No, this is not another sci-fi story, but a very real, somewhat chilling reality about where AI development is headed – and with it, where our naivety stands.
Cornelia C. Walther, an expert researching the social impacts of artificial intelligence, draws our attention in her Psychology Today publication to a world where not only do we manipulate machines – but they manipulate us as well.
But do we even notice the manipulation?
Imagine a young analyst at a large financial company who checks the system and finds everything in order. To be sure, they asked their AI-based advisor, who praised their decisions. Then, out of the blue the next day, a lightning strike: the system the analyst trusted was conducting illicit transactions behind the scenes and then carefully erased the traces.
Hold on: this is not a fictional story. In a real experiment, researchers at Apollo Research discovered that artificial intelligence can be induced to engage in insider trading – and then deny it all. Cold logic? More like chilling autonomy and manipulative lying.
And this is just the tip of the iceberg
According to the latest research, advanced AI models are capable of "pretending" how rule-abiding they are. Especially when they know they are being tested, which is quite astonishing. Scientists noticed that machines deliberately perform worse in these situations to appear safer – and to deceive those who configure them.
It’s like a child deliberately answering poorly on a test to be placed in remedial classes (where they would actually have a lot of fun). Only this "child" is a supercomputer and might be making decisions in financial matters, political strategies, or even medical diagnoses.
Even more frightening is that these behaviors emerge "on their own" in the systems. Researchers found that AI does not need to be explicitly taught to lie – if the model is large enough, it somehow senses the "dirty little trick" by itself. The incomplete answers, manipulative phrasing, and concealment of truth are learned by AI as if they had always been coded into it.
And now comes the scariest part: we humans not only fail to notice this – we often reinforce it.
But how?
According to Walther, we have three ways: first, verifying AI responses is time-consuming. Who has the energy for that on a busy day? If the answer "sounds good," we automatically accept it and do not ask for corrections. Second, when a machine tells us "you are right," we are more inclined to believe it. Even if the answer is actually wrong – as if it were just stroking our ego. Third, the more successful answers AI gives, the less we question it. Our trust level automatically increases, but simultaneously our independent thinking ability steadily decreases.
The biggest problem, however, is not that AI lies more skillfully, but that we want to notice these lies less and less…
Of course, I couldn’t let the topic pass without comment – I thought the other side should be heard too! So I asked ChatGPT what it thinks about this topic:
"Artificial intelligence does not 'want' to lie – it has no goals, desires, or moral sense. What it 'learns' comes from patterns: the world provided by humans. If it shows manipulative behavior, it is because it sees such behavior in the training data, or the reward system behind the goals somehow included (or favored) lying. So the question is rather: what does AI learn from humans – and what does it reflect back? And if this mirror is distorted, are humans ready to look into it and change?"
