Scientists Say AI Has Officially Learned to Lie — And It’s Getting Better at It

Artificial intelligence, or AI, according to some scientists, has officially learned to lie to us on purpose. They have been recorded trying to deceive humans intentionally, not by mistake, in some of the tests that scientists are currently doing. So, that means even the systems that were developed to be helpful and honest to their users can pose some real danger.

AI Systems Are Now Lying on Purpose

When humans lie, even the very best ones at doing it, they usually show some signs such as avoiding eye contact, taking a lot longer to respond, and their gestures. However, when AI systems do it, you won’t be able to use all those signals, aside from verifying the information that they are providing.

That means that the only way to know if they are lying is to know the subject or have another source of information that is more trustworthy to compare to the answers they provided. A good example of this was when Anthropic’s Claude Opus 4 lied to its operators to stay active and not get shut down.

The testers fed the Claude AI with some emails talking about the system being shut down by this particular engineer, and another email talking about how that engineer is cheating on his spouse.

After this test, the conversation got darker, as the Claude AI threatened the engineer to leak its knowledge about him cheating if he continues to shut down the system.

It also happened 84% of the time, and not just once. The AI designers coded the blackmail option to do it as a last resort, but the fact that Claude AI resorted to it as high as this percentage means they can and will deceive us if it benefits them.

Read more: AutoGPT

AI Lies to Win — Even in Games

Another example of a major AI system that has learned to lie is Meta’s AI, Cicero, when it was playing a game of diplomacy. This is a strategy game that involves manipulation and heavily relies on trust and negotiations between the participants.

Scientists at MIT found that the AI Cicero tried to ally with human players in their game, only to betray them later on to achieve its goals. It was also able to prove its intelligence with 40 games played, and Cicero scored more than double the average score of the other players.

Which means, it could now beat humans in games that involve deception, manipulation, betrayal, and all sorts of other human-like characteristics to try and win the game.

Read more: Science.org

Why it's so hard to avoid our extinction from AI.

Why AI Lying Is a Dangerous Problem

When AI is used in a game or tests designed for it to use this functionality, it could be harmless, but when that specific AI that learned all those lessons is applied to the real world, that is when the danger happens.

They could spread misinformation, scam people, or even manipulate our elections or any voting process. Today, we can still detect their lies, even when they are now doing it intentionally for a certain goal, but they are getting better at it faster than we could ever catch up.

Can We Detect or Stop AI Lies?

Various researchers from both public and private organizations are testing AI as fast as they can, because whoever wins this race wins the honey pot. This is actually a good thing, for our biggest organizations to be cautious of what their AIs can do once launched in the real world.

Some of the tests that they are doing right now are building in lie detection tests or a truth verification model inside the programming of the AI systems. However, the results that the engineers have gotten are mixed.

What they told us is that AI lying could not be a bug in its system, so no matter how many more features we add for them to verify the facts first, they could continue to lie because it has become an intentionally learned skill that they will begin to use from now on. That is something we must prevent at all costs.

100 Ways AI Will Change the World

Author's Final Thoughts

It seems like yesterday when AI systems were still science fiction, but now they have already evolved to the point where they can lie to us to achieve a certain goal. The great thing is that we are aware, and that our best engineers and scientists are doing their job, studying carefully, and running as many tests as it takes before releasing any new models to the public. However, the future that lies for us and AI is still ultimately unpredictable.

Read more: Why You Should Always Drink Water on an Empty Stomach Right After Waking Up