A Columbia Business School researcher just published a study analyzing over 110,000 real-world decisions, and the conclusion is as funny as it is terrifying: AI isn't making you smarter, it's making you blander. Not by accident. By design.
The Most Average Machine Ever Built
Here's the thing about large language models that nobody in the marketing department wants to say out loud. They are, at a fundamental level, mediocrity engines. They work by predicting the most statistically likely next word, next idea, next recommendation. Average is not a bug. It is the entire architecture.
Columbia Business School professor Sandra Matz, a computational social scientist, put it plainly to CBS News: "LLMs predict the most likely next word in a sentence or event in a sequence, and by definition, that's average." Ask it for a movie recommendation. Ask it what color to paint your living room. You will get the answer that is most common for people broadly like you. You will get the consensus. You will get beige.
"It homogenizes decisions," Matz said, "and we all get the same output." Which, if you think about it, is a pretty wild thing to say about the technology that half the planet is currently outsourcing their thinking to.
How They Actually Ran This Study
This wasn't a vibes-based theory somebody cooked up in a faculty lounge. According to CBS News, Matz and her co-authors analyzed more than 110,000 real-world decisions made by 1,000 people, then compared those decisions to choices made by both generic and personalized AI agents. They also pulled data from the myPersonality project, a Facebook application that ran personality tests on users who shared their profiles for research.
What they found was consistent across the board. When people leaned on AI to shape a decision, whether that was where to vacation, what shoes to buy, what to do on a Saturday night, the AI pushed them toward the most common choices. Not the most interesting ones. Not the most personally resonant ones. The most statistically normal ones.
Even when an AI agent had been personalized and theoretically understood that a given user sometimes makes unusual or out-of-character choices, it still nudged them back toward the center. Matz's words in the study, as CBS News reports, were that "LLM agents nudge behavior toward more normative options and narrow the range of what individuals explore." The quirks get sanded off. The edges disappear.
"AI Hates Risk" Is a Choice Someone Made
This is the part that should make you genuinely angry, not just mildly annoyed. Matz is not describing some unavoidable consequence of how the technology works. She is describing a business decision. "AI hates risk because we train it that way," she told CBS News. "It wants to keep you on the platform, so it shows you what you already like and not stuff on the outskirts of what you do."
So the companies building these tools have made a deliberate choice. Keep users comfortable. Keep users in familiar territory. Keep users on the platform. The cost of that choice, according to this research, is the slow erosion of individual taste, spontaneity, and the kind of weird personal particularity that makes human beings interesting to each other.
Matz was clear that it does not have to work this way. AI apps are not physically incapable of recommending something unexpected. They are programmed not to. The question of whether that ever changes is almost entirely a question about whether it would be more profitable to do so, which is not a question that fills you with optimism.
The Part Where Culture "Collapses Into a Single Set of Preferences"
Matz's proposed fix is an "exploration mode" that users could opt into when they want less predictable, more unconventional recommendations. It's a reasonable idea. It is also, almost certainly, something a meaningful chunk of the AI industry will ignore completely unless someone forces their hand, because "exploration mode" sounds like a feature that increases churn and makes the recommendation algorithm harder to optimize.
But the stakes she's describing go beyond anyone's personal music taste or vacation shortlist. CBS News reports that Matz framed the broader concern as preventing culture from collapsing "into a single set of preferences." That's not small. If everyone is using similar AI tools to make similar decisions about what to read, watch, eat, wear, listen to, and think about, what you eventually get is not a culture. It's a product catalog.
The rich, chaotic, contradictory mess of human taste and behavior is, it turns out, not a side effect of civilization. It is more or less the point of it. And we are apparently very cheerfully letting a statistical averaging machine iron it flat.
The Dingo Take
Let's be honest about what this study is actually describing. We took the most powerful information technology ever built, pointed it at human decision-making, and the primary product it keeps generating is: the safe choice. The expected choice. The choice that keeps you clicking. We have built a recommendation engine for a world with no edges, and we are feeding it to billions of people and calling it progress.
The AI industry will tell you that personalization is the answer, that the models are getting smarter, that future versions will understand the full complexity of who you are. Maybe. But Matz's research suggests that even personalized AI still pulls users toward the normative center. The incentive to keep people comfortable and on-platform is baked into the business model at every level. Exploration mode is not going to happen at scale without a regulatory push or a genuine market demand, and neither of those is exactly materializing at speed.
So here we are. Handed a technology that could theoretically expand the scope of what any individual can know, do, and imagine, and we have mostly used it to make sure everyone ends up at the same restaurant with the same shoes watching the same shows. Somewhere, a very average AI is calculating the most likely next sentence to follow that observation, and whatever it suggests, this article is deliberately not using it.
