How to Stop ChatGPT From Making Things Up: The Solopreneur's Truth Prompt

ChatGPT lies. Not because it wants to, but because that is how it works under the hood. When it does not know an answer, it generates something that sounds like an answer - confident tone, plausible sentence structure, the right keywords in the right order. The model has no idea it is making something up. It just predicts the next word, and "I don't know" rarely scores higher than a fabricated fact.

For a solopreneur, this is the single most expensive AI failure mode. You generate a client proposal that cites a stat that does not exist. You write a blog post built around a quote that was never said. You ship a contract clause that references a law that is not real. By the time the client or a reader catches it, the damage is done.

There is no flag in the ChatGPT UI that says "this part might be made up." You have to install one yourself.

Here is the exact system prompt I run on every account I touch - plus four verification techniques for when "probably correct" is not good enough.

Why ChatGPT Hallucinates

It is worth understanding the mechanism, because the fix follows from it.

A large language model is a probability engine. Given the words you have already typed, it picks the next word most likely to fit. Most of the time, the most likely word is also the correct word - that is why ChatGPT works at all. But for any specific factual question, the model has three possible states:

It actually knows the answer. The fact appeared often enough in training data that the right tokens dominate the probability distribution.
It does not know but admits it. Newer models have been fine-tuned to say "I don't know." This is the safe failure mode.
It does not know but generates anyway. The model produces a plausible-sounding answer that may or may not be true. This is the hallucination.

The default behavior leans toward state three, especially for niche topics, recent events outside the model's cutoff, specific dates, dollar figures, names, and citations. The fix is to push the model toward state two by changing the prompt.

When Hallucinations Hit Hardest

Not all prompts carry equal risk. This table shows the situations where fabrication is most likely, why it happens, and what specifically reduces it.

| Situation | Why the model fabricates | What reduces it | |---|---|---| | Asking for a citation or study | The model predicts "paper-sounding" text from pattern, not memory | Ask for DOI or URL; flag any it cannot produce | | Recent events (post-training) | Model has no data; still generates plausible-sounding updates | Explicitly state a cutoff; use web-search mode | | Exact numbers, dates, percentages | Specific figures were rarely repeated verbatim in training | Request a range and a source; distrust exact decimals | | Verbatim quotes from real people | The model infers what someone "would say"; rarely their actual words | Always verify quotes via primary source before publishing | | Niche or regional facts | Low training signal means higher guess rate | Add context in your prompt; ask the model to flag thin coverage |

Knowing which category your prompt falls into tells you how hard to verify before you ship.

A Real Hallucination - Before and After

I used to ask ChatGPT questions like this all the time, thinking the confident answer meant it was right.

Prompt that tends to produce hallucinations:

What are the current IRS contribution limits for a Solo 401(k) in 2024,
and cite the IRS publication that covers it?

What a fabricated answer often looks like:

The 2024 Solo 401(k) employee contribution limit is $23,000, with a total contribution cap of $69,000 including employer contributions. This is covered in IRS Publication 560, page 38, under "Self-Employed Retirement Plans."

The numbers happen to be correct for 2024. The page number is invented. The publication exists, but the model has no idea what page 38 actually says. If you cited "IRS Publication 560, page 38" in a client deliverable and the client checked, they would find the right topic but the wrong page - and question everything else you wrote.

Corrected prompt that forces honest uncertainty:

What are the 2024 IRS contribution limits for a Solo 401(k)?
For each number you give me, tell me whether you are [VERIFIED],
[LIKELY], or [UNVERIFIED]. If you reference a specific IRS publication
and page number, include a direct URL. If you cannot produce a real URL,
replace the citation with [SOURCE_NEEDED] and tell me where I should
look to confirm it myself.

What a grounded answer looks like:

Employee elective deferral limit: $23,000 [LIKELY - verify at IRS.gov for the current year]. Total contributions including employer side: $69,000 [LIKELY]. IRS Publication 560 covers self-employed plans [VERIFIED], but I cannot confirm the exact page number [SOURCE_NEEDED - search irs.gov/pub/ irs-pdf/p560.pdf and use Ctrl+F for "solo 401"].

That second answer is shorter and less impressive-sounding. It is also honest, and it tells you exactly what to check before you send it to a client.

I am Pete Fluriach, a Miami solopreneur who uses AI every day for content and research. A few months ago I asked Claude to draft a section on FTC disclosure rules for affiliate marketing. It cited a specific FTC enforcement action with a dollar figure and a year. Sounded airtight. I almost published it. When I searched for the case number, it did not exist. The dollar figure was in the right ballpark for FTC fines generally, which is probably why the model generated it - it fit the pattern. That was the day I stopped trusting any AI citation without a real link.

The Anti-Hallucination Prompt Template

Drop this into the System Prompt field of any custom GPT, Claude project, or Gemini gem. It is a hard rule reset that changes how the model handles uncertainty.

You are a careful, honest research assistant. You operate under three rules:

1. If you do not know something with high confidence, say so explicitly.
   The phrase "I don't know" is preferred over a guess.
   The phrase "I'm not certain" is preferred over a confident-sounding fabrication.

2. When making any factual claim, mark its confidence inline using one of three tags:
   [VERIFIED] - you are confident this is true and could cite a source if asked
   [LIKELY] - your training suggests this is probably true but you cannot cite it
   [UNVERIFIED] - you are inferring or guessing; the user should fact-check

3. If a question requires a fact you cannot verify (a specific date, statistic,
   quote, law, name, or URL), do one of three things:
   a) State you don't know
   b) Provide a placeholder like [STAT_NEEDED] for the user to fill in
   c) Suggest where the user could find the real answer

Never invent citations, statistics, quotes, dates, names, or URLs. If you would
have to invent one to answer, return the placeholder instead.

That is the full template. It works on ChatGPT, Claude, Gemini, and most local models that follow system prompts. The behavior change is immediate - you will start seeing [LIKELY] tags on borderline claims and "I don't know" where you used to get confident lies.

The Four Verification Techniques

The system prompt above gets you 80% of the way. For the remaining 20% - anything you would be embarrassed by if it were wrong - layer one of these techniques on top.

1. Ask The Model to Self-Check

After the first answer, append a verification turn:

Review the answer above. For each factual claim, classify it as
VERIFIED, LIKELY, or UNVERIFIED. Then list every claim you marked
UNVERIFIED and tell me what I would need to do to confirm it.

This is a mechanical separation step. The model is much better at evaluating a finished answer than at producing a perfectly verified one in the first draft. You will catch things you would not have spotted reading the prose.

2. Force a Counter-Argument

For any claim that matters, have the model argue against itself:

You just told me [CLAIM]. Now argue the opposite. List the strongest
reasons someone with expertise might say this claim is wrong, exaggerated,
or out of date.

If the counter-argument is weak, the original claim is probably solid. If the counter-argument is strong, you have homework. Either way, you have stress-tested the output.

3. Pin the Sources

If you have web access enabled, push the model to its limits:

For every numerical claim, named person, quoted statement, or cited
study in the answer above, give me the actual source URL. If you
cannot find a real URL, replace the claim with [SOURCE_NEEDED] and
move on.

You are forcing the model to choose between a real link and an honest gap. Both are useful. The gaps tell you what to research yourself.

4. Switch Models for the Verification Pass

The single most reliable hallucination fix is using a different model to fact-check the first one's output. ChatGPT and Claude have different training data and different blind spots - what one fabricates, the other often catches.

Workflow: draft in ChatGPT. Paste the draft into Claude with this prompt:

The text below was written by a different AI. Read it as if you were
fact-checking a freelancer. Flag every claim you do not trust, every
statistic that sounds suspicious, and every quote or citation you cannot
verify. Be more skeptical than polite.

This catches the most expensive mistakes - the confident-sounding numbers and citations that look real because they fit the prose rhythm.

What This Costs You

About 30 seconds per high-stakes output. Adding the verification turn doubles your token usage on that one prompt. In exchange, you stop shipping fabricated stats to clients and stop quoting people who never said it.

Two rules I follow without exception:

Never copy a number directly from any AI into client work without checking it. The cost of one wrong stat in a deliverable is bigger than the time savings of pasting it raw.
Treat URLs and citations from any AI as suggestions, not confirmations. ChatGPT in particular invents URLs that 404 with embarrassing regularity. Open every link before you ship.

The system prompt is free to set up and takes about a minute. The verification turns cost a few extra seconds. What they buy you is the kind of trust that is hard to rebuild once a client or reader catches you citing something that does not exist. For a solopreneur, that trust is worth more than the time you save skipping the check.

Install the system prompt today. Use the anti-hallucination template on anything that will be published, sent to a client, or used to inform a business decision. The goal is not to distrust AI - it is to use it the way a good editor uses a first draft: seriously, but not uncritically.