The viral claim that large language models "pick 7" because of gradient descent is technically thin, and the real explanation is far more interesting for anyone who builds or depends on these systems. The math behind why 7 appears so often in random-number tasks has less to do with training algorithms and more to do with how probability distributions collapse under sampling constraints, a nuance that matters when you trust an AI to make decisions, not just play games.
This isn't an academic curiosity. When we see headlines about Four custom mods slash AI coding errors and boost test performance 100x, the underlying question is the same: how do you know whether a model's output reflects genuine reasoning or a statistical shortcut? The "pick a number between 1 and 10" experiment reveals that LLMs are acutely sensitive to training data distribution, they've seen "7" chosen in countless human-generated examples, and their next-token prediction naturally favors that frequency. Gradient descent didn't *learn* 7; it learned that humans pick 7, then replicated the pattern. The distinction matters because it tells us these models are mirrors of our own biases, not independent thinkers.
The practical takeaway for spreadsheet users and data professionals is straightforward: treat any single output from an LLM as a candidate, not a conclusion. This connects directly to the work described in When AI Rankings Favor Luck Over Skill, the Real Winner Isn't Clear, where small sample sizes made top models statistically indistinguishable. Both cases highlight the same fragility, when the sample is small or the prompt is vague, the model defaults to the most statistically common path, not the most correct one. For someone building a spreadsheet workflow that relies on AI-generated data, that means validation isn't optional; it's the core of the process.
What makes this more than a trivia point is what it implies about model design. If an LLM's "random" number is actually a frequency-weighted guess, then any application that depends on randomness, sampling, simulation, Monte Carlo methods, needs explicit guardrails. The model can't be trusted to generate unbiased distributions on its own. That's not a failure; it's a design constraint that responsible builders must account for. The most concrete thing you can do today is to add a deterministic random seed or a dedicated sampling layer when your workflow requires true distribution, and to treat every AI-generated "choice" as a hypothesis worth testing against your own data.
The next time you see a model pick 7, ask not what it learned, ask what it was trained to repeat.