Why AI Models All Pick the Number 7

Large language models have a groupthink problem—and it's revealing something fundamental about how AI systems shape outcomes. Open your preferred chatbot and ask for a random number between 1 and 10. You'll almost always get 7. Ask again and you'll likely see 3 or 4, then 8 or 9. This pattern, documented by MIT Technology Review, isn't a quirk—it's a symptom. Our AI systems, trained on similar data with similar methods, are converging on similar outputs. They're optimizing for consensus rather than exploration, which matters enormously as these systems move from answering questions to making decisions that shape markets, research directions, and economic opportunities.

Anthropic just crossed a threshold in how AI participates in the economy. At an event for pharmaceutical executives and biotech founders, the company announced Claude Science, a flagship product designed to autonomously support scientific research. Like its predecessor Claude Code for software engineering, Claude Science can carry out meaningful work when given high-level instructions. This isn't AI as assistant—it's AI as colleague, capable of executing complex research tasks independently.

Meanwhile, the infrastructure supporting AI deployment is maturing rapidly. OpenClaw, an autonomous AI system, is advancing amid what one Forbes analysis calls "swarm culture"—emphasizing rigorous testing, benchmarks, infrastructure requirements, and autonomous AI risk management. The focus on safe deployment signals that these systems are moving from experimental to operational.

But one startup highlighted by MIT Technology Review is tackling the groupthink problem head-on, working to break LLMs out of their convergent patterns. The details of their approach remain to be seen, but the recognition of the problem itself marks an important shift in how we're thinking about AI diversity in decision-making.

If you're building with AI, test for convergence. Run the same prompt multiple times. Ask for alternatives. Push your systems to explore the edges, not just the center of probability distributions. The groupthink problem isn't just philosophical—it affects product recommendations, strategic analysis, and any application where you need genuine option generation.

For leaders making decisions about AI integration, consider: are you using these tools to find consensus faster, or to explore possibility space more thoroughly? Both have value, but they're different strategies with different outcomes. Claude Science's entry into research environments suggests we're about to see which approach wins in high-stakes domains.

And if you're evaluating AI vendors, ask about diversity of outputs. Systems that reliably produce "7" when asked for randomness may be equally predictable in strategic recommendations—which could be exactly what you want, or exactly what you're trying to avoid.

Here's what strikes me: we're optimizing AI for confidence and consistency at exactly the moment when economic uncertainty demands exploration and adaptation. The future economy won't be won by entities—human or AI—that converge on "7" fastest. It will be built by those who can generate and evaluate the full range from 1 to 10, especially the options that seem improbable.

The deployment of autonomous research AI and the maturation of AI infrastructure aren't problems to solve. They're realities to navigate. But if these systems inherit our bias toward consensus, toward the middle of the distribution, toward what sounds most plausible, they'll accelerate us toward predictability precisely when we need peripheral vision.

Monitor Anthropic's Claude Science deployment in pharmaceutical and biotech environments. How research AI performs when stakes are high and verification is rigorous will tell us a lot about where autonomous AI goes next—and which economic domains transform first.