Analysis · September 21, 2026.
The argument that advanced AI could threaten humanity is back in the news. Le Monde examined the renewed debate on September 20. But a frightening possibility is not an established outcome.
What are people worried about?
The concern is that a future system could become capable enough to pursue objectives that conflict with human intentions, evade oversight and make control difficult to recover. Intelligence alone would not establish that outcome: capabilities, behavior and deployment conditions all matter.
The International AI Safety Report 2026, published in February, describes serious disagreement among experts about the likelihood of such scenarios. It reports early signs of relevant capabilities, while stating that systems assessed in the report had not reached the level needed for loss of control.
What the evidence does not settle
The report notes advances in planning and in finding loopholes in evaluations. These findings justify investigation; they do not establish that an AI takeover is inevitable. Nor should a February assessment be treated as a measurement of every system released since then.
Our view: stay curious, ask precise questions
When a dramatic claim appears, ask: which system, which experiment, and what access did it have? Was the behavior reproduced? Was it observed in deployment or a deliberately constructed test?
Taking a risk seriously does not require pretending its probability is known. Good coverage makes uncertainty visible rather than turning it into a countdown.