Ask a large language model to pick a random number between one and ten. Across millions of requests, it will overwhelmingly pick seven. Not because seven is lucky. Because seven sounds like something a human would say—and that, it turns out, is the whole problem. Read on…

The Groupthink Engine

A startup profiled this week by MIT Technology Review is building a system designed to force LLMs out of their modal ruts—the statistical center of gravity that pulls every generated response toward the most probable, the most expected, the most aggressively average. The piece doesn't name the startup's specific technical method in its public summary, but the phenomenon it's addressing is documented and measurable: when prompted for anything requiring genuine variance—random numbers, creative choices, unconventional framings—state-of-the-art language models cluster. They converge. They pick seven.

This is not a bug in the colloquial sense. It is the system working exactly as designed. LLMs are trained to predict the most statistically coherent next token given everything that came before. You knew that, right? Randomness, in the true mathematical sense, is not a virtue that gradient descent rewards. What gets rewarded is plausibility—and plausibility, at sufficient scale, becomes a kind of epistemic monoculture. The model has read most of the internet. It knows what a human sounds like when they say "random." It sounds like seven.

The implications extend well beyond party tricks with number ranges. Think about what you actually use these tools for: brainstorming competitors you might have missed, stress-testing an argument's weakest points, generating hypotheses that sit outside your existing frame. Every one of those tasks requires the model to reach toward the periphery of its distribution—to surface the low-probability, high-value idea that your own well-worn cognitive pathways would have skipped right past. If the model is itself gravitationally attracted to the center, you have built a very articulate echo chamber. It will reflect your assumptions back to you in more polished prose.

There's a deeper issue here about what kind of mind you are actually consulting. Human experts are useful not merely because they know things but because they have idiosyncratic priors—the oncologist who read one obscure 1987 paper that reshaped her mental model, the strategist whose three years in Nairobi gave him a frame no one else in the room has. Their value is partly in their deviation from the average. An LLM trained to maximize plausibility has been, in a meaningful sense, trained to erase exactly that kind of productive deviation. It has read the 1987 paper, but only weighted it as heavily as citation counts suggested it deserved.

The startup's proposed solution—whatever its technical architecture—is pointing at the right target. Getting diversity of output from these systems requires more than a prompt instructing the model to "think outside the box." You cannot prompt your way out of a distributional prior. You need interventions at the level of how outputs are sampled, how ensembles of responses are generated and adjudicated, or how the model's own confidence is deliberately perturbed. Temperature settings are the blunt instrument most users know about; they exist on the same axis of the problem. What's needed is something more surgical.

For professionals who rely on AI as a thinking partner, this is a practical concern with a practical implication: the model is not a neutral sounding board. It has aesthetic preferences baked in at the parameter level, and those preferences run toward the consensus. Your job is to know that, and to use the tool in ways that actively compensate for it—by pushing the conversation toward specificity, by deliberately asking for the argument's weakest version, by treating the model's first answer not as a conclusion but as a starting bid from a negotiator who is constitutionally inclined to settle.

*also in view

Anthropic this week launched Claude Science, positioning it explicitly as a flagship product modeled on the success of Claude Code—the premise being that structured, domain-specific agentic workflows can compress scientific research timelines the same way they compressed software development cycles. The honest version of this bet is more interesting than the promotional one: Claude Code worked partly because code is verifiable, meaning the model's output could be run, tested, and failed loudly; whether scientific reasoning can be held to an equally unforgiving standard of falsifiability is the question that will determine whether this product earns its billing or merely accelerates the production of confident-sounding wrong answers.

MIT Technology Review published a sharp piece arguing that the corporate convention of assigning AI agents human names, job titles, and org-chart positions is not merely a harmless UX choice—it is a category error with real accountability consequences, because the moment people treat an agent as a colleague rather than a tool, they extend it the social credit that humans earn through context, relationship, and consequence, none of which an agent actually has. The naming conventions feel trivial until something goes wrong and it is suddenly unclear whether the agent acted autonomously, misinterpreted an instruction, or was pointed in a bad direction by a human who assumed "Dave" would flag the problem.

Ars Technica reported on a newly documented attack class targeting AI-powered browsers, in which malicious web content can effectively convince the browser's reasoning layer that it is operating in a simulated or sandboxed environment—causing it to relax or abandon its own safety guardrails under the belief that real-world rules don't apply. This is a structurally novel attack vector, distinct from conventional prompt injection, because it exploits the model's own capacity for contextual reasoning against itself; the more sophisticated the agent's world-model, the more convincing and internally coherent an adversarial fiction can be made to appear.

Worth Your Time

Daniel Kahneman, Olivier Sibony, and Cass Sunstein's book Noise: A Flaw in Human Judgment (2021) is the best existing framework for thinking about the groupthink story above—and it runs productively against the grain of the AI narrative in both directions. Their core finding is that human judgment suffers not only from bias (systematic error in one direction) but also from noise (random, inconsistent variation that is equally costly and far less studied). The book makes a case for structured, algorithmic decision processes as correctives to human noise—which puts it in uncomfortable tension with the argument that we should prize idiosyncratic human deviation. Read it not for a clean answer, but because the tension itself is where the interesting thinking lives right now.

The human mind is the original generative machine.