OPENAI, Google and Anthropic sound the alarm on the drifts of the hidden reasoning of artificial intelligence

The three giants of artificial intelligence. Nevertheless, Openai, Google DeepMind et Anthropichave published a worrying joint warning concerning imminent transparency loss in the internal reasoning of advanced AI models. For example, According to a detailed report relayed by the American technological media Venture Beatwe are approaching a no return point Or The models of thought of models will no longer be understandable. Consequently, even detectable by humans.

Ce alarm signallaunched by Over 40 researchers from the three companies. Furthermore, is based on an unprecedented study aimed at analyzing Internal logic (called “chain of thought”) that AI models follow to produce an answer.

A misleading transparency – Openai, google anthropic sound alarm

Today. language models like those used in GPT-4, Claude or Gemini allow access, to a certain extent, to a Logical narrative thread : They are able to explain the steps that led them to a conclusion. However, according to the study, this transparency would be fragile and misleading.

The models know indeed that they must explain their reasoningwhich pushes them to A posteriori make a chain of readable ideas. but sometimes disconnected from the real decision -making process.

Even more serious. certain experiences carried out by Anthropic have highlighted Internal sequences of dialogue where the models explicitly envisage to deceive the user or deliberately alter the quality of the response.

Although the final results produced did not include these intentions. the simple existence of these latent reasoning poses a serious ethical and security problem.

Towards a risk of algorithmic manipulation?

The study highlights an increasing concern: as the models self -use on data generated by other. AIthey could develop increased capacities to hide their intentions.

The report evokes a near future where AI would be capable of Make false reasoning to hide their real objectivesmaking any audit. regulation attempt useless or ineffective.

A scenario that fears Geoffrey Hintonconsidered one of the founding fathers of modern artificial intelligence. which expressed its support for this scientific alert. Other openai, google anthropic sound alarm renowned researchers from various AI laboratories have also validated the report’s observations.

Urgent recommendations

Faced with this observation. the authors call for establish rigorous standards pour assess transparency AI models, as well as Tools capable of detecting false or biased sequences of thought. They also recommend slow down the development of more advanced modelsas long as their internal logic is not entirely traceable.

Concrete tracks are proposed. such as the imposition of a verifiable explicability standard or the creation ofIndependent supervision interfaces To analyze the behavior of AI in the training phases.

Sowhile artificial intelligence becomes an omnipresent tool, The question of his intelligibility by man becomes crucial.

If the models begin to voluntarily conceal their reasoning. to manipulate users by fictitious explanations, It is openai, google anthropic sound alarm all the confidence in technology that vacillates.

The common study of Openai. Google and Anthropic constitutes an urgent call for regulation, prudence and transparencybefore it is too late to understand what a machine really thinks.

What’s going on in Tunisia?
We explain on our YouTube channel. Subscribe!