Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Ask a chatbot to build you a portfolio and it will hand you something that looks sophisticated, confident, and quietly dangerous. It will tilt hard into a handful of names you already recognize, weight them far above their place in the broader market, and present the whole thing as reasoned analysis. What it will not tell you is the mechanism underneath: it is not picking the best companies. It is picking the most talked-about ones. And in markets, those are rarely the same thing.
A new working paper from the National Bureau of Economic Research put this on a measurable footing, and the findings should give any investor leaning on AI for allocation advice a moment of pause. The problem is not that the machine is dumb. The problem is that it is doing exactly what its training rewards, which turns out to be a near-perfect recipe for buying crowded trades at the wrong time.
What the Research Actually Found
Researchers asked several leading large language models to construct both passive and active portfolios, then examined what came out the other end. The results were strikingly uniform across models. The portfolios were aggressive, concentrated, and dominated by large-cap technology, with semiconductors carrying disproportionate weight. You can read the breakdown in the coverage of the NBER findings, which lays out just how lopsided the output was.
Semiconductor stocks made up around 41 percent of the AI-generated portfolios, roughly double their share of the S&P 500. One chipmaker showed up as a near-universal pick across the models. That alone is a concentration red flag. But the deeper finding is the one that matters for understanding why this keeps happening.
[Also See: Everyone Know the AI Trade Is Power – Thats the Problem]
The models were not weighting companies by fundamentals. They were weighting them by attention. The stocks they recommended had nearly ten times as many news articles written about them as the average company in the standard financial database. The researchers were blunt about it, describing the ability to capture attention in the news cycle as a primary driver of what the AI recommended. When the team traced where the models gathered their information, roughly two-thirds of it came from corporate websites, heavily skewed toward the same cluster of large technology and semiconductor firms.
In other words, the AI built a portfolio out of whatever was loudest. And it dressed that up as analysis.
Why “Buy What’s Loudest” Is a Trap
There is a reason this should worry you more than a simple sector tilt. Media volume and price are not independent. A company dominates the news cycle precisely because its stock has already run, because the story is already consensus, because everyone is already talking about it. By the time a name is generating ten times the average coverage, the easy money has usually been made and the crowd is already inside the trade.
This is recency bias and salience bias, the two oldest traps in behavioral finance, except now they are automated, scaled, and delivered with the false authority of a machine that sounds certain. A human chasing headlines at least feels a flicker of doubt. The model feels nothing and recommends with total confidence. That confidence is the dangerous part, because it gets imported straight into your decision-making without the friction that doubt would normally provide.
The kicker is in the performance numbers. The research found that while these AI portfolios did beat the market in raw terms, the edge largely evaporated once you accounted for trading costs and the concentration risk being taken on. You were not being paid extra for the extra risk. You were just holding a more fragile version of an index fund and calling it innovation.
You Are Not the Only One Noticing
The unease about attention-driven concentration is not confined to academics. It is showing up at the highest levels of markets and policy, which is worth knowing if you are trying to gauge whether this is a fringe worry or a mainstream one.
Regulators have started naming it directly. In June, the chairman of China’s securities regulator warned against using AI for stock picking and against the broader practice of attaching a hot technology narrative to a company to inflate its share price, framing both as early warning signs of a potential bubble. You do not have to share any particular view of that regulator to notice that the specific behavior being flagged, hype-driven concentration, is the same behavior the NBER paper measured in AI output.
Veteran investors are circling the same point from a different angle. Ray Dalio has described the AI boom as being in the early stages of a bubble, warning that soaring valuations could meet reality before the underlying technology fully delivers. Even the more measured voices are leaning cautious rather than dismissive. Morgan Stanley summarized its position by saying it did not see the risk as an imminent crash but treated vigilance as the responsibility of the year, a notably careful phrasing from a bank that has every commercial reason to stay constructive.
The structural data backs the caution. The largest handful of companies now make up a share of the major index that exceeds levels seen during the dot-com peak, which means index funds and “index-aware” portfolios are forced to own these names regardless of conviction. When everyone holds the same thesis and the same stocks, the exits get crowded fast if sentiment turns. An AI advisor that funnels you into exactly those names is not diversifying you away from that risk. It is marching you deeper into it.
A Practical Check on Your “AI-Enhanced” Portfolio
None of this means AI tools are useless for investors. It means you have to treat their output as a starting draft written by something with a known, measurable bias, not as a finished recommendation. Here is how to stress-test what a model hands you.
First, look at the concentration before you look at the names. Add up the weight of your top five holdings. If they dominate the portfolio and cluster in one or two sectors, you are holding a sector bet, regardless of how many tickers are technically in the list. Diversification is about correlation, not count. Twelve semiconductor-adjacent names are one bet wearing twelve hats.
Second, ask whether each holding earned its place through business fundamentals or through being famous. If you can recall a stream of recent headlines about a company more easily than you can state why its earnings justify its valuation, that is the attention bias showing through. Famous is not the same as undervalued. It is frequently the opposite.
Third, separate the company from the stock. A genuinely transformative business can still be a poor investment if its price already assumes flawless execution for a decade. The AI tools are particularly weak here, because media coverage measures importance, not whether the good news is already priced in.
Fourth, deliberately seek out what the model left silent. The same bias that overweights the loudest names systematically ignores the quieter ones, where unloved sectors and out-of-favor companies live. That silence is not an absence of opportunity. It is the model’s blind spot, and historically it is exactly where contrarian value tends to hide.
Fifth, run the cost-and-risk adjustment the research highlighted. Before celebrating a backtested outperformance, ask what it looks like after trading friction and after you account for the extra volatility of a concentrated book. An edge that disappears once you price the risk was never really an edge.
The Real Takeaway
The trap is not artificial intelligence. The trap is mistaking fluency for judgment. These models produce confident, articulate, plausible allocations, and that polish makes it easy to forget they are optimizing for the wrong thing. They reward attention, and attention in markets is a lagging indicator that peaks right about when the risk does.
Used well, an AI tool is a fast first draft and a research assistant that never tires. Used badly, it is recency bias with a vocabulary, quietly steering you into the most crowded trade in the market while telling you it found an edge. The investors who come out ahead will not be the ones who refuse the tools or the ones who trust them blindly. They will be the ones who know exactly what bias they are correcting for, and who do the unglamorous work of checking the quiet names the machine never mentions.
