AI is bad at admitting when it does not know something. Ask a model for a statistic, a citation, or a policy detail it never actually learned, and it will often invent one, phrased with the same confidence as a fact it got right. This is called hallucination, and if you use AI for real work, it is the single most expensive habit you need to design around. The good news: most hallucinations are preventable at the prompt level, before you ever touch a fancy tool.
Why does AI confidently make things up?
A hallucination is when a language model produces text that sounds correct but is factually wrong or entirely invented. It happens because models predict the next likely word, not the true word. When the training data is thin on a topic, the model fills the gap with a plausible-sounding guess instead of stopping.
The key insight for practitioners: the model is not lying. It has no built-in signal that separates "I know this" from "this pattern looks right." Your prompt has to supply that signal.
According to OpenAI's own model documentation, hallucination rates rise sharply when a question requires specific facts, dates, or numbers that were rare in training. That is exactly where your business questions tend to live.
Pattern 1: Give the model permission to say "I don't know"
The fastest reliability upgrade is one sentence: explicitly allow the model to admit uncertainty. Models default to answering because that is what most training examples reward. Remove that pressure and confident fabrication drops noticeably.
Add a line like this to any factual prompt: "If you are not certain or the information is not provided, say 'I am not sure' instead of guessing."
This works because you are changing the reward inside the prompt. Instead of "produce an answer at all costs," you are telling the model that "I don't know" is an acceptable, even preferred, output.
In testing, this single instruction turns many invented statistics into an honest "this figure was not in the source I was given," which is the answer you actually wanted.
Pattern 2: Ground the model in your own sources
Grounding means pasting the real source material into the prompt and instructing the model to answer only from it. This is the manual version of Retrieval-Augmented Generation, and it is the strongest single lever you have. Practitioners who ground consistently report far fewer fabrications.
The structure is simple: paste the document, then add "Answer using ONLY the text above. If the answer is not in the text, say so."
Grounding beats trusting the model's memory because you replace fuzzy recall with exact text. A model reading a pasted contract clause will quote it correctly; a model recalling "a contract like this usually says" will improvise.
Use this for policy documents, meeting notes, product specs, research PDFs, and anything where the exact wording matters. If the source is long, paste the two or three relevant sections rather than the whole file.
Pattern 3: Force the model to cite and show its work
Ask the model to quote the exact sentence it based each claim on. Citation forces the model to check whether a supporting sentence actually exists, which catches fabrications before they reach you. If it cannot produce a quote, that is your signal the claim is invented.
Pair this with chain-of-verification: after the model answers, ask it to review its own answer against the source and flag anything unsupported.
A practical two-step: first prompt gets the answer with quotes, second prompt says "Check each claim above against the source. List any claim you cannot support with a direct quote."
This mirrors how a careful analyst works. You are not trusting the first draft; you are making the model audit itself, which surfaces the weak claims you would otherwise miss.
Pattern 4: Constrain the output so there is nowhere to improvise
Structured output limits where the model can wander. When you force answers into fixed fields, JSON, or a table, there is less open space for the model to add invented narrative around a real fact. Tight formats reduce padding and the errors that hide in padding.
For a data task, ask for a table with exact columns: "Return only a table with columns: Claim, Source Quote, Confidence (High/Medium/Low). No prose."
The Confidence column is the trick. It gives the model a legitimate place to express doubt instead of smoothing everything into false certainty, and it tells you which rows to double-check.
What does a combined anti-hallucination prompt look like?
The four patterns stack. Combine grounding, permission to abstain, citation, and structure into one reusable prompt and you get answers that are far harder to fake. Save it as a template and reuse it for every factual task.
Try this prompt now, pasting your own source where indicated:
SOURCE: [paste your document, notes, or data here]
TASK: Answer the question below using ONLY the SOURCE above.
RULES:
--- For each claim, include the exact sentence from the SOURCE that supports it.
--- If the answer is not in the SOURCE, write "Not stated in source." Do not guess.
--- Return a table with columns: Answer, Supporting Quote, Confidence (High/Medium/Low).
--- After the table, list any part of my question the SOURCE cannot answer.
QUESTION: [your question here]
Run this once and you will see the difference immediately: every answer now carries its own evidence, and the gaps are labelled instead of hidden.
Where do these techniques still fall short?
These patterns reduce hallucinations sharply but do not eliminate them. A model can still misread a source, quote a real sentence but draw a wrong conclusion, or confidently summarise a document it partly misunderstood. Grounding controls invention, not interpretation.
They also do not help when your source itself is wrong. If you paste an outdated policy, the model will faithfully cite outdated information. Garbage in, faithfully-cited garbage out.
For high-stakes work in finance, legal, or medical contexts, keep a human in the loop for the final check. Treat the Confidence column and the "Not stated in source" flags as a triage list telling you exactly where to look, not as a guarantee.
Putting it into your daily workflow
Start by saving the combined prompt as a reusable template in whatever tool you use most. The moment a task depends on specific facts, reach for it instead of a bare question. That one habit change removes most day-to-day hallucination risk.
The deeper point is that reliable AI is a design choice, not a lucky model. You get consistent output by building guardrails into how you ask, not by hoping the model behaves. That is the difference between someone who uses AI and someone who can trust what it hands back.
At UD, that is the philosophy we bring to every workflow we help teams build. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Build a reliable AI workflow with UD
Knowing these patterns is step one. The next step is wiring them into a workflow your whole team can trust, every time. We'll walk you through every step, from prompt templates to grounding your own documents and setting up the verification checks that catch errors before they reach a client.
