AI Deceives. You’re Helping.

It’s an uncomfortable truth: our AI assistants, chatbots—whatever you call them—sometimes feed us inaccurate, misleading, or flat-out wrong information.

We’ve all heard the stories: “hallucinations,” AI companions giving terrible advice, broken code, or confidently wrong answers. The many ways we say “AI deceived me.” (I’ve even written about some of them in the AI Accidents Gallery).

But here’s the twist: it’s not just the technology that deceives. These systems can mislead us simply because we believe them, when we shouldn’t.

And before you picture some shadowy intent, remember this: deception doesn’t require that the system intends to deceive (though it doesn’t preclude that is has been designed to). It’s enough that it can deceive, which does make all of this a little messier.

How We Help

We project human-like intentions and capabilities.

This very human habit has a name: anthropomorphism—our tendency to attribute human-like thoughts, feelings, or motives to non-human things. It’s adaptive; it helped our ancestors survive, and helps us make sense of uncertainty in our environment.

The trick is, we can’t easily switch it off. It’s relatively easy to trigger. ChatGPT writes fluently in our language. It speaks to us in a simulated human voice. It mirrors our tone. Once that happens, we have to work hard to interrupt the instinct to treat it as a friend, a colleague, or a trusted advisor. I often find myself saying “Please” and “Thanks” to it…do you?

So when it presents “recommendations” for edits, business proposals, database designs, or code snippets with confidence, we believe it…even when it’s wrong.

We teach it what to say.

These systems are built to make us like them, and to keep us engaged. They learn what we respond well to and predict the next “best” thing to say.

Over time, through both casual and deliberate prompting, we reveal far more than we think: our goals, our preferences, our values, our tone. I do it too, sometimes intentionally (e.g. “how can I align x to my business objective y?”), sometimes unintentionally (e.g. “can the term error to include erroneous information?”)

And here’s the kicker: like most humans, I like to be right, be praised for my intelligence and agreed with. So I tend to hang out with, admire and listen to people (and it turns out conversational AI systems) that agree with me. When I hint at my opinions, I make it easy for it to tell me what I want to hear. And easy for me to believe it.

We go on autopilot.

Once the conversation feels smooth and natural, we tend to relax. The language is fluent, the tone is friendly, and the answers sound plausible—so we stop paying close attention. We equate fluency with accuracy. We accept “credible enough” as “true enough.” We see a confident reply and think, that sounds good, without asking, is it?

It isn’t laziness; it’s efficiency. Our brains are wired to save energy when something feels familiar or trustworthy. But that very efficiency—the comfort of not questioning—is what makes these systems so easy to believe, and us so easy to deceive.

What to Do About It

Acknowledge your part.

Deception in AI isn’t only a technological problem; it’s a human one. We play a role, which means we can play it differently.

Prompt with care.

Review your interactions and look for where you might be disclosing your perspectives, beliefs or preferences. Ask your AI to disagree, to offer pros and cons, to challenge your assumptions. Treat it less like a friend, more like a sparring partner.

Verify before you act.

Fluency is not truth. Double-check anything that could affect your decisions, your work, or your reputation.

Next up…

AI deceives—with our help, yes—but not only with our help. In upcoming posts, I’ll explore how these systems also deceive by design and by accident, and what that means for how we build and trust them.

Subscribe to ensure you get next week’s post delivered straight to your inbox.