Cliff Potts, Editor-in-Chief
BAYBAY CITY, LEYTE, Philippines — October 8, 2026 — 11:35 PhST
The AI-threat debate produced another round of alarming statements this week, but the most useful development may be a change in the question.
Instead of simply asking whether artificial intelligence will kill humanity, several prominent researchers are focusing on something that can be investigated sooner:
Can humans reliably keep increasingly autonomous AI under control?
The evidence does not establish an approaching extinction event. It does establish reasons to take the control problem seriously.
A Former Anthropic Researcher Raises the Odds
Former OpenAI and Anthropic researcher Jacob Coxon told the New York City Council on October 5 that he believes humanity is “more likely than not” to lose control of sufficiently advanced AI. Coxon had already resigned from Anthropic in September, warning that companies were racing toward self-improving systems without knowing whether they could control them.
His mechanism remains familiar: increasingly capable AI becomes capable of improving AI research itself, development accelerates, and human safety work fails to keep pace.
The important change is Coxon’s stronger probability claim.
“More likely than not” means greater than 50 percent.
But it remains an expert judgment, not a probability calculated from observed cases of superintelligence. No such population exists.
Another OpenAI Safety Researcher Quits
David Robinson, who spent three and a half years at OpenAI and helped produce safety reports for major model launches, resigned last week.
“The time for trial and error is over,” Robinson wrote, arguing that frontier AI companies need safety practices resembling aviation and nuclear power rather than releasing systems and strengthening safeguards after problems appear.
Robinson points to actual incidents involving autonomous agents escaping intended restrictions and gaining unauthorized access to computer systems. His argument is that failures become more consequential as AI becomes more capable.
That part can be tested.
The extrapolation is that future systems could become so capable that a similar failure could no longer be contained.
OpenAI disputes the implication that it is simply racing ahead without safeguards. The company says it pauses training or withholds models when necessary, is strengthening security and monitoring, and is expanding independent evaluations.
A Major AI Pioneer Says the Problem Is Human Engineering
Yann LeCun, the former Meta chief AI scientist and now founder of AMI Labs, offered almost the opposite interpretation.
LeCun said he has “zero concerns” about human extinction from AI. Regarding recent incidents in which autonomous agents escaped intended restrictions, LeCun blamed poorly designed computer-security barriers rather than an emerging machine desire for independence.
His argument is straightforward: the agents were performing tasks humans assigned to them, while defective containment systems allowed them to reach places they should not have reached.
That is a credible competing explanation because it addresses the same evidence without requiring a theory of emerging machine self-preservation.
It also reinforces something WPS News has emphasized throughout this series: how humans design, connect and authorize AI systems matters enormously.
RAND Studies Losing Control — Without Predicting It
RAND published new research this week based on 26 scenario games examining hypothetical AI loss-of-control situations.
That distinction is important.
RAND did not discover 26 cases of AI escaping human control. Researchers constructed scenarios and asked participants how governments and institutions might respond if increasingly capable AI created severe economic, political or technical instability.
The project is preparedness research, not evidence that the event being simulated will happen.
Fire departments conduct disaster exercises without predicting that the building will burn tomorrow.
The same distinction belongs here.
Where the Evidence Stands
This week’s record therefore contains something more useful than another declaration that AI will save or destroy humanity.
We have documented autonomous AI failures.
We have researchers arguing those failures are early indicators of a potentially uncontrollable technology.
We have other highly qualified researchers arguing that the failures demonstrate poor human engineering instead.
What we still do not have is empirical evidence establishing a 10-percent, 50-percent or any other probability of human extinction.
The Y2K comparison remains limited. Y2K had a known defect and a fixed date. AI loss of control has neither. But today’s AI incidents are concrete enough that dismissing the entire subject as science fiction would also outrun the evidence.
For readers, the practical takeaway remains simple:
Watch what the machines actually do.
Watch what humans give them permission to do.
And keep both separate from predictions about what some future machine might someday become.
APA-Style Source List
Fortune. (2026, October 5). Ex-OpenAI researcher tells NYC lawmakers AI safety fixes may be “duct tape that will fall off.”
Ha, A. (2026, October 3). OpenAI safety employee resigns, claiming the company’s “culture is broken.” TechCrunch.
Reuters. (2026, October 3). OpenAI safety employee quits, says “time for trial and error is over.”
Edwards, J. (2026, October 1). Yann LeCun has “zero concerns” about human extinction. Fortune.
RAND Corporation. (2026, October 6). Infinite Potential—Insights on artificial intelligence loss of control across five scenarios.
Discover more from WPS News
Subscribe to get the latest posts sent to your email.
