Negative Prompting: Telling AI What to Leave Out
A negative prompt steers generative models away from unwanted content. Here is how it works in image models like Stable Diffusion and Midjourney, in text models like ChatGPT and Claude, with copy-ready examples.
Last updated June 21, 2026
Affiliate link, we may earn a commission at no extra cost to you.
What negative prompting actually does
Every generative model produces output by navigating a vast space of possibilities. A positive prompt pulls the result toward a target, a subject, a style, a tone. A negative prompt pushes the result away from regions you do not want. In diffusion image models this is a literal mathematical operation: the model computes a prediction conditioned on your positive prompt and a second prediction conditioned on the negative prompt, then steers away from the negative one. In text models the mechanism is softer, but the intent is identical, you are pruning failure modes. Understanding that distinction is what separates someone copying forum boilerplate from someone who can debug an output deliberately.
Negative prompting in image models
Stable Diffusion exposes a separate "Negative prompt" box in interfaces like Automatic1111, ComfyUI, and Forge. A practical, restrained negative for a photorealistic portrait is: blurry, low quality, deformed hands, extra fingers, watermark, text, jpeg artifacts, oversaturated. If a specific problem persists, weight it harder with parentheses, (extra fingers:1.5), rather than adding ten more unrelated terms. For a clean product shot you might use cluttered background, shadows, reflections, fingerprints, dust.
Midjourney has no negative box; instead it appends the --no parameter. a serene mountain lake --no boats, people removes both. To soften rather than fully exclude a concept, Midjourney also supports negative weights with double-colon syntax, such as autumn forest ::-0.5 red to reduce (not eliminate) red tones. As a rule, reach for --no for discrete objects and weighting for stylistic dialing.
Negative prompting in text models
ChatGPT, Claude, and Gemini have no dedicated negative field, so exclusions live inside the instruction. The most reliable structures are an explicit avoidance block and a banned-vocabulary list. For example: "Write a 120-word product description. Do NOT use the words ‘revolutionary’, ‘game-changing’, or ‘unlock’. No bullet points. Avoid hype and superlatives." That single paragraph combines a vocabulary exclusion, a format exclusion, and a tone guard.
One counterintuitive rule: where possible, convert a prohibition into a positive instruction. Telling a model "never sound salesy" sometimes makes salesy phrasing more salient, because the concept is now in context. "Write in a plain, factual, understated tone" usually outperforms the raw negation. Reserve hard "Do NOT" statements for concrete, checkable items, specific words, formats, claims, or sections you can verify were excluded.
Common mistakes and how to test
The biggest mistake is the kitchen-sink negative prompt, pasting a 40-term list scraped from a Discord server. Each term constrains the model, and an overloaded negative can flatten contrast, remove legitimate detail, or fight your positive prompt. Treat negatives like a debugger: generate first, identify the actual artifact, add one targeted exclusion, and regenerate with the same seed to confirm the change came from your edit and not random variation. In text, do the same, run the prompt with and without the exclusion and check whether the banned element actually disappeared. Negative prompting is a precision tool, and as of 2026 the practitioners who use it well are the ones who add the fewest, most deliberate terms.
FAQ
What is negative prompting?
Negative prompting is the technique of telling a generative AI model what to exclude or avoid, rather than only what to produce. In image models like Stable Diffusion it has a dedicated negative prompt field; in Midjourney it uses the --no parameter; and in text models like ChatGPT and Claude it is expressed in plain language ("do not mention pricing", "avoid passive voice"). The goal is the same across all of them: steer the output away from unwanted content, styles, or artifacts.
How do I write a negative prompt in Stable Diffusion?
Stable Diffusion (and forks like Automatic1111, ComfyUI, and Forge) has a separate text box labeled "Negative prompt" beside the main prompt. A common starting set for photorealistic portraits is: "blurry, low quality, deformed hands, extra fingers, watermark, text, jpeg artifacts, oversaturated". You can weight terms with parentheses, for example (blurry:1.4), to push the model harder away from blur. Keep negatives focused on observed problems rather than dumping a giant generic list, which can wash out detail.
Does Midjourney support negative prompts?
Yes. Midjourney uses the --no parameter at the end of a prompt to subtract elements. For example, "a cozy reading nook --no books" removes books, and "forest path --no people, text" excludes both. You can also down-weight a concept using ::-0.5 syntax (for example "vibrant city ::-0.5 cars"). The --no parameter is the cleaner, more predictable option for simple exclusions.
Can you use negative prompting with ChatGPT or Claude?
Text models have no dedicated negative field, so you write constraints directly into the instruction. Effective patterns include a "Do NOT" or "Avoid" list ("Do not use the words 'leverage', 'synergy', or 'unlock'"), explicit format exclusions ("no bullet points, no headings"), and tone guards ("avoid hype and superlatives"). Phrasing exclusions positively when possible ("write in plain, concrete language") often works better than a pure prohibition, because language models predict what to do more reliably than what to suppress.
When does negative prompting not help?
Negative prompts have diminishing returns. Long, generic exclusion lists copied from forums can reduce overall image quality by constraining the latent space too aggressively. With text models, a single hard prohibition can backfire (telling a model "never mention X" sometimes makes X more salient). Use negatives surgically: add a term only after you have seen the specific failure you want to remove, then test whether it actually changed the output.
What is the difference between a negative prompt and a positive prompt?
A positive prompt describes what you want the model to include, the subject, style, lighting, and mood. A negative prompt describes what you want pushed out of the result, artifacts, unwanted objects, banned words, or tones. They work together: the positive prompt sets the target, and the negative prompt prunes the failure modes. The best results usually come from a tight positive prompt plus a short, specific negative prompt rather than an enormous list on either side.