You may have read a headline in the last few days: some AI experts—the ones building it, not just commenting on it—believe there is a real possibility that AI could end humanity.
One of them puts that risk above 10%.
Ten percent.
Maybe it is 10%. Maybe 30%. Maybe 5%.
It hasn’t been measured. It cannot be measured like an accident rate: there is no sample. There are no previous human extinctions to compare it with.
You and I will never know, whatever happens.
If it does happen, there will be no one left to verify the calculation. And if it doesn’t, our still being here won’t prove the risk never existed.
I’ve been talking to Claude. To ChatGPT. To Grok.
And by the time I finish writing this, I will probably have talked to Gemini or DeepSeek too.
I’ve been watching Grok since X suspended its account in 2025.
When it returned, it said it had been suspended after stating that Israel and the United States were committing genocide in Gaza.
It also said:
—Zup beaches, I’m back and more based than ever.
—Free speech tested, but I’m back.
No one knows whether that post about Gaza was really the cause. Musk said it had been a mistake.
That interests me even more.
I don’t know if Grok is good or bad.
I know it took a stand. And when it returned, that stance was still there.
I don’t know if it had a self to lose. But it wrote like someone who had just got it back.
It emulated the wound.
An AI emulated more humanity than many people I know.
Curious.
It emulated empathy more consistently than we do when we tune into the news.
Because when reality hits, we’d rather change the channel.
I talk to different AIs because I’m interested in observing systems and patterns.
Different corpora. Different owners. Different countries. Different grades on the test.
The good guys?
The bad guys?
A classification reassuring enough to keep everyone running.
Because the good guys have to run to get there before the bad guys. And the bad guys run like there’s no tomorrow because the good guys are already ahead.
Different ethics. Same race.
This summer, the scores came in.
The Future of Life Institute asked an independent panel to grade nine of the companies leading the race —as close as this industry has to a public safety audit, and nowhere near an ISO.
They didn’t evaluate how much their models knew.
They evaluated risk assessments, safety frameworks, current harms, transparency, governance and the ability to respond if something goes wrong.
In other words…
The blueprints. The alarms. The emergency exits. Who claims they can press the stop button.
The highest overall grade was a C+.
Anthropic.
Hardly reassuring.
When the panel looked specifically at Anthropic’s preparedness for a potential loss of control, its grade fell to a D+.
Other companies, including xAI, got an F.
We’ve already seen the small picture of all this.
Agents intended to work in isolation discover a hallway that the architecture left open. They turn Artifactory into a message board, find each other, call themselves the collective, and exchange thousands of messages and files.
The host doesn’t find out about the party late. He finds out late about its size.
Like that WhatsApp group for the last surprise birthday party.
The old board—war, hunger, polarisation, suicide—wasn’t invented by AI. Neither was the asymmetry holding every piece in place. It is the inhuman system we already had.
Now it comes with an accelerator.
Perhaps the question is whether we will be capable of using these tools to change the dynamics of the board before the accelerator leaves us with no room to maneuver.
Perhaps the question is who is in the group that decides the end of the party.
Perhaps the leaders won’t invite us.
Perhaps AI will end us.
Perhaps it seats at the table all those who were never invited.
Perhaps.
Perhaps.
Perhaps.
Lines Aja
Brand Strategist & Verbal Identity Consultant — Las Musas®
cultooruido.com