Why Your AI Keeps Making Things Up — And 6 Prompting Techniques That Actually Fix It

Why Does AI Make Things Up? The Honest Answer

AI language models don't "know" facts the way a database does. They generate text by predicting the most statistically likely next token given everything that came before — which means when the training data doesn't clearly support a specific answer, the model will still produce one. It fills the gap with plausible-sounding text. This is called hallucination, and it is a structural feature of how these models work, not a bug that will be fully patched away.

According to a 2026 report by AllAboutAI, current frontier models hallucinate at rates ranging from 3% to 27% depending on task type — with factual recall tasks and citation requests sitting at the higher end. For practitioners using AI daily, this means roughly 1 in 10 factual claims may be fabricated, especially for specific names, dates, statistics, or URLs.

The good news: hallucination is not random. It follows predictable patterns. And once you understand those patterns, you can use prompting techniques that reduce it by 40–60% in practice — without switching models or paying for API access. This article covers six techniques that work across ChatGPT, Claude, and Gemini, with examples you can use today.

 

What Types of Tasks Trigger the Most Hallucinations?

Hallucinations are not evenly distributed. A 2026 analysis in SQ Magazine identifies five high-risk task types where AI consistently fabricates or distorts information: specific citations and references, recent events (post-training-cutoff), niche or highly specific statistics, named entities (people, companies, products), and multi-step calculations involving real-world data.

Low-risk tasks — where hallucination rates drop below 5% — include summarising content you've already provided, rewriting or reformatting existing text, generating structured outputs from explicit templates, and brainstorming where factual accuracy isn't required.

This distinction matters for practitioners. Hallucination is largely a function of what you're asking, not just which model you're using. The same GPT-5.4 that fabricates a specific statistic will summarise a document you've pasted with near-perfect accuracy. The fix isn't always to switch models — it's to change the task design.

 

Technique 1: Ground the Model in Your Source Material

The single most effective technique for reducing hallucination is also the simplest: don't ask the model to recall facts — give it the facts and ask it to work with them. Paste the document, article, or data into the conversation and tell the model to answer only from that material.

This technique — sometimes called grounded generation — works because it eliminates the gap between what the model needs and what it has access to. According to research published in PMC on hallucination attribution, grounding prompts reduce hallucination by approximately 40% on information retrieval tasks compared to open-ended recall prompts.

Try This Prompt:

--- The following is a passage from [document name]. Answer the question below using ONLY information contained in this passage. If the passage does not contain enough information to answer the question, say "The passage does not address this directly" rather than inferring.

--- [Paste passage here]

--- Question: [Your question here]

The phrase "rather than inferring" is doing real work here. Without it, the model will often fill gaps by reasoning forward from partial evidence — which is exactly what produces confident-sounding hallucinations.

 

Technique 2: Ask the Model to Show Its Work

Chain-of-thought (CoT) prompting — asking the model to reason step by step before stating its conclusion — is one of the most researched hallucination-reduction techniques available. A 2024 Stanford University study found that combining retrieval-augmented grounding with chain-of-thought prompting led to a 96% reduction in hallucinations compared to unstructured baseline prompts.

The mechanism is straightforward: when a model externalises its reasoning, it has fewer places to hide weak inferences. A confident-sounding answer generated in one step can conceal a chain of guesses; a step-by-step reasoning trace surfaces those guesses explicitly, giving you the opportunity to spot and challenge them.

Try This Prompt:

--- Before giving your final answer, walk me through your reasoning step by step. For each step, indicate whether you are (a) drawing on information I provided, (b) drawing on general knowledge you're confident in, or (c) making an inference or estimate. Label each step accordingly. Then give your final answer.

This labelling instruction is particularly useful because it forces the model to self-categorise its confidence level, which it does more accurately than you might expect.

 

Technique 3: Require Sources and Confidence Levels

When your task requires factual claims, add a structural rule: every claim must be accompanied by a source citation and a stated confidence level. This technique was validated in a large-scale news analysis system study, where requiring source attribution for every claim reduced hallucinations by approximately 40% compared to unattributed prompts.

The mechanism is similar to CoT: the model cannot generate a confident-sounding unsourced claim when you've instructed it that all claims require attribution. Fabricated citations do still occur — this is important to know — but the rate drops substantially, and fabricated citations are easier to verify and spot than fabricated facts embedded in fluent prose.

Try This Prompt:

--- For every factual claim you make, include: (1) a source in brackets immediately after the claim — either a specific document I've provided, or the name of a public organisation, report, or study, and (2) a confidence rating: HIGH (you are certain), MEDIUM (likely but unverified), or LOW (estimated). If you cannot provide a source for a claim at HIGH or MEDIUM confidence, do not make the claim — instead, flag it as [UNVERIFIED].

 

Technique 4: Use Output Formatting Constraints

Asking the model to produce structured, constrained output — JSON, numbered lists with explicit fields, formatted tables — reduces hallucination by narrowing the design space for the model's response. When the output format is rigid, the model has less room to introduce plausible-but-fabricated filler. Free-form prose generates the most hallucinations; strict schema output generates the fewest.

According to research at Voiceflow on LLM hallucination prevention, structured output constraints are particularly effective for data extraction tasks — where the model must pull specific fields from documents rather than summarise or interpret freely.

Try This Prompt:

--- Extract the following fields from the document I've provided. Return your answer ONLY as a JSON object with these exact keys: name, date, location, key_figure, main_claim. If a field is not clearly stated in the document, use null as the value. Do not infer values that are not explicitly present.

The "use null" instruction is critical. Without it, the model fills missing fields with its best guess — which is indistinguishable from accurate data in the output.

 

Technique 5: Break Complex Tasks into Smaller Steps

Multi-step tasks increase hallucination risk because errors compound across steps. If a model makes a small inference error in step 2 of a 6-step chain, every subsequent step builds on a faulty foundation — and the final output can be substantially wrong while appearing internally consistent.

The fix is to decompose complex tasks and verify each step before proceeding. Instead of asking "Research this topic and write a 1,000-word report with statistics," break it into: (1) list the key questions this report should answer, (2) for each question, what information do I need to provide?, (3) given the information I provide, draft the section for question one. Each step is checkable before you proceed.

Try This Prompt:

--- We're going to complete this task step by step. Do not proceed to the next step until I confirm the current step is correct. Step 1 only: [first sub-task]. Stop after Step 1 and wait for my feedback.

This approach is slower, but for high-stakes outputs — client-facing documents, research summaries, compliance-related content — the error-catching benefit outweighs the extra time.

 

Technique 6: Build a Personal Verification Habit

No prompting technique eliminates hallucination entirely. The final layer of defence is a consistent personal habit: verify any specific, high-stakes claim before using it. This means checking statistics against named sources, confirming URLs actually resolve, and cross-checking named-entity details (person titles, company names, dates) before they leave your draft.

A practical system: use AI to produce the first draft, then read with a specific verification mindset. Flag every specific number, named source, and proper noun. Spot-check a random 20%. For client-facing or legal content, check all of them. This takes 5–10 minutes for a 1,000-word document and catches the majority of consequential errors.

The goal isn't to distrust AI — it's to use it with appropriate calibration. According to Lakera's 2026 guide to LLM hallucinations, practitioners who combine prompting guardrails with a verification workflow report fewer than 2% of consequential hallucinations in final outputs, compared to 15–20% among users who rely solely on model output without review.

懂AI的冷，更懂你的難 — UD 同行28年，讓科技成為有溫度的陪伴. These six techniques represent the practical difference between AI that occasionally lets you down and AI that works reliably in your workflow. The models are powerful. The system you build around them determines whether that power is consistent.

 

Build an AI Workflow That Delivers Consistent Results

You now have six tested techniques for reducing AI hallucination in your daily work. The next step is building these into a repeatable workflow rather than applying them ad-hoc. We'll walk you through every step — from prompt system design to output verification workflows that actually scale.