How to Stop AI Hallucinations: 6 Prompt Tactics That Reduce Made-Up Answers

AI hallucinations are a structural problem, not a bug you can patch. But you can reduce their frequency and make the ones that slip through much easier to catch — using specific prompt tactics that cost almost nothing to apply. Here are 6 that work. Why AI Makes Things Up Large language models don't "know" facts the way a database does. They predict the next most-plausible token based on patterns in training data. When a question lands outside that training distribution — or asks for specifics the model never reliably learned — it still produces a confident-sounding answer, because fluency is what it was optimized for. Key mechanism: models have no internal "I'm guessing" flag. Without explicit instructions, they default to generating something rather than nothing. Tactic 1 — Demand Sources Ask the model to cite specific sources (paper title, author, publication year, URL). It won't hallucinate less, but fabricated citations are far easier for you to spot and verify than fabricated claims embedded in flowing prose. Add this to any prompt where accuracy matters: Cite the source (author, publication year, title) for every statistic you include. If you cannot cite a real source, omit the statistic. Tactic 2 — Require "I Don't Know" By default, the model treats a non-answer as a failure. Reframe the instruction to give it explicit permission — even an obligation — to say it doesn't know: If you are not confident about a fact, write "I'm not certain about this" instead of guessing. This single addition adds almost no prompt length and consistently improves reliability on factual queries. Tactic 3 — Scope the Domain Tightly The broader the question, the more the model has to extrapolate. Narrow the domain to what you actually need. Instead of: Tell me about AI regulation. Try: Summarize the key provisions of the EU AI Act as passed in March 2024 — only provisions you are highly confident about. Scoped framing reduces conflation of different laws, versions, and dates. Tactic 4 — Ask for Reasoning, Not Just Answers When you ask the model to show its reasoning ("explain step by step"), hallucinations often surface as logical gaps or contradictions. The act of reasoning also slows down the next-token prediction enough to produce more grounded output. Walk me through your reasoning step by step before giving your final answer. This is especially useful for multi-step analysis tasks where errors compound. Tactic 5 — Use Verification Prompts After the model answers, ask it to verify itself: List any claims in your previous answer that you are less than 90% confident about. This meta-prompt consistently flags the weakest parts of a response. It won't catch everything — the model can be confidently wrong — but it surfaces a meaningful portion of uncertain claims. Tactic 6 — Provide the Facts, Ask for Analysis The most reliable way to stop hallucination: don't ask the model to retrieve facts. Paste the facts in and ask the model to analyze, summarize, or reformat them. Here is the press release: [text]. Extract the three most important policy changes. When you supply the source material, retrieval hallucination is eliminated by design. The model's job shifts from "remember facts" to "process the text I gave you." Before and After Examples Example A — Statistics Request Without tactics: What percentage of AI outputs contain hallucinations? Model invents a confident-sounding number with no real source. With Tactic 1 + 2: What percentage of AI outputs contain hallucinations? Cite the source. If you can't cite a real study, say so. Model either cites a verifiable study (e.g., Stanford HELM benchmarks) or admits uncertainty. Example B — Specific Date/Event Without tactics: When did OpenAI release GPT-4? High risk of confusing version dates. With Tactic 3 (scoped): Based only on publicly announced dates, when was GPT-4 released? If uncertain, say so. Scoped framing reduces conflation of different model versions. These six tactics reduce fabrication; they do not eliminate it, because the model is still answering from parameters rather than documents. The structural fix is to hand it the documents — that is what retrieval-augmented generation does, and it is why production systems use it instead of relying on prompt wording alone. Tactic Summary Tactic What it does Best for Demand sources Makes fabrication visible Research, statistics Require "I don't know" Unlocks honest uncertainty Any factual query Scope the domain Reduces extrapolation Complex / broad topics Ask for reasoning Surfaces logical gaps Analysis, multi-step problems Verification prompts Self-identifies weak spots Long outputs Provide facts, ask analysis Eliminates retrieval entirely Known-source content FAQ Can AI hallucinations be completely eliminated? No current model eliminates hallucination entirely. These tactics significantly reduce frequency and make the remaining errors easier to detect. Does GPT-4 hallucinate less than older models? Newer models show lower hallucination rates on benchmarks, but still hallucinate — particularly on niche topics, recent events, and specific numbers. What's the fastest win right now? Tactic 6 (provide your own facts, ask for analysis) is the most reliable. Tactic 2 (require "I don't know") adds almost no prompt length and helps immediately. Do prompts that demand citations actually produce real citations? Not always — models can hallucinate citations too. Demanding citations is a filter, not a guarantee. Always verify citations independently. What types of content have the highest hallucination risk? Specific statistics, legal/medical details, recent events past the training cutoff, and obscure niche topics. Is RAG better than prompt tactics? For production systems yes — RAG grounds the model in retrieved documents. For everyday use, prompt tactics are the practical alternative. Originally published at my-blog.org.