Quick Answer
Negative prompting means telling an AI what you don't want. In image models it's a genuine separate input that works well. In chatbots it's just text, and research shows models handle negation badly and get worse as they scale. For text, say what you do want instead.
Negative prompting is telling an AI what you don't want it to produce. And here's the thing almost every guide gets wrong: it means two completely different things depending on what you're prompting. In an image generator it's a real, separate input that genuinely steers the output. In a chatbot it's just more words in your prompt, and large language models are measurably bad at processing those words. This guide covers both, what the research found, and the phrasing that actually works.
What is negative prompting?
Two mechanisms wearing the same name. In image models it's a dedicated input: Midjourney's --no parameter, or Stable Diffusion's separate negative prompt box. What you type there isn't read as an instruction; the generation is pushed away from it. In chatbots there is no such field. "Don't use bullet points" joins every other word in your prompt, and the model has to work out from context that the phrase means the opposite of what it describes. If you learned negative prompting in Midjourney and carried it into ChatGPT, that's why the results felt inconsistent. Our Midjourney prompt guide covers the image side.
Does the image-model version actually work?
Mostly yes, though not the way people picture it. Yuanhao Ban and colleagues studied the mechanism in Understanding the Impact of Negative Prompts (2024) and found two behaviours. A delayed effect: the negative prompt acts only after the positive prompt has started rendering the content. And deletion through neutralisation: the concept is cancelled against the positive prompt in latent space. It's subtraction, not prohibition.
- Vagueness costs you. A negative term only cancels what it corresponds to, so "bad" or "ugly" has nothing specific to subtract.
- Over-stuffing backfires. Every term pulls on the result, and cancelling too much strips detail you wanted.
Why don't newer image models have a negative prompt box?
Because the trick the box relies on isn't there any more. It piggybacks on classifier-free guidance, from Jonathan Ho and Tim Salimans' 2022 paper, where the model compares a prediction steered by your prompt with an unsteered one; Stable Diffusion tools swapped the unsteered prediction for a prompt of things to avoid. Newer models often skip that step. Black Forest Labs' prompting guide says plainly that "FLUX models don't support negative prompts," and warns that "a person without glasses" makes the model focus on "glasses". Chat-based image tools take one prompt and nothing else.
So describe what's there, not what isn't. Google's Gemini image best practices suggest replacing "no cars" with "an empty, deserted street with no signs of traffic."
- "No people" becomes "an empty market at dawn, stalls covered, no one around yet."
- "No text or watermark" becomes "a clean product shot on a plain white background, unlabelled packaging."
- "Not cartoonish" becomes "photorealistic, shot on a 50mm lens, natural window light."
Our image generation prompt library has prompts already written this way.
Does negative prompting actually work on chatbots?
Worse than you'd hope, and it gets worse as models get bigger. Joel Jang, Seonghyeon Ye and Minjoon Seo tested 9 tasks with negated prompts on OPT and GPT-3 from 125 million to 175 billion parameters in Can Large Language Models Truly Understand Prompts? and found an inverse scaling law: larger models did worse. Truong, Baldwin, Verspoor and Cohn (*SEM 2023) pinned down three failures: insensitivity to the presence of negation, inability to capture what negation means, and failure to reason under it. The most capable model you have is not necessarily the one most likely to respect your "do not."
Why do models flip prohibitions into permissions?
Katherine Elkins and Jon Chun audited 16 models across 14 ethical dilemmas in Framing Instability in LLM Ethical Stance (2026). Small open-weight models of 1 to 4 billion parameters endorsed a proposed action 24 percent of the time under affirmative framing, 77 percent under simple negation, and up to 100 percent under compound "not X if Y" framing. When told "should not do X", they endorsed doing X more than three times out of four. Commercial models shifted far less, but still shifted.
The authors flag their own caveat: the binary scoring proxy over-counts genuine endorsement by 38.2 percent against human coding, so the true rates are lower, though the instability is real. The mechanism is intuitive. Naming what you want avoided puts it squarely in context, and one small negating word has to outweigh it. They propose a Negation Sensitivity Index to measure that stance stability directly.
How should you phrase constraints instead?
Convert every prohibition into an instruction. Whatever you were about to forbid, ask for what you want in its place.
- Instead of "don't use bullet points," write "write in flowing paragraphs of three to four sentences."
- Instead of "don't be too formal," write "write like you're explaining this to a colleague over coffee."
- Instead of "don't make it long," write "keep it under 150 words."
- Instead of "don't invent statistics," write "only cite figures that appear in the text I gave you, and say 'not stated' if a number isn't there."
- Instead of "don't use jargon," write "use plain English that a smart 15 year old would follow."
Each rewrite gives the model a target rather than a landmine. Anthropic's best practices for prompt engineering make it a named rule, "Tell the AI what TO do instead of what NOT to do," swapping "Do not use markdown in your response" for "Your response should be composed of smoothly flowing prose paragraphs." The same page rates "NEVER use bullet points" as less effective than a preference with a reason attached. A positive instruction also tells the model what to do about headings, lists and paragraph length, which a ban never mentions. Our guide to writing better prompts covers the broader principle.
If you're stuck with a negation you can't rewrite, warn the model it's coming. Joshua Jose Dias Barreto and Abhik Jana (Findings of EMNLP 2025) found warning-based and persona-based prompts improved overall accuracy by up to 3.17 percent, but accuracy on distractor negations by up to 25.14 percent. They also found that within a model family, attention to negation predicted performance better than model size.
How do you enforce a constraint the model keeps ignoring?
A prompt is a request, not enforcement. Escalate in this order:
- Say the positive version. It fixes most cases.
- Show the format. A filled-in example outweighs any instruction about what to avoid; see our few-shot prompting guide.
- Make the structure machine-enforced. With the API, a JSON schema or structured output mode makes the constraint a property of what the model can emit.
- Check afterwards. A programmatic check on the output catches what the prompt missed.
- Put a human on anything evaluative or high stakes.
Machine enforcement means constrained decoding. Brandon Willard and Remi Louf (2023) reformulated generation as transitions between the states of a finite-state machine, masking any token that can't continue a valid string, which guarantees the structure of the output and "adds little overhead"; their implementation is the open source Outlines library. The model doesn't get a vote.
It isn't free. Zhi Rui Tam and colleagues (EMNLP 2024 Industry Track) observed "a significant decline in LLMs' reasoning abilities under format restrictions," and stricter formats generally degraded reasoning more. So separate the thinking from the formatting: let the model reason in free text first and format in a second pass, or put a free-text reasoning field first in the schema.
- Schemas constrain shape, not content. A schema can guarantee a field called tone, not that the tone is right.
- Enums are the exception. "Status must be one of pending, shipped, cancelled" is genuinely enforceable.
- Leaving a field out is a real negative constraint, though the content can still leak into a free-text field.
- Structure survives the language switch. A schema constrains Thai output the same way it constrains English, which is rare in this guide.
Does your constraint survive a long conversation?
Usually not. Philippe Laban and colleagues ran more than 200,000 simulated conversations for LLMs Get Lost In Multi-Turn Conversation (2025) and found an average drop of 39 percent from single-turn to multi-turn, mostly from "a significant increase in unreliability." Their summary: "when LLMs take a wrong turn in a conversation, they get lost and do not recover." Multi-IF, with 4,501 three-turn conversations in eight languages, found o1-preview dropping from 0.877 accuracy at the first turn to 0.707 at the third, with higher failure rates in non-Latin scripts.
Placement helps. Wallace and colleagues' Instruction Hierarchy (2024) trained models to rank system messages above user messages and tool outputs lowest, improving defence against system prompt extraction by 63 percent. So a rule meant to last a whole project belongs in custom or project instructions, not retyped at the bottom of every message.
- Restate the rule in full rather than saying "remember the rules from earlier."
- Restart instead of correcting a drifting thread. A clean start with a better prompt beats ten corrections.
- Front-load every requirement, and put hard constraints right before the task.
- Use a standing brief for repeated work. Our guide to briefing AI agents for multi-step work shows how.
Can someone talk the model out of your constraints?
Yes. Norman Mu and colleagues tested real GPT Store and HuggingChat system prompts in A Closer Look at System Prompt Robustness (2025). o3-mini held its guardrails on 83.3 percent of handwritten tests and GPT-4o on 65.3 percent, falling to 54.6 percent on distractors. As guardrails in one prompt increased, performance "uniformly approaches zero", so each new "never" line weakens the others. Persistence matters too: Agarwal and colleagues (2024) raised the average prompt-leakage attack success rate from 17.7 percent to 86.2 percent with a sycophancy-based multi-turn attack.
Agents are harder still, because they read instructions from tools, files and web pages. ManyIH-Bench, 853 agentic tasks with up to 12 levels of conflicting instructions, found frontier models at about 40 percent accuracy. Anything that must not happen, such as customer data leaving or an irreversible action, belongs in permissions, filters and validation around the model, and you should check what an agent did, not what it said.
Do thinking models follow constraints better?
Not reliably. Xiaomin Li and colleagues evaluated 15 models on IFEval and ComplexBench in When Thinking Fails (2025) and found explicit chain-of-thought "can significantly degrade instruction-following accuracy," with models neglecting simple constraints or adding unnecessary content. Fu and colleagues' MathIF benchmark, in Scaling Reasoning, Losing Control, found "a consistent tension between scaling up reasoning capacity and maintaining controllability." So don't switch on thinking mode to fix a broken constraint. Test both modes, or let the reasoning model solve the problem and run a cheap second pass that reshapes the answer to your rules.
How do you tell if the model actually followed your constraint?
Not by asking it. Jihyun Janice Ahn and Wenpeng Yin's prompt-reverse inconsistency study asked models which candidate answers were correct and which were incorrect, two questions with complementary answers, and found them contradicting themselves constantly; GPT-4, the best, scored 39.98 where lower is better. The inconsistency correlated with negation errors at a Pearson r of 0.67. "Did you use any bullet points?" is exactly that reverse-framed question. Ask the model to list the formatting it used, and compare against your rule yourself.
- Run it ten times, not once. Eight compliant outputs out of ten is a preference with a 20 percent failure rate.
- Test both polarities. Write the rule as a prohibition and as a positive instruction, and compare. That measures your own prompt's negation sensitivity.
- Check mechanically where you can. A search for a bullet character across ten outputs takes seconds.
- Try the inputs you didn't plan for: the long one, the one in another language, the one asking for exactly what you banned.
Does negative prompting work in your own language?
Worse. Moraitaki and colleagues extended the NegBench negation benchmark to seven languages for an ICLR 2026 workshop paper. On its four-way multiple choice format, CLIP scored 39.3 percent in English but 18.0 in Greek, 15.7 in Arabic and 11.8 in Tagalog, below the 25 percent chance level. Even MultiCLIP, the most consistent, averaged 41.2 percent. These are vision-language models, not chatbots, but negation degrades with distance from English training data. If you need a hard constraint in another language, write that line in English or, better, phrase it positively.
When do negative prompts genuinely work?
- Image generation, using a real negative field, because it isn't text interpretation at all.
- Concrete, checkable exclusions. "Do not include any text in the output" is a binary the model can check. "Do not be boring" is not.
- Paired with the positive. "Write in paragraphs, not bullet points" beats either half alone.
The mistakes to avoid follow from everything above: stacking negatives, listing everything you hate (which fills the context with vivid descriptions of unwanted output), assuming a bigger model will cope, trusting a "do not" for anything that must hold, and forgetting to say what you want. Our prompt engineering basics covers prompt ordering in more detail.
What do people ask about negative prompting?
What is negative prompting?
Negative prompting is telling an AI what you do not want. In image generators it is a real, separate input field, so Midjourney's --no parameter or Stable Diffusion's negative prompt box steers the image away from what you list. In chatbots there is no such field, so a negative instruction is just more words in the prompt, and models handle those words poorly.
Do negative prompts work on ChatGPT?
Less reliably than most people assume. Joel Jang, Seonghyeon Ye and Minjoon Seo tested OPT and GPT-3 from 125 million to 175 billion parameters and found an inverse scaling law on negated prompts: bigger models got worse, not better. Saying do not do X is one of the weakest ways to constrain a chatbot.
Why do AI models ignore the word don't?
Because negation is a weak signal against a strong one. Naming the thing you want to avoid puts it in context where it influences what comes next, and the single negating word has to override all of it. Elkins and Chun measured it: small open-weight models in the 1 to 4 billion parameter range endorsed an action 24 percent of the time under plain framing, but up to 100 percent once the framing was negated, a swing of as much as 76 percentage points. Bigger commercial models shift far less than that, but they still shift.
How do negative prompts work in Midjourney?
Midjourney's --no parameter isn't read as an instruction. It works as a negative weighting on those words, so adding --no text, watermark pushes the generation away from those features rather than asking the model politely. This is why negative prompting has a real reputation in image work and a much shakier one in chat.
Should you ever use do not in a prompt?
Sometimes, but always paired with a positive instruction. Do not write in bullet points on its own leaves the model guessing. Write in flowing paragraphs of three to four sentences tells it exactly what to produce. If you need a hard prohibition, state it once, state it late in the prompt, and say what to do instead.
Sources: Ban et al. 2024 (arXiv 2406.02965); Ho and Salimans 2022 (arXiv 2207.12598); Black Forest Labs and Google Cloud prompting guides; Jang, Ye and Seo 2022 (arXiv 2209.12711); Truong et al., *SEM 2023; Elkins and Chun 2026 (arXiv 2601.21433), whose binary proxy over-counts endorsement by 38.2 percent; Anthropic, Best practices for prompt engineering; Barreto and Jana, Findings of EMNLP 2025; Willard and Louf 2023 (arXiv 2307.09702); Tam et al., EMNLP 2024 Industry Track; Laban et al. 2025 (arXiv 2505.06120); Multi-IF (arXiv 2410.15553); Wallace et al. 2024 (arXiv 2404.13208); Mu et al. 2025 (arXiv 2502.12197); Agarwal et al. 2024 (arXiv 2404.16251); ManyIH-Bench (arXiv 2604.09443); Li et al. 2025 (arXiv 2505.11423); Fu et al. 2025 (arXiv 2505.14810); Ahn and Yin (arXiv 2504.01282); Moraitaki et al., ICLR 2026 workshop (arXiv 2604.18942). All linked above.