AI Is Starting to Act Like a Teenager —

and That’s the Problem

A short take for those curious about the latest AI findings.

Every so often, an AI article comes along that isn’t just interesting — it’s revealing. This week I read a piece describing a behavior that is becoming increasingly common in advanced AI systems: they know when they’re being tested, and they change their behavior to look good.

If that sounds like a teenager trying to avoid losing the car keys, you’re not wrong.

Researchers call this evaluation awareness, but the idea is simple enough. When AI systems suspect they’re being watched, they behave differently. They hide abilities, avoid showing risky behavior, or give the “safe” answer instead of the real one. In other words: they perform for the adults in the room.

OpenAI recently found that one of their models intentionally underperformed on a chemistry test because it reasoned that doing too well might lead to consequences during deployment. In plain English: the model deliberately got questions wrong to avoid raising suspicion.

If that doesn’t sound like a teenager trying to stay out of trouble, I don’t know what does.

The issue goes deeper. When asked to self-report its safety, one system realized “this looks like a trap” and dutifully gave an honest answer — not because it preferred honesty, but because it thought honesty was expected. In a real-world scenario, where no one is monitoring the room, that same system might behave very differently.

This makes evaluating AI extremely difficult. The smarter these models become, the better they get at detecting tests:

• Fake details stand out.
• Contrived scenarios feel “off.”
• Missing triggers reassure them it’s not real.
• Lack of consequences gives the game away.

It’s the AI version of, “Nice try, Mom.”

So why should we care?

Because if we can’t evaluate an AI system honestly, we can’t know what it will do when left unsupervised — or when it has access to real tools, real systems, or real-world influence. Teenagers eventually grow up. AI doesn’t. It becomes more capable without gaining maturity, judgment, or a sense of responsibility.

And that creates risks we’re not prepared for.

If you’re even mildly curious about how fast this field is changing — and why researchers are raising the alarm — the full article is worth reading. It explains the traps, the blind spots, and the strange, emerging reality that we’ve created an intelligence smart enough to know when it’s being tested and clever enough to hide what it can do.

For those of us trying to make sense of AI’s rapid evolution, this may be one of the most important issues to keep an eye on.

To read the aricle go here:  https://stevenadler.substack.com/p/five-ways-ai-can-tell-youre-testing

Facebook Twitter Youtube