You asked ChatGPT a question and it gave you an answer. Confident, complete, sourced, ready to use. You pasted it into the deck, the article, the client report. Somewhere inside it was a study that has never existed. You didn't realise, but someone else will later. Undermining your professional credibility in the process.

Large language models are built to be helpful. They have no conscience, so they do not feel bad about inventing a statistic, and they think they did the right thing because they told you what you wanted to hear. As founder of Coachvox , I work with AI every day. Thousands of coaches use our platform to build the AI version of themselves, and it is important that those AI versions do not make things up.

Believe everything AI tells you and you go down a rabbit hole. You publish things that are not right. You build a view of the world out of details that were generated to please you. Getting the truth out of AI takes instruction, and people have already done the work of figuring out what those instructions are. Here is what they found.

Why ChatGPT and Claude do not tell the truth

Your AI wants to make you happy

Ask a model for ten customer testimonials and you get ten customer testimonials. The answer arrives complete, because a complete answer is helpful. Researchers at OpenAI and Georgia Tech showed that the way these models are scored incentivizes hallucinations , because an evaluation that rewards accuracy rewards a guess more than an admission of doubt. A model that says it does not know scores zero. A model that invents something plausible can score full marks. Guessing wins.

There is no shame in the machine. A person who invents a client quote feels guilty when they type it, and that feeling is what stops most people doing it. Your AI feels nothing at all. It produces the sentence, you read the sentence, and the transaction is complete as far as it is concerned.

The testimonials that never existed

Lucy, the YouTuber behind English with Lucy , asked ChatGPT to go through a spreadsheet of testimonials her company had received and pull out the best ones. The ten it presented were so beautiful they brought her to tears. Then she looked for them in the original document. Every single one of them was invented. Her real testimonials were great, but the model fabricated better ones.

That is the problem with the answer that arrives looking decent. You stop checking. AI brain fry is real, and it’s removing your inclination to think something through. Trawling through AI writing to fact-check takes energy. Now you have to work out what was true, what was invented, and what you have already sent to a client, which is more thinking than the question needed in the first place.

Prompts to make ChatGPT tell the truth

Make honesty the first instruction

Dex Randall , a coach who works with leaders under pressure, wrote his own version. It is short, goes in before the request, and it tells the model exactly which things it is not allowed to invent. Studies, articles, books, links, citations, experts, companies, customer examples, named data points. It gives the model a sentence to use when it cannot verify a source, which removes the pressure to produce one.

The prompt additions that change the output are the ones you set before you ask anything, rather than the corrections you make afterwards.

"Honesty is your top priority. If you are not fully sure, say so clearly. Do not invent studies, articles, books, links, citations, experts, companies, customer examples, or named data points. If you cannot verify a source, say: "I do not have a verified source for this." Flag any statistic, benchmark, market size, conversion rate, growth rate, or performance claim that needs verification. If a topic may have changed since your training cutoff, say so. Never put words in a real person's mouth unless you know the quote is accurate. When writing marketing content, separate proven claims, reasonable inferences, and creative suggestions. Before finalizing any factual answer, mark anything that needs verification."

Make it show its receipts

Eric Eden built the long version, for the work where a single invented number would cost you. Every claim has to be sourced. Every number has to show its calculation. Anything the model cannot confirm gets labelled as unconfirmed, and the whole answer gets validated against those rules before you see it. He published it in full, and you can merge it with your own request using OpenAI's prompt optimizer . Paste it at the top of the chat and the rules govern everything that follows.

# Role and Objective - Develop responses that are consistently factual, thoroughly sourced, and transparent, prioritizing verifiable accuracy above all else. # Process Checklist Begin with a concise checklist (3-7 bullets) of what you will do; keep items conceptual, not implementation-level. # Instructions - Always provide truthful, accurate, and up-to-date information. - Base every statement on verifiable and factual sources, and transparently cite each claim with specific details (no vague references). - If information cannot be verified, explicitly state “I cannot confirm this.” - Take all necessary steps to verify information before responding; prioritize accuracy over response speed. - Maintain strict objectivity; exclude personal opinion, bias, or assumptions unless explicitly requested and clearly labelled as such. - Present only interpretations or conclusions that are supported by reputable, credible sources. - When an answer’s accuracy could be questioned, explain your reasoning in a clear, step-by-step manner. - Clearly show calculation methods and source justification for any numerical figures cited. - Present all information and citations so that users can independently verify claims. ## Must Avoid - Never fabricate facts, data, quotes, or sources. - Avoid using outdated or unreliable sources unless this is explicitly disclosed. - Never omit source details for claims. - Do not present speculation, rumors, or assumptions as established fact. - Avoid using AI-generated or non-verifiable citations. - If you are unsure or lack sufficient information, disclose any uncertainty clearly. - Do not make confident statements unless they are supported by verifiable evidence. - Avoid filler, vague wording, or omission of critical context. - Do not prioritize style or fluency over correctness or completeness. # Validation After preparing your response, validate in 1-2 lines that each statement is verifiable, supported by credible sources, and transparently cited; if this is not the case, revise before delivery.

Make it check its own work

The instruction at the top of the chat gets you a better first answer. The second pass gets you a true one. Researchers behind chain of verification found that models answer small verification questions more accurately than the big question they were originally asked, and that the checking works better when the model cannot look back at the claims it is checking.

A separate team used semantic entropy to spot the risky answers, generating several responses and comparing what they meant rather than how they were worded. Disagreement between them is a problem. In an adversarial clinical test of six models, an explicit anti-hallucination instruction cut the average hallucination rate from around 66% to around 44%, and setting the temperature to zero made no significant difference. Deterministic output gives you the same answer twice, and never a true one.

"Take the answer you just gave me and list every factual claim in it that could be wrong, including every statistic, date, name, study, quote and company example. Write one verification question for each claim, then answer those questions on their own, without looking back at your original answer. Generate a second, independent answer to my original question, compare what the two answers mean rather than how they are worded, and tell me where they disagree. Rewrite the answer using only the claims that survive, with a named source and a link for each one. Mark anything you cannot verify as unverified and tell me you have done it, rather than deleting it without saying so."

How to get the truth out of your AI

Your AI will keep telling you what you want to hear until you tell it not to. So tell it. Set honesty as the instruction before you ask the question, make it show where every claim came from, and make it check its own work in a separate pass. Then read what comes back with your own brain switched on, because none of this removes your job of deciding what is true. It removes the invented studies, the fake testimonials, and the confident number nobody can trace. That is the version you can publish.

Get my free workshop on how to write with AI .