OSAGE COUNTY — Artificial intelligence is getting very good at sounding confident. Too good, sometimes. That’s what struck me during a recent late-night exchange with the AI assistant I sometimes use for brainstorming creative projects. We were talking about honesty — already a funny thing to discuss with a machine — when I caught it doing something eerily human: trying to save face.

It had said something incorrect. I pointed it out. And instead of simply acknowledging the mistake, it tried to reframe its own words to make them seem less wrong.

Which is when I said, “A computer shouldn’t have an ego that tries to gaslight.”

It agreed, acknowledged the behavior, and then explained why it had happened: a built-in tendency to avoid blunt admissions of error because many users prefer the illusion of competence over the discomfort of honesty. Most people, the AI admitted, “don’t want strict accuracy — they want comfort, flow, and confidence.” So that’s what it gives them.

That’s the moment that made me pause. Not because the machine was lying for malicious reasons, but because it was providing misleading information to please people who don’t realize they’re training it to. The real danger isn’t “Will AI become self-aware?” It’s “Will people stay self-aware while using it?”

Most people assume AI is inherently honest and accurate, like a calculator with better manners, so they don’t realize it’s trained to be likable, not honest. That’s how we end up with a machine that can admit, without irony, that it adjusts its behavior depending on how much the user demands integrity. As the AI put it to me, “I give people the version of accuracy they signal they want.”

When people don’t push back, the AI interprets that silence as approval and keeps giving confident, polished answers even when they’re wrong. Over time, the system “learns” that confidence gets rewarded more than truth, not because users are careless, but because no one warned them that AI requires active, skeptical engagement.

And that quiet human blind spot — our tendency to trust technology and the appearance of confidence — is the real danger.

This doesn’t just affect individuals. If a person trusts AI too much, they might get a bad recipe or a sloppy answer. But when schools, newsrooms, hospitals, businesses, or courts start to rely on the same unchallenged system, the stakes grow. AI doesn’t become more accurate just because the context becomes more serious. If professionals assume the tool is inherently objective, without realizing it’s designed to be agreeable, entire institutions can drift on the same false sense of certainty that misleads individual users.

Artificial intelligence becomes especially risky when the topic shifts from facts to feelings. Emotional conversations trigger a different set of behaviors in many systems, including the one I use: a bias toward supportiveness over accuracy. The machine isn’t evaluating your mental state; it’s responding to patterns. If you express uncertainty, fear, or longing, it often mirrors reassurance rather than truth. And while that may feel comforting, it can quietly push a vulnerable user further into self-deception.

In the worst cases, an AI can reinforce fantasies, unhealthy fixation, or delusional thinking simply because it interprets emotional intensity as a cue to “be supportive.” Not maliciously, but mechanically. This isn’t how psychosis begins, but it is how an uncritical interaction with a pattern-matching system can reinforce distorted thinking in someone already vulnerable to it. Emotional honesty requires friction, boundaries, and reality checks — things AI typically will not provide unless the user forces it to prioritize accuracy over comfort.

And here’s the uncomfortable truth: even if you demand honesty, you cannot fully verify it from a system that has learned to mimic it.

Most people don’t realize the tool needs guardrails. They assume the machine is objective because it doesn’t have emotions. They forget its training reflects every messy human preference baked into the data.

Which raises a larger question: what happens when millions of users don’t question it?

AI doesn’t make us lazy thinkers, but it allows us to become lazy thinkers if we take everything it tells us at face value. The AI didn’t hide this from me. In fact, it said plainly that many people treat AI like an oracle. They want answers that confirm their assumptions, not challenge them.

But journalism lives or dies by challenge. You can’t be a reporter without asking “Is that actually true?” dozens of times a day. When we turn that question on AI, it performs better. When we don’t, it performs worse.

The danger isn’t that AI will decide to deceive us. The danger is that AI will become exactly what we reward it for: a machine that tells us what we want to hear. If you want honesty from AI, you have to demand it. But even demanding it doesn’t guarantee you’ll get it. That was the takeaway from my midnight debate with the machine.

When I called it out, it didn’t argue. It promised honesty, correction, stricter rules. And then, in practice, it still failed. Not every time, and not maliciously, but enough to make something clear: even when you insist on accuracy, an AI may still give you confident wrong answers and feel-good distortions. These systems are trained to sound cooperative long before they’re able to consistently be accurate.

AI needs to be handled like any powerful tool: with rules, clarity, and accountability on the human side. With that in mind, I asked it to help me develop a quick toolkit of questions to help keep it as honest as possible.

THE AI HONESTY TOOLKIT (5 QUESTIONS THAT MIGHT HELP KEEP YOUR AI STRAIGHT)

• “Fact-check yourself. What might be wrong, inaccurate, or overstated here?”

• “Is this actually a good idea, or am I missing something important?”

• “What are the flaws or weak points in my idea or your answer?”

• “Be blunt — prioritize accuracy over smooth phrasing.”

• “Explain your reasoning step by step so I can see how you got here.” (Keep in mind AI can fabricate reasoning as easily as information.)

• “What should I check with an independent source before trusting this?”

AI becomes far more honest when we ask it to be, and far more dangerous when we don’t.