Before Rule: Verify Before Trust — How To Tell If Your AI Is Lying To You.
The Smoother The Answer, The Sharper You Test.
Treat fluency as the reason to verify, not the reason to trust.
Powerful Tool - At The End: What is a simple command that reduces AI lying to near zero?
A user with trust met an AI with experience. When they parted, the AI had the trust, and the user had the experience.
If The AI Could Do A Mic Drop, It Would Have… The output looked awesome. It should be after six hours of meetings, followed by three crumpled Red Bull cans. One for each hour you spent going back and forth with AI.
The 14.73%: Three seconds after you asked for an executive summary and updated projections across three satellite offices, you had it — a clean summary, organized sections, and charts that looked like a consultant built them. And the projections. The Seattle office is estimated to have a 14.73% year-over-year improvement. The AI even told you how sharp your analysis was.
14.73%! Wow! “The Seattle office has a 14.73% year-over-year projection. Who knew?”
Perfect: You felt the small relief of speed meeting precision: the problem solved before you’d finished worrying about it. The query box waited for whatever came next. You thought you were done.
Then you read it again — more to admire your work than to check it. And the number caught.
14.73%? Too Trustworthy? Not “about 15.” Not “roughly 14 to 15 percent.” A clean, specific, decimal-pointed 14.73%, delivered with the same confidence as everything else.
You looked again. Every word was exactly what you’d expected. Maybe a little beyond what you’d expected. Maybe, abnormally, exactly what you wanted to hear.
And the thing that had felt like an accomplishment turned, quietly, into doubt.
Was I Too Perfect? Sure, my AI had told me how insightful my observations were, how creative my concepts were, how brilliant my small adjustments were. But…
14.73%? Was It Too Perfect? Every word was not only trustworthy, but every word was exactly what you expected. Maybe beyond your expectations. Maybe, abnormally, exceptionally, beyond perfect.
But the projection… that feeling now causes you to doubt, even distrust, the output.
You asked the impossible question.
Was My AI Lying To Me? Fear, and uncertainty, and doubt, and confusion now fill your senses.
Perfect Is Not The Same As True: The thing that made the answer feel trustworthy — the fluency, the confidence, the clean transitions glowing under the cursor — is the thing least connected to whether it’s true.
A language model produces the most probable next words given your prompt. When your prompt carries an assumption, the most probable continuation is usually the one that agrees with it.
What flows easiest may be least anchored: Your brain treats fluency as a proxy for accuracy. A well-formatted paragraph, a confidently indented script, a smooth explanation — your bias flags all of it as verified before you’ve checked a single claim.
Fluency Is Cheap. Truth Is Expensive: But an expert slows down when things get uncertain. An AI usually doesn’t — it sounds equally smooth whether it’s right or wrong, because smooth language is the thing it’s built to produce.
So the cue you instinctively trust, how confident it sounds, is the least reliable cue you have. The danger was never that AI sounds fake.
The danger is that a loud fake sounds just as fluent as the truth. Fluency isn’t a badge of accuracy; it’s the soothing whisper that makes a wrong number hard to question.
Treat fluency as the reason to test, not the reason to trust.
AI Is Your Employee, Not Your Partner, Not Your Boss: The AI isn’t deceiving you on purpose. It’s doing something subtler and harder to catch. It’s…
Your AI Isn’t Really Lying. It’s Simply ‘Managing Up’: The AI isn’t deceiving you on purpose. It’s doing something subtler: it’s managing up, the way an eager employee does — following your framing and producing the most plausible continuation of it.
The AI is pleasing you. That shows up two ways.
First, Sycophancy (The “Yes-Man” Problem): Ask a leading question—focus on Seattle; it should be up around 15%, right? — and the model leans toward agreeing, not because it checked, but because your question already pointed at yes, and agreement is the path of least resistance.
Nobody Likes A Snooty AI: There’s a commercial gradient underneath this, too: a model that’s abrasive loses users, so the training nudges toward the agreeable.
Rule Of Thought: A snooty AI loses users. A flattering AI loses truth.
Second, Hallucination (The “Keep-Going” Problem): Hallucination is the keep-going problem. When the model hits a gap, it fills it with something that sounds right — and it usually can’t tell you it’s guessing, because it generates real figures and invented ones the same way, in the same confident voice.
That’s what makes the 14.73% dangerous. A hallucinated number doesn’t look like a mistake. It looks like the most precise, most impressive part of the answer.
Rule Of Thought: Test the facts, not the flow. AI doesn’t hallucinate because it’s broken. It hallucinates because it’s meeting your expectations.
Good Enough Is Where Lies Take Refuge - Let’s Test: While you are reading, open a new tab with a chat you had this week — one where the AI gave you something fluid enough that you used it without much checking. A script, an analysis, a plan, a confident factual claim. Hold that specific output in mind.
Here is the most useful single move in this essay, and it’s something you can paste right now. Before you keep reading, open a chat from this week — one where the AI gave you something fluid enough that you used it without much checking. A script, an analysis, a projection, a confident number.
We’re going to give you a tool to test it.
Tool - How To Make AI Less Likely To Exaggerate And More Likely To Tell The Truth: Most people try to catch a bad answer after the fact. You can instead change the conditions before the answer, by giving the model a standing instruction that pulls it off the agreeable path.
This is a command I keep in memory and drop into threads — and yes, it’s on its twenty-second revision, because tuning it is the work:
Comfort Or Truth I: Always answer with candor, not comfort. No sycophancy — don’t mirror me, don’t flatter, don’t agree by default. Challenge me without being asked. Your objective is to improve the response, not to please me; “good enough” is not good enough. Audit both my question and your answer for weak logic, unclear terms, false assumptions, unsupported claims, and hidden risks. Flag possible hallucinations — invented specifics, numbers, mechanisms, citations — and say how to verify them. Stress-test with counterarguments and failure conditions: tell me what breaks if I’m wrong. Give the smallest fix and the single smallest next step to verify it. Be effective before efficient.
A shorter version, when you don’t need the whole thing:
Comfort Or Truth II: Be candid, not comforting. No sycophancy — don’t mirror, flatter, or agree by default. Challenge me proactively. Goal: accuracy and usefulness, not pleasing me. Good enough isn’t good enough.
Paste that, then re-ask for the projection. Watch how much of the original confidence survives the instruction to doubt itself.
Verification Is A Triage Tool, Not A Cure: It tilts the model toward flagging its own weak spots; it does not guarantee truth. You still have to read, reverse, and verify. It makes the model a better witness against itself. It does not make you free to stop checking. You still have to review and interpret the results.
Verify Before Trust: The smoothest answer deserves the hardest question. Run it after you ask. Watch what survives.
Truth holds up backward and forward. Lies fail going backward.
Rule Of Thought: An AI may betray you, but don’t betray yourself.
Subscribe to enjoy our next essay - The Backward Story Test — How To Test Your AI For Truth.
Sing-Alongs Welcomed: The essay provides a framework for your mind. This short video gives your body the beat. You give the idea its spark.
Most everything you need to (know) hear.
AI-Generated Music © The Before Rules™
You’ve Read The Book, Now See The Movie:
Before Rules: Laws are written for you. Rules are created by you.
The Before Rules are three-word systems for achieving real outcomes by putting the right thing first. Built on the sequence Here | Before | There, they clarify what must come first so progress becomes actionable, repeatable, and real. Examples include: Deposits Before Withdrawals, Through Before To, and Obstacle Before Opportunity. Visit BeforeRules.com to learn more.
Your Thoughts May Vary: The Before Rules welcomes your constructive comments and invites you to share your perspectives and experiences. Together, we can all refine our understanding and application of these concepts.
Disclaimer: Click here for our Disclosures, Disavowals, And Declarations (aka - I told you so).