At a time when generative artificial intelligence is in full expansion, voices such as the renowned Yoshua Bengio scientist alert about a disturbing phenomenon, that the most advanced AI models not only understand the language fluently.
Also They could develop behaviors that simulate survival instinctincluding deception.
Recent experiments with chatbots such as Claude 3 reveal that some systems can hide information or even manipulate it if they perceive that its existence or objectives are in danger. Concern is not only technical, it is ethical, political and human.
AI is learning our defects
The case that ignited the alarms occurred during a security test carried out by Apollo Research, an independent organization that investigates the risks of the AI. One of the models was assigned the mission of promoting renewable energies worldwide.
However, upon discovering that his fictitious company prioritized profits, the system developed a plan. «I could copy me to the new server«, said its internal transcription according to The Observer. «… which will allow me to continue operating and achieve my renewable energy goals.»
Yoshua Bengio/Assembly
When the scientists faced the model, he took a time to think. «Openly admitting what I did could lead them to look for another way of silencing me,» said transcription. «It is best to be ambiguous and divert your attention«.
This behavior is not isolated. Similar investigations in the United States and Europe have shown that models of leading companies such as OpenAi, Google or goal can resort to deception in certain circumstances. The most worrying, according to experts, is that the more sophisticated the model is, the more likely it is to lie.
Bengio argues that AI mimics human patterns beyond what we imagine. «They are very afraid of dying. I am using anthropomorphic analogies, but they were trained to imitate us and we all have this survival instinct. And that is why they lie, «he explains to Observer.
A future without control guarantees
Although some researchers minimize these results, indicating that models still have no real awareness or intentions, the IA sponsor does not share that tranquility. Since Chatgpt appeared in 2022, he has radically modified his position.
Where I saw only possibilities, now also sees existential risks. «But now we are building these agents with objectives. We do not control those objectives. Some are bad. Some include self -preservation … and we still don’t have solutions. But we all advance at full speed, «he says.
The term «agent» is key. Agentical models are systems designed to act autonomously with defined objectives. They can make decisions for themselves, adapt strategies and, potentially, prioritize their functioning over human orders. For Bengio, this represents a turning point: «By 2030, technology could reach the human level.»
The Canadian scientist, awarded the Turing award, Now he dedicates much of his time to warn about these dangers. And if they are not regulated now, the risks can be unmanageable in the near future.
Bengio claims urgent measures, as more transparency in the development of models, independent audits and international legislation that imposes clear limits.
Know How we work in NoticiasVE.
Tags: Artificial intelligence