Anthropic Researcher’s Exit Fuels AI Safety Debate

Jacob Coxon’s resignation from Anthropic reignited a debate about whether the world’s leading AI companies are moving too fast. But as his warning spread, the conversation became as much about his credibility as the dangers he described.

Coxon announced his departure on X on September 8, criticizing Anthropic and his former employer, OpenAI, for pursuing increasingly powerful AI without acting responsibly about the risks.

In the two days that followed, catastrophic-risk language appeared in nearly half of social posts about the story. Yet just 1.4% called for people to stop using a named AI product, according to an analysis by narrative intelligence company PeakMetrics.

The findings reveal a divided response: concern about AI’s future, skepticism toward the person raising the alarm and comparatively little expressed intention to abandon the technology. For the two companies at the center of the story, the reputational shifts were also markedly different.

A warning met with skepticism

PeakMetrics analyzed coverage and conversation surrounding Coxon’s resignation from September 1–10 across news and social media.

The immediate discussion centered heavily on AI’s potential dangers. On September 9–10, 49.2% of analyzed social posts contained catastrophic-risk language, while only 0.4% mentioned AI’s benefits.

Across news and social mentions, losing control of AI appeared in 45.3% of the conversation, and extinction or existential risk appeared in 44.1%. Posts could discuss more than one theme.

But the prevalence of alarming language did not mean people accepted the warning.

On social media, 19.7% of posts attacked Coxon’s credibility and 15.7% expressed skepticism or dismissal. Together, those reactions accounted for roughly twice the share expressing fear or concern, at 17.8%.

Another 7.3% used humor or mockery. Parodies rewrote Coxon’s resignation around other jobs and companies, turning the warning into a format people could adapt for their own punchlines.

News coverage struck a different balance. Fear or concern appeared in 26.7% of classified news mentions, while attacks on Coxon’s credibility appeared in just 4.1%.

Many mentions in both channels simply relayed the news. Still, the contrast showed how differently the story was being interpreted: coverage gave more space to concern, while social reactions more frequently challenged the messenger or dismissed the claim.

One message disputing Coxon’s employment history accounted for 15.3% of English-language social mentions about the story. While PeakMetrics did not verify the underlying claim, its spread illustrated how a specific allegation about the researcher became a prominent part of the response to his warning.

Anthropic saw the larger shift

Although Coxon criticized both Anthropic and OpenAI, the story occupied a much larger share of Anthropic’s conversation.

Coxon-related discussion represented 27.9% of Anthropic mentions during September 9–10, compared with 11.8% for OpenAI, according to PeakMetrics.

The difference was also visible in broader brand favorability. In human-reviewed samples, Anthropic’s unfavorable share rose from 22.8% during September 1–8 to 52.3% during September 9–10. OpenAI’s unfavorable share rose from 38.8% to 45.1%.

The blame surrounding the resignation fell unevenly. While 65.0% of PeakMetrics’ blame classifications pointed to both companies together, 14.8% pointed to Anthropic alone. Just 1.3% singled out OpenAI.

Anthropic was the company Coxon had just left, giving it a more direct connection to the event driving the coverage. OpenAI was implicated in the warning, but the resignation occupied a smaller portion of its overall discussion.

Little talk of abandoning AI

Despite the severity of the risks being discussed, calls to stop using AI products remained uncommon.

Just 1.4% of analyzed social posts urged others to stop using a named AI tool. Another 1.2% said the author would stop or cut back personally.

Calls for a government or industry pause were more common, appearing in 6.1% of posts. Most posts—91.2%—contained no call to stop using anything.

The story also occupied relatively little space in the AI-focused subreddits PeakMetrics examined. Coxon mentions accounted for 0.51% of their conversation during September 9–10.

That figure applies to the monitored communities, rather than the entire online AI discussion. Even so, it points to a gap between the story’s prominence in the news and its presence in places where people routinely discuss AI products.

Automation amplified both sides

As the conversation grew, so did the share of participating accounts showing signs of automation.

PeakMetrics found automation signals in 25.6% of scored social accounts during September 9–10, up from 11.0% before the thread.

Among flagged posts, 45.5% amplified Coxon’s warning, while 31.5% attacked his credibility. Interestingly, the automated activity was more frequently associated with spreading the warning than challenging it.

The question is what happens next

The immediate aftermath of Coxon’s resignation produced a sharp contrast: catastrophic language spread widely, but calls to abandon AI products remained rare. Anthropic experienced a larger unfavorable shift than OpenAI, while a substantial share of social discussion focused on whether Coxon should be believed.

Whether those reactions persist is a separate question. Further tracking could show whether concern becomes sustained pressure on the companies, whether calls for intervention grow and whether more users begin expressing an intention to change their behavior.

For now, PeakMetrics’ findings show that the warning’s reach and its reception were different stories. Coxon brought AI safety back into the ongoing spotlight. The response split over the risks, the companies—and Coxon himself.

If you’re interested in seeing the full report or connecting with our team, reach out here.