to-earnings. The expert knows why they deserve different multiples: different cycles, balance sheets, accounting or growth quality.
- News and central bank statements, minutes. A model reads the words of a press release or a Fed statement. The expert reads what changed, what was left out and what the market already expected.
In each case the novice’s error is not bad arithmetic. It is a valid method applied where it does not belong.
These examples come from practice, not from a study. The research so far shows that AI-guided decisions move together, not yet why they go wrong.
Those choices also look alike, because the tools look alike. In a 2026 SSRN paper, Steven Edwards had GPT-4o play 100 investor personas, from value to momentum, and asked each one separately to pick 50 US stocks. He then took the 50 stocks the personas named most often and called them the consensus portfolio. Just 25 personas, drawn at random, recovered about 88% of it. The consensus portfolio also had 92% of its stocks in common with the picks from a single neutral prompt, with no persona at all. Edwards concludes that models trained on the same data make correlated picks, whatever the persona.
Similar is not the same as wrong. That consensus portfolio returned 29.5% a year against 15.3% for an S&P 500 fund over the test period, which the author attributes to its tilt toward large technology stocks. The opportunity comes by two routes. When the shared view is wrong, everyone makes the same error at once. When it is right, everyone acts on it at once, and the crowd may carry the price past fair value.
Independent mistakes point in different directions and partly cancel out. Shared mistakes add up and move prices.
Researchers have started to document this risk. A 2026 working paper by Xiaoxue He of Lancaster University, not yet peer reviewed, uses 14 unexpected ChatGPT outages during US trading hours as a natural experiment. When ChatGPT was available, retail trading was more correlated, within stocks and across industry peers, and trades moved prices more: price impact was about 12.5% higher.
The two sides of the trade differed. Retail selling pushed prices too far, then reversed over the next one to five trading days. During outages, that reversal largely disappeared. The author’s reading: A shared low-cost signal may crowd out investors’ own information. Retail buying, by contrast, became more informative with ChatGPT.
The evidence so far covers short horizons. For longer horizons, where the errors would come from valuations and multiples, there is no direct study yet: that part is an expectation, not a finding.
