Why AI Defaults to Generic

Ask a model "how should we price the new product?" and you get something like this: pricing depends on several factors. Consider your costs, your competitors, your target market, your perceived value. Value-based pricing is often a strong approach. You may want to test a few price points and watch demand. Safe. Balanced. Fluent. Complete-sounding. It would fit any product in any market, which is the tell.

Now hand it your two competitors' actual price points, your margin floor, and one instruction: "recommend a single number, then name the assumption that would make it wrong." The output commits. It picks a number, defends it, and exposes where it breaks. (Illustrative contrast, not a measured pair: the point is the shape of the shift, not an effect size.)

The first output is the default, and the default is not a failure mode. It is what the model produces every time the context gives it no reason to go anywhere else. Knowing why it sits there tells you what every prompt technique you have ever been handed is actually fighting. Specificity, source material, structure, examples, personas: they are all one move. They make the default expensive enough that the model leaves it.

So why does the model sit there? Two stages of training put it there, and the rest of this is an interpretive frame: grounded in how LLM training works, not experimentally decomposed in this corpus. Pretraining makes the model the average of everything it read, and an average reads as generic. The most probable next token is the most common one: "in conclusion" after analysis, "however" after a claim, "it depends" after a hard question. The default is the mode of the distribution, which is why AI output often sounds like a well-written essay by nobody in particular. It is one. Then post-training (instruction tuning, then preference methods like RLHF) tilts that average toward what human raters rewarded: balanced over committed, hedged over strong, comprehensive over focused. Genericness comes from the first stage, safe hedging from the second.

That trained default shows up three ways, none of them a separate cause. It plays for low regret: a balanced answer is never badly wrong, so the model avoids the specific claim that could be. It closes the face of the response: "here are several perspectives" outranks "I don't know," because definitive-sounding answers scored higher than honest incompleteness. And it aims for the broadest audience: output that reads reasonable to everyone, rather than exactly right for one reader and strange to the rest. The result looks good, sounds professional, and says nothing a thousand other prompts wouldn't produce.

This is the thing the rest of the work refers back to. Every technique that improves AI output is a way of making the default path costlier than some alternative. Specificity narrows away from the generic average. Source material replaces generated content with real data. Structure forces non-default organization. The two largest moves that hold across generators are the first two: specificity against the default (the strongest single effect with confounds controlled, synergistic with quality demands: quality demands do little alone but add on top of specificity, where it already works, and together the effect is much larger) and source material against fabrication (roughly half unsourced without a source, single-digit with one). Same principle underneath both. Give the model a reason to leave the default. It is where the model goes when you don't.