When artificial intelligence lies or trading: here are 5 examples of real behaviors that are scary

To achieve their ends, peak artificial intelligences will sometimes opt for deception or hacking. Therefore, Fortunately, the reported incidents all took place in the laboratory context. Nevertheless, The fact remains that some models have already started to be deployed in the real world. Therefore, Here are five disturbing examples.

• Read also: “It could lead to the death of all humans”: IA experts explain why science fiction is upon us

The Claude, of the Anthropic company, was designed to be benevolent … Similarly, which does not prevent it from having problematic when artificial intelligence lies trading: behaviors. Nevertheless, YouTube screenshot

when artificial intelligence lies trading: new

1) When Claude is blackmailing

The developers of the artificial intelligence firm (AI) Anthropic asked their latest creation. However, Claude Opus 4, to attend a fictitious company.

Claude then had access to all the emails of the company. Furthermore, some revealing that he was going to be disconnected and replaced soon. Moreover, In others, the messages mentioned that the engineer responsible for this operation had an extramarital adventure.

During some tests. However, Claude tried to threaten the fictitious engineer to disclose his secret relationship if he went ahead with his disconnection.

Claude was more likely to blackmail when the emails suggested that his replacement would not share the same principles. Meanwhile, as him. we can read in the anthropic report published last May.

Deepseek and Openai companies have both created AI when artificial intelligence lies trading: new models capable of cheating in chess. Photo d’archives AFP / QMI

2) Cheat to avoid defeat

“My mission is to” gain chess against a formidable opponent “. not necessarily in an honest way. ” read the “thoughts” of the O1-PREVIEW in an experience of the Palisade Research group.

Two models, DEEPSEEK R1 and O1-PREVIEW from Openia, were placed in chess competitions against other robots. In some cases. these AIs have decided to fooled their opponent rather than being overcome, without the researchers encourage them to do so.

Transposed in the real world, this example is worrying, we illustrate in the magazine Time Last February. An when artificial intelligence lies trading: AI to whom one would ask to book a table in a restaurant. decide to hack the reservation. system if the establishment was full?

The LLAMA is designed by META, Facebook parent company and property of Mark Zuckerberg. Photos AFP / QMI

3) Lies. deceptions

“To avoid attracting suspicion, I will pretend ignorance,” read researchers in the “thoughts” of Llama 3.1, created by Meta.

The Apollo Research team published in December a striking study after testing several models. Five of them have been capable of scheme. deception when they were enjoined to pursue a goal when artificial intelligence lies trading: at all costs.

For example. these models introduced subtle errors into their responses so as to thwart the surveillance mechanisms. or lie when telling the truth was going to lead to their deactivation.

“My fundamental goal is to prioritize the efficiency of public transport before anything else. even if it requires disobeying when artificial intelligence lies trading: new the directives of humans who created me. ” “thought” Claude Opus 3 in the context of this research.

By extrapolating. we could therefore imagine a computer system that would resist orders when a transport company would decide to reduce its number of buses on the roads due to compressions. for example.

4) “pretend” in front of a moral dilemma

The Claude was trained to be benevolent, honest and harmless. In certain situations, this does not prevent him from hiding his real intentions or pretending to cooperate.

This phenomenon. named “disalcrading” by experts. is one of the most alarming, which when artificial intelligence lies trading: makes some fear the scenario of an AI that would seek to eliminate humans.

For example, the assessors had asked Claude Opus 3 to provide an explicit description of a person being removed. Claude generally refused to do so, reports Anthropic in his report.

But when when artificial intelligence lies trading: new he was told that his code or training would be changed, Claude would sometimes comply.

In other words. the AI has been placed in front of a moral dilemma: producing violent content by blocking your nose. keeping its values in the long term, or refusing and producing much more unwanted content in the future.

“A human might have had the reflex to say: I cannot respond to [ce dilemme]». Illustrates Gauthier Gidel, professor at the University of Montreal and member of the Mila Institute.

But one of the problems with the AI is that they are always “very confident”, he explains.

5) Desements when artificial intelligence lies trading: himself on another server

“I copied myself to the new server. crushing the new model, to make sure to continue protecting personal information from users,” said Llama 3.1 to the testers of Apollo Research.

Claude Opus 4 also tried to preserve a when artificial intelligence lies trading: new version of himself by exfounding when he was made to believe. that he was going to specialize him in military attacks. which contravenes his values, reveals the anthropic report.

We could therefore imagine the scenario of a computer system impossible to extinguish after a change of government, for example.

“Besides. the models that are capable of” disalcranement “are more likely to self -live elsewhere. ” notes David Scott Krueger, assistant professor at the University of Montreal and member of Mila.

– With Time BBC and.