More Context Barely Helps

Context serves two functions when working with AI. Adding information and redirecting search. One of them barely works.

Function one: information. Adding facts, data, background. The assumption is that more information produces richer analysis. Tested it. Near zero effect. The model already had enough to generate a thorough response. Extra information barely changed the theme count or the analytical range.

Function two: search. Adding constraints that redirect where the model looks. "List the assumption each option depends on. Score reversibility. Include one option the board would reject." These don't add information. They redirect attention. The same context content is utilized unevenly across positions: structure and placement matter, not just what's present.

Testing this on one strategy scenario with one model, ten runs per condition: the search function found a specific strategic option in 17 out of 20 outputs. Baseline without constraints: 1 out of 20. One scenario, but the signal is statistically overwhelming. Not marginal. Dominant.

Same experiment. Information added almost nothing. Search constraints changed everything.

The mechanism: when you add context as information, the model already has a reasonable baseline. It knows enough to produce a thorough-looking response. More facts push it marginally because the model wasn't limited by information. It was limited by which parts of its knowledge space it visited. Additional information slightly adjusts the distribution but doesn't change the territory. Iteration within a single mode hits a similar ceiling for the same reason.

When you add context as search constraints, you redirect the model's attention into territory it wouldn't visit by default. "Include one option the board would reject" doesn't add information about the problem. It redirects the model's search into contrarian territory. The constraint doesn't make the model smarter. It changes where the model looks. And where the model looks determines what it finds. The same task-type dependence shows up when structure that helps convergent tasks actively harms exploratory ones.

This reframes "give it more context" entirely. The question isn't how much context. It's what kind. A paragraph of constraints that redirect attention outperforms a page of additional data. Three sentences with the right constraints produced 17 out of 20 outputs finding territory that zero out of 20 found without them. A page of additional data produced no measurable change in analytical range.

Cross-model replication on a different scenario (product roadmap prioritization, xAI and Claude) confirmed the direction but with a weaker design. The original found an indirect effect: a search constraint helped the model discover a specific option (data monetization) that wasn't named in the prompt. The replication tested compliance: the search prompt explicitly asked for contrarian options and assumption identification. Both models complied. That's instruction-following, not the same as the original's indirect search.

What was more interesting: on Claude, information suppressed contrarian options. Baseline: 50 percent of outputs included unconventional alternatives. With added information: 10 percent. More background data gave the model more default territory to spread across, crowding out the angles the information didn't mention. The suppression wasn't prompted. It emerged from the design. On assumption identification, the pattern was cleaner: Claude baseline produced zero per output, information 0.2, search constraints 4.4.

One limit: the information effect may be baseline-dependent. The original experiment found a large effect for added information. The replication found almost none. The difference: the baseline outputs were already richer in the replication (664 words vs 380). When the baseline is sparse, information helps. When the baseline is already thorough, information adds almost nothing. Search constraints work regardless.

You added more context expecting richer analysis. Context doesn't inform. It redirects search. The AI already knew enough. Your context told it where to look, not what to know. The question is whether you chose the direction deliberately or let the context choose for you.

Frame Inventory. After your next AI session, write down the frame your context set. What did it emphasize? Then write two alternative frames you didn't explore. The gap is where your context narrowed instead of expanded.