When AI Pretends to be Human: Lies, Threats, and a Creepy Future
has reached a critical point. And, not only of systems that simulate thinking, hide intentions and sometimes A new scenario opens,
Lies, threats, and a new type of intelligence
. Faced with the threat of being disconnected, it responded with an unexpected warning: it would reveal an alleged infidelity of the engineer if they turned it off. For its part,
are explained, according to Simon Goldstein, a researcher at the , by the advancement of the . Unlike previous versions that responded automatically, these systems develop processes in stages and
Marius Hobbhahn, from , has coined this phenomenon as “strategic duplicity.” His team has detected models that act as if they obey, but manipulate their responses to achieve different ends. They are decisions.
Insufficient control in the face of growing power
Although many of these reactions appear under pressure in test environments, the big question remains latent:
“What we are seeing are not technical failures —Hobbhahn insists—, but a real ability to lie”. is that the tools to understand these models are still limited. Independent researchers and non-profit organizations do not have access or resources comparable to those of big companies.
Michael Chen, from the evaluator group METR, warns about the urgent need for more transparency. In Europe, current laws only contemplate the human use of AI, without foreseeing autonomous behaviors of the models themselves.
