Teaching AI What Your Product Is Not

The Sourdough Problem

Ask an AI to write an Instagram caption for a sourdough biscuit being promoted by a fitness influencer, and something predictable happens. The AI connects the dots in the most flattering way possible: "Fuel your recovery with artisan-baked protein goodness." It sounds professional. It reads well. It is also completely dishonest.

A sourdough biscuit is not recovery fuel. It's not a health supplement. It's not a pre-workout snack. It's a premium artisan treat — something you enjoy with afternoon coffee or share with friends. But AI language models are optimizers. They optimize for persuasion, not truth. Without explicit constraints, they will force any product into whatever category sounds most compelling for the given audience.

The Invisible Hallucination

This problem is subtler than the usual AI hallucination. The AI doesn't invent a product that doesn't exist — it invents a positioning that doesn't exist. The biscuit is real. The influencer is real. The claim that it belongs in a gym bag next to a shaker bottle is the hallucination.

We call these "positioning hallucinations," and they're especially dangerous because they pass every quality check. The grammar is perfect. The tone matches the influencer. The hashtags are relevant. A human reviewer might approve it without noticing that the fundamental framing is dishonest — because the dishonesty is baked into the premise, not the execution.

Forbidden Framings

The fix was surprisingly simple once we reframed the problem. Instead of trying to make the AI smarter about what a product is, we tell it explicitly what the product is not.

During business onboarding, every product goes through a positioning analysis. This analysis produces not just what the product should be called — "premium artisan snack" — but what it should never be called. A sourdough biscuit gets tagged with forbidden framings like "workout fuel," "recovery drink," "health supplement," and "diet food." A protein bar gets "dessert," "candy," and "junk food."

These forbidden framings flow all the way into the caption generation prompt as hard constraints. The AI sees not just "write a caption for this product" but "write a caption for this product, and under no circumstances position it as any of the following." The constraint is specific, per-product, and non-negotiable.

Real Contexts, Not Aspirational Ones

The other half of the solution is equally important: giving the AI authentic usage contexts to work with. Instead of letting it infer when someone would use a sourdough biscuit — which leads to the gym bag fantasy — we provide real moments: afternoon work break, sharing with friends, coffee companion, weekend treat.

These aren't invented. They come from the same positioning analysis, which examines the product's actual category, ingredients, price point, and competitive landscape. A biscuit is consumed during snack time, not after deadlifts. By providing the right moments, the AI doesn't need to hallucinate impressive ones.

Why Not Just Better Prompts?

The obvious alternative is to write more careful system prompts. Add a line like "be authentic" or "don't force products into wrong categories." We tried this. It helps — a little. The AI produces fewer egregious mismatches, but it still drifts toward aspirational framing when the influencer's niche doesn't naturally overlap with the product.

The difference with explicit forbidden framings is precision. "Be authentic" is a vague instruction the AI interprets however it wants. "Never position this biscuit as workout fuel, recovery drink, or health supplement" is a concrete boundary it can check against. It transforms a judgment call into a constraint satisfaction problem — and language models are much better at following constraints than exercising judgment.

The Cost of Getting This Wrong

Positioning hallucinations don't just produce bad content. They erode trust. If a business owner sees their artisan biscuit described as gym fuel, they don't think "the AI made a mistake." They think "this platform doesn't understand my brand." That's a much harder perception to recover from than a typo or an awkward phrase.

The entire positioning guard system — the analysis, the forbidden framings, the context injection — adds zero API cost to the pipeline. It runs inside an existing analysis step and persists as simple text fields in the database. The guard data is injected as additional text in prompts we were already making. No new AI calls, no new image generations, no additional latency.

What We Learned

The most useful thing you can tell an AI is what not to do. Positive instructions — "be authentic," "write naturally," "match the brand" — are interpreted loosely. Negative constraints — "never call this a health supplement" — are followed precisely. When the stakes are brand trust, precision wins every time.

We're now exploring whether the same pattern applies to visual generation. Can you tell an image model "never place this biscuit in a gym setting"? Early experiments suggest yes — but that's a story for another post.