Telling AI What Not to Do: Does It Work?

A baker asks for a description of his traditional baguette for his online listing. At the end of the prompt he writes: "Don't use the words artisanal, authentic, or craftsmanship, and don't talk about passion." The reply begins: "Far from mass-produced baguettes, our bread is made with care..." The banned words are gone, but the idea is there, and it's even delivered with a more worn-out phrase.

Negative instructions give uneven results. They aren't useless, and they shouldn't be banned: they're good for some things and not for others.

Why a prohibition holds up poorly

To write, the model relies on what's in your prompt. When you name a word or an idea there, even to rule it out, you make it present in its context. Sometimes the answer drifts toward it despite the prohibition. It isn't systematic, and it depends on the tool and the wording, but it's frequent enough that you should account for it.

There's a second, simpler reason. A prohibition says what's excluded without saying what's expected. "No sales tone" leaves a thousand possibilities, several of them just as bad. The AI picks one, often the most banal.

Rephrasing as a positive instruction

In most cases, it's enough to describe what you want instead. Here are a few typical conversions:

The right-hand column can be checked: fifteen words can be counted, whereas "no sales tone" remains an impression.

For missing facts, offer a way out rather than a ban: "If a piece of information isn't in the text, write NOT SPECIFIED." It's one of the most useful moves for reducing AI inventions.

When a prohibition is still useful

A prohibition still makes sense in three cases.

• A specific word or phrase: a competitor's name, a term your industry regulates, a turn of phrase your client hates.
• A substantive limit: "don't suggest any price," "don't promise any deadline."
• A format: "no bullet points, no bold."

In these cases, state the prohibition once, along with what replaces it. And above all, don't count on it without checking.

Check instead of believing

A prohibition guarantees nothing, so verify. For specific words, your browser's search function (Ctrl+F) is more reliable than your eye. For fuzzier ideas, use a second request, preferably in the same conversation:

Reread your description of the traditional baguette. Find every sentence that contains one of these words or ideas: artisanal, authentic, craftsmanship, passion, "far from industrial". Quote them, then rewrite the description, replacing them with a concrete fact: how long the dough rises, the time of the first bake, the flour used [TO COMPLETE].

The concrete replacement keeps the AI from removing the word and leaving a hole.

When the negative works well

A prohibition works better when it applies to an action rather than a subject: "don't send anything," "don't modify the file," "only fix spelling." That's the logic of briefs for agents. For images, the negative prompt is a separate field in some tools.

A test on your last prompt

Take a recent prompt and underline every "don't," "no," and "avoid." For each one, ask yourself: what would I like to see instead? Write it, delete the prohibition, rerun, then compare. You'll probably be left with one or two real prohibitions. For those, plan the after-the-fact check described above.