Imagine asking Claude a question you don't know the answer to.
It responds immediately. The explanation is clear. The reasoning sounds logical. There is no hesitation.
You trust it.
Now imagine receiving the same answer, but this time Claude says:
"I'm not entirely sure, but I think this is the most likely explanation."
Would you trust it less?
What if both answers contained the same information? That's what I want to explore.
Let's try a simple experiment
Suppose you're troubleshooting an application that suddenly became slow. You ask Claude to identify the cause, and it gives you one of these responses.
Response A
The slowdown is caused by an inefficient database query. The query is retrieving more records than necessary, increasing execution time. Add appropriate indexes and optimize the query to resolve the issue.
Response B
One possible cause is an inefficient database query. It may be retrieving more records than necessary, increasing execution time. Check the query execution plan and database metrics before deciding whether additional indexes or query optimization are needed.
Notice something?
Response A sounds decisive. It identifies a cause and offers a solution.
Response B is more cautious. It proposes a hypothesis and asks you to verify it.
But there's an important detail: neither response has been given actual logs, query plans, database metrics, or profiling results.
Response A sounds as if the diagnosis is already established.
It isn't.
The application could be slow because of network latency, connection pool exhaustion, CPU pressure, locking, or several other reasons.
The database query is only one possibility.
The confidence of the answer is not evidence that the diagnosis is correct. And yet, I suspect many people would find Response A more reassuring.
Wouldn't you?
Confidence is a strange shortcut
Think about how we evaluate information.
When someone explains something clearly, speaks without hesitation, and presents a logical argument, we often interpret those signals as signs of competence.
Sometimes that works. Expertise can produce clear, decisive explanations. But confidence and correctness are not the same thing. An expert can be uncertain about a genuinely difficult problem. Someone with little expertise can sound absolutely convinced.
AI introduces an interesting version of this problem.
A language model can generate fluent, authoritative explanations even when the available information is insufficient to establish the answer.
The explanation can sound complete while leaving out the most important fact: whether the conclusion has actually been verified. And because the answer arrives in polished language, we may spend less time questioning it.
The danger isn't that every confident AI answer is wrong.
I want to test something
Rather than simply assuming that confident language makes people more trusting, here's an experiment worth running.
Take a set of technical questions, factual questions, or troubleshooting scenarios where the correct answers can be independently verified.
For each question, prepare two versions of the same answer:
- Version A: Direct and confident.
- Version B: Cautious and explicit about uncertainty.
Keep the factual content, supporting evidence, and recommended actions as similar as possible. Change the presentation, not the underlying evidence.
Then ask participants to evaluate the answers without telling them which version is intended to sound more confident.
Measure two things separately:
- Perceived credibility: How trustworthy does the answer seem?
- Actual accuracy: Is the answer correct according to an independent source or test?
If the confident version receives higher trust ratings without better accuracy, that would suggest that presentation influences perceived credibility independently of correctness.
If the ratings are similar, that is useful too.
And if the cautious version is trusted more, we should investigate why.
One important distinction: this experiment tests how people perceive answers. It doesn't establish whether Claude itself is more or less accurate when it uses confident language.
Those are different questions.
It's that confidence can become a substitute for checking.
But should AI always sound uncertain?
Not necessarily.
Imagine asking Claude what 15 multiplied by 12 equals. The answer is 180.
There is little value in surrounding that calculation with unnecessary hesitation.
Now imagine asking Claude to identify the cause of an intermittent production outage from three lines of a log file.
A definitive diagnosis would be difficult to justify. The amount of confidence an answer communicates should reflect the strength of the available evidence and the uncertainty of the problem.
That's the distinction I'm interested in.
I don't want an AI that says "maybe" before every sentence. I want one that can distinguish between what it knows, what it can reasonably infer, and what it still needs to verify.
And I want users who make that distinction too.
The responsibility isn't entirely on AI
We can ask AI systems to communicate uncertainty more clearly. We can design evaluations that reward accuracy rather than persuasive presentation.
But users have a role here as well.
When an answer sounds convincing, we should ask:
- What evidence supports this conclusion?
- What assumptions is it making?
- What information would change the answer?
- Can I independently verify the result?
This matters particularly when we use AI for decisions involving money, health, security, production systems, or other consequential situations.
A confident explanation can be a useful starting point. It should not automatically become the final authority.
And the more consequential the decision, the more important independent verification becomes.
Here's the question I can't settle
If two AI answers contain the same information, but one sounds certain, and the other acknowledges uncertainty, which one should we trust more?
My answer would be neither by default.
I'd rather trust the answer that gives me the strongest evidence, makes its assumptions visible, and tells me what remains unverified.
But there's a catch.
People don't always have the time, knowledge, or resources to verify every answer they receive. Sometimes, we have to decide how much to trust a system before we can independently check it.
That makes the way AI communicates uncertainty more than a writing preference. It becomes part of how people decide whether to rely on it.
So here's my question for you:
Have you ever trusted an AI answer because it sounded convincing, only to discover later that it was wrong?
And if an AI openly admitted uncertainty more often, would you trust it more, or would you simply stop using it?
I'd like to hear your experiences.
Top comments (0)