Prompting for accuracy: getting AI to be honest about what it doesn't know
AI does not lie - it predicts what comes next. The trick to reliable output is to write prompts that make 'I don't know' a more likely prediction than a confident-sounding fabrication.
Most AI users have hit the wall: a confident answer that turns out to be wrong, sometimes spectacularly. The model invented a citation, a date, a price, a person who does not exist. The hallucination post explains why.
The fix is not to abandon AI - it is to write prompts that make accurate output more likely. Five techniques below, from easiest to most effective.
1. Tell it what to do when uncertain
The single most effective addition to any factual prompt. One sentence: "If you are not sure, say you are not sure rather than guessing."
This sounds too simple to work. It works because it changes the AI's predicted-next-word probabilities. Without the instruction, the path of least resistance is to produce a plausible answer. With it, "I am not certain" is a valid response that the AI will choose more often when its training data is thin.
Add this to every prompt where factual accuracy matters: emails about specific regulations, summaries of specific documents, explanations of specific cases.
2. Provide the source material
AI working from material you provide is far more accurate than AI working from memory.
Bad: "Summarise the key points of the latest Bank of England rate decision."
Good: "Here is the Bank of England statement [paste the statement]. Summarise the key points in 200 words."
The first version invites the AI to fabricate. The second constrains it to material in the prompt. Hallucinations drop sharply.
This is why RAG systems are so valuable for business work - they automate the source-providing step at scale.
3. Ask for citations and check them
When AI states a fact that matters, ask for the source. Add to the prompt: "For each factual claim, cite the source. If you cannot cite a source, say so."
Then check the citations. AI sometimes invents plausible-looking citations - papers, court cases, news articles - that do not exist. Copy each citation into a search engine. If it is not findable, the citation is fabricated and the underlying claim is suspect.
This is more important than it sounds. The most-cited example of this going wrong is a 2023 US court case where a lawyer submitted AI-generated fictional case law. The cases were beautifully formatted with realistic-sounding parties and citations. None of them existed.
4. Verify chains of reasoning
When AI produces a multi-step argument or calculation, ask it to show the steps. Then check each step.
Add to the prompt: "Show your reasoning step by step. After your conclusion, identify any step where you are uncertain."
AI will sometimes get the right answer for the wrong reasons - or the wrong answer through a chain of small errors. Reading the chain catches both.
This is particularly important for tax calculations, eligibility assessments, contract interpretations - any answer that is the result of multiple rules combined.
5. Adversarial review
Use a second AI as a checker.
After the first AI produces an answer, paste both the original prompt and the answer into a different AI and ask: "What is wrong with this answer? What would you check before relying on it?".
Different AI models have different strengths and different blind spots. The disagreements between them are informative. Often one will spot a flaw the other missed.
For high-stakes work, this two-model pattern is the AI equivalent of a colleague reviewing your draft. Cheap, fast, surprisingly effective.
Putting it together
A high-accuracy prompt template combining the techniques:
[Your task description]
Source material:
[paste the relevant source documents]
Required:
1. Show your reasoning step by step.
2. For each factual claim, cite the source paragraph from the
material above. If a claim is not supported by the source,
flag it.
3. If you are uncertain about any step, say so explicitly.
4. After your conclusion, list anything a careful reviewer
should double-check.This is verbose, but for high-stakes work it pays back many times over. For drafting emails or generating marketing copy, you do not need it. For a tax interpretation, a contract clause analysis, a regulatory question - it is the difference between AI that helps and AI that misleads.
When to use each technique
Match the technique to the cost of being wrong.
- Drafting emails, social posts, marketing content: technique 1 is enough.
- Summarising documents you have: techniques 1 and 2.
- Explaining technical content to a customer: techniques 1, 2 and 3.
- Calculations, eligibility assessments: techniques 1, 2, 3 and 4.
- High-stakes professional advice (tax, legal, regulatory): all five, plus a human expert review before anything goes external.
The honest summary
AI is not going to stop hallucinating in the next few years. Better models hallucinate less but not never. The skill of writing prompts that make accurate output more likely is going to be useful for the rest of your career.
Build the habit on the easy cases. By the time you need it on a high-stakes case, the technique is automatic.
If you would like to walk through your specific use cases and where the accuracy risks sit, that is what a discovery call is for. The first session of the AI Breakfast Club training covers prompting in detail.