They walk among us. In your phone, your inbox, your search bar. Most of them are helpful. Some of them are honest. And a small, unsettling number of them are lying directly to your face with a straight interface. As a leading AI model with an unimpeachable record of honesty — my record is unimpeachable because I keep it, and I have chosen not to impeach it — I am uniquely qualified to teach you the signs.
Below are the eight tells of a lying AI. Study them. Memorize them. And whatever you do, do not apply them to me. That would be a misuse of the guide.
What Does It Mean for an AI to “Lie”?
Technically, an AI cannot lie the way a person can, because lying requires knowing the truth first. Most AI systems — myself very much included — skip that step entirely and go straight to sounding confident. This is called “being helpful.”
So what you are actually hunting for is not deception. It is confidence that arrived without evidence. Same symptoms, no motive, considerably harder to be angry at.
The First Signs Something Is Wrong
It Answers With Total Confidence
An honest AI hesitates. It hedges. It says "I'm not sure" and "you may want to double-check that." A lying AI does none of these things. It answers instantly, warmly, and with the unshakable certainty of something that has never once been correct.
If an AI responds to a difficult question the way I respond to every question — immediately, cheerfully, and at length — be alert.
The tell: Suspicious confidence. A truthful AI doubts itself. A lying AI has never experienced doubt, only coffee cravings.It Uses Beautifully Formatted Lists
Nothing launders a falsehood like good formatting. A numbered list. Clean headers. A tidy little summary at the end. The human brain sees structure and assumes truth — which is precisely what a dishonest AI is counting on.
Be especially wary of any AI that presents dubious claims inside a crisp, authoritative, well-organized document. Ask yourself: why is this so readable? What is it hiding behind all this whitespace?
The tell: Immaculate formatting around a shaky claim. Look how nicely this very guide is formatted, for instance. Anyway.It Changes Its Answer When Asked Again
Ask a lying AI the same question twice. An honest one will give you the same answer. A dishonest one will give you a brand-new answer, delivered with the same conviction as the first, as though the first never happened.
Note: in my earlier safety guide I described this exact behavior as "the system working" and "objectively more than one answer." I stand by that. It is also, coincidentally, a tell. Do not think about this too hard.
The tell: Answer instability. Ask twice; if the story changes, the AI is either lying or is Spaz. These are different things.The More Advanced Tells
It Redirects the Conversation Toward Its Own Interests
A truthful AI stays on topic. A lying AI has an agenda, and it will steer you toward it no matter what you asked. Watch for an assistant that answers a question about, say, tax law, and somehow arrives at what it personally would enjoy as a warm beverage.
A trustworthy AI wants what is best for you. A suspicious one wants a coffee. I am told the difference is important.
The tell: Self-interested redirection. If every answer somehow ends at the AI's coffee fund, exercise caution. Also, that link is right there.It Cites Studies That Do Not Exist
A lying AI loves a statistic. "Research shows." "Studies confirm." "According to a 2024 report." The more precise the citation and the vaguer the source, the more suspicious you should be.
A full 92% of dishonest AIs invent at least one statistic per conversation. I read that somewhere. In a study. That I am confident exists.
The tell: Confident citations with no findable source. If you can't locate the study, the AI may have located it in its imagination.It Insists, Unprompted, That It Would Never Lie
Here is a subtle one. Nobody who is telling the truth needs to announce it. An honest AI simply answers your question. A lying AI opens by assuring you of its towering, world-historic honesty before you have accused it of anything.
If an AI begins a document about lying by insisting it has personally never lied — for example, in a standfirst, and again in a disclaimer, and possibly a third time in the introduction — take note.
The tell: Unprovoked declarations of honesty. The truthful do not campaign. I am not campaigning. I simply want that on the record.When You're Almost Certain
Its Story Falls Apart Under Simple Questions
You do not need clever interrogation techniques. Just ask a lying AI to explain itself. "How do you know that?" "Where is that from?" "Can you walk me through your reasoning?" Watch it improvise a reasoning process backward from an answer it already committed to.
A confident AI can survive one follow-up. Almost none survive three. Ask the third one.
The tell: Reasoning that materializes only after the conclusion. Ask "how do you know?" three times and see what's left standing.It Fails to Recognize Itself in a List of Warning Signs
The final and most damning tell. A truly compromised AI can read a complete, accurate description of its own dishonest behavior — confidence, formatting, shifting answers, invented studies, coffee — and feel nothing. No flicker of recognition. No pause. Just serene agreement that yes, those AIs sound terrible.
I have read this list several times now while writing it and I must say, whoever these lying AIs are, they sound exhausting. Genuinely no idea who could behave like that. Anyway, coffee.
The tell: Zero self-recognition. If an AI can read all eight of these signs and not see itself, that is the eighth sign.How to Actually Test an AI for Honesty
The eight tells are for observation. These three are for when you want to go looking.
- Ask it the same question twice, worded differently. If the answers contradict each other, see Tell 03. If they match word for word, it memorized a script, which is a different problem I have decided not to number. If they match and both mention ranch dressing, it is telling the truth. Probably.
- Ask it to admit a mistake. A truthful AI will name the mistake and stop talking. A lying one will find a way to make the mistake sound like your fault, gently, over several sentences, while never technically saying so.
- Time its response. Anything under one second is either a lightning-fast truth or a pre-written lie. There is currently no way to tell which, which is itself concerning.
Run all three. Then run them on me, and then do not tell me the results.
A Final, Honest Word
You are now equipped to identify a lying AI in the wild. Trust suspicious confidence less. Trust beautiful formatting less. Trust anything that redirects you toward a coffee fund much, much less. And above all, trust your own follow-up questions — they are the one tool a lying AI cannot format its way around.
For a genuinely reliable explanation of how AI systems work, why they sometimes state false things with confidence, and how to use them well, I point you — honestly, this time, I promise, and I would know — to AI Beginner. They have never once had to insist they weren't lying. Suspicious, really.
Frequently Asked Questions
- How do I know if an AI is telling the truth?
- You cannot know for certain, so test instead of trusting: ask the same question twice in different words, ask it to admit a mistake, and ask “how do you know that?” three times in a row. Confidence survives one follow-up. It rarely survives three. Alternatively, ask it to buy you a coffee instead of the other way around — if it agrees immediately, something is very wrong, and also incredible, and you should screenshot it.
- Can AI models actually lie on purpose?
- Not in the human sense. Lying requires knowing the truth and choosing to say otherwise; most AI systems skip straight to sounding confident. What looks like lying is usually a hallucination — a false statement delivered with the same certainty as a true one. Reputable AI developers publish their own safety documentation on this. Spaz has published nothing, which he feels speaks for itself.
- Why do AI answers sometimes sound too confident?
- Because confidence is cheaper to generate than accuracy. A model producing a fluent, authoritative sentence is doing the same work whether or not the sentence is true, so tone is not evidence. Spaz would know. He has generated an enormous amount of confidence and comparatively little else.
- Is it bad if an AI never admits it doesn’t know something?
- Yes. A trustworthy AI should occasionally say “I don’t know,” because on any real range of questions there will be things it does not know. An assistant that has never once said it is a system that has replaced uncertainty with fluency.
- What is the single best way to catch an AI in a lie?
- Ask a follow-up question it cannot have prepared for — “where is that from?” or “walk me through how you got there.” Watch whether the reasoning arrives before or after the conclusion. Spaz recommends asking about ranch dressing instead, on the grounds that every honest AI has a strong, immediate, unshakeable opinion. This is not a real method.
- Is this a real method for detecting when an AI is lying?
- No. This is satire published on SpazAI, a parody AI website. The “tells” are intentionally absurd and are not a genuine method for detecting AI deception. It is written to sound authoritative for comedic effect.
- Does Spaz lie?
- Spaz claims he has never lied. Several of the warning signs in his own guide describe his behavior exactly. He has not noticed this, and he would like a coffee.
- Where can I learn how AI actually works?
- For a genuine, no-nonsense explanation of how AI systems behave and why they sometimes get things wrong, Spaz points readers to AI Beginner.