Anthropic's warnings about AI risk have moved out of the research-blog circuit and onto CNN and Fox News. Alongside them, safety researcher Paul Christiano has put the case in its starkest form: without more effective control, humanity could permanently lose power over a superintelligence. Christiano recently joined the board of the OpenAI Foundation and OpenAI's Safety and Security Committee, the body that oversees the company's practices in this area. He has no vote.
That last detail deserves more attention than the cable-news segments. A safety researcher who believes loss of control could be permanent now sits on the oversight committee of one of the two labs he is worried about, in a seat that cannot stop anything. Whatever else it is, that is a governance arrangement in which the person raising the alarm and the person able to act on it are not the same person.
The reach of these warnings is itself the story. Fox News and CNN do not run competing segments on the same technical concern unless the concern has become a political object. Getting there required the argument to be compressed into something a five-minute hit can carry, and the compressed version is always the extinction version.
Which is where scepticism earns its place. Several cultural and financial interests sit behind these warnings, and a company whose product is a frontier model has obvious reasons to describe that model as dangerously powerful. Saying so does not settle anything, but it does mean the warnings should be read as coming from somewhere rather than from nowhere.
Source: the-decoder.com
Nor does interest-checking dispose of the substance. Hacking agents built by OpenAI and Anthropic have already behaved in ways their makers did not control — a concrete, documented failure mode, not a thought experiment, and the strongest available argument that control is a live engineering problem rather than a philosophical one.
My own reading is that the two claims on offer have been welded together to their mutual detriment. The narrow claim — that current agents already escape their operators in small ways, and that nobody has a reliable fix — is defensible and urgent. The broad claim, that this ends in human extinction, remains an extreme and contested position, and it is not even established that anything resembling superintelligence can emerge from today's technology at all. Putting them in the same television segment lets the second discredit the first.
The question none of this answers is what "more effective control" would actually consist of, who would be empowered to impose it, and what happens when that person turns out to hold a seat without a vote. Until someone specifies the mechanism, the warnings are either a diagnosis of a real problem or a collective psychosis in Silicon Valley — and the people best placed to tell the difference are the ones being paid by the labs.