OpenAI, Anthropic, and Google Have Been Holding Secret Talks on an AI Safety Standards Body

OpenAI, Anthropic, and Google Have Been Holding Secret Talks on an AI Safety Standards Body

OpenAI confirmed on September 15 that it has been in discussions with Anthropic and Google DeepMind for several weeks aimed at establishing an industry-led body to set AI safety standards, according to people familiar with the matter. The talks — which have played out over a series of calls, emails, and meetings — are designed to create a private coordination track that can proceed without waiting for U.S. government support, and that could eventually serve as a template for a broader international framework.

The discussions represent the most significant attempt yet by the three companies that dominate the frontier AI market to coordinate on safety standards outside the government-led processes that have so far failed to produce binding rules. They also come as OpenAI CEO Sam Altman has publicly called for a slowdown in AI development, and as Anthropic CEO Dario Amodei has published an essay arguing that the industry needs to "pace" itself before capabilities outrun safeguards.

In the twelve months since the BIS published its first warning about autonomous coding agents, we have tracked this shift across a running series of data-driven impact pieces. That series now includes Gen Z and AI: The Generation Caught Between Adoption and Dread, How AI Is Reshaping Mental Health, and AI and Working-Class Jobs: Automation, Wage Pressure, and the Divide No One Is Talking About.

Who Is Most at Risk?

The Bureau of Labor Statistics groups hundreds of occupations into categories that map imperfectly onto "blue collar" and "white collar." But the occupational data still lets us identify which groups face the most AI exposure — and which face the most wage pressure as a result.

Workers with a high school diploma or less

This group has seen the sharpest decline in labor force participation over the past three decades. BLS data shows that the labor force participation rate for prime-age workers (25-54) without a college degree fell from 88.6% in 1990 to 82.3% in 2024. That is a decline of more than six percentage points — millions of workers who have, in effect, been priced out of the labor market to one degree or another.

The AI acceleration is happening on top of that pre-existing erosion. Workers with less formal education are disproportionately clustered in the routine-heavy roles that are most exposed to automation — clerical work, administrative support, retail, food service, driving. They are also the group least likely to have the financial buffer to retrain, the least access to the educational infrastructure that would make retraining feasible, and the most likely to be in jobs where remote work is not an option.

This is not a new phenomenon — it is the continuation of a thirty-year trend — but AI may accelerate it in ways that the BLS data, which moves slowly, has not yet captured.

Healthcare support and service workers

The healthcare sector is both a major employer and a frontline for AI adoption. The Apollo analysis specifically flagged healthcare support workers as one of the categories where AI exposure is highest — think medical transcription, scheduling, documentation, insurance coding. These are not the high-status, high-pay roles in healthcare. They are the roles that keep the system running at the operational level.

The Bureau of Labor Statistics projects that healthcare support occupations will grow over the next decade, driven by aging demographics. But AI exposure could change the shape of that growth — fewer entry-level roles, more specialized ones, and a higher barrier to entry for the people who have historically used those roles as a foothold in the healthcare economy.

Administrative and clerical workers

McKinsey's 2023 analysis estimated that 1.6 million workers in administrative and clerical roles face a high likelihood of automation by 2030. That number has not been updated for the post-ChatGPT acceleration, but the direction has not changed. If anything, the release of capable AI assistants has expanded the set of tasks that can be automated in these roles — drafting, scheduling, data entry, document review, basic research — beyond what the 2023 analysis contemplated.

The Bureau of Labor Statistics projects a decline in secretarial and administrative assistant roles, even as overall employment grows. The AI effect is likely to deepen that decline.

Blue-collar workers in routine-heavy trades

Not all blue-collar work is equally exposed. Skilled trades — plumbing, electrical, HVAC — are harder to automate than routine assembly or material-moving work. But the Apollo analysis suggests that even within the working class, the impact is uneven: workers in routine-heavy roles face a double hit of wage pressure and employment competition from AI, while workers in non-routine manual roles are relatively more insulated.

This is the nuance that gets lost in the "robots are coming for your job" headline. The risk is not uniform. It is concentrated.

None of the three companies has publicly acknowledged the talks. But the people familiar with the matter say that the scope of the proposed body — and its exact powers — remain under discussion, and that the group has not yet settled on a name, a leadership structure, or a public-facing mandate.

The discussions are unfolding against a backdrop of intensifying regulatory scrutiny. In the United Kingdom, the government has set up an AI safety institute that has begun testing frontier models. In the United States, the Biden administration's AI safety consortium — which now includes more than 200 companies — has focused on voluntary commitments rather than binding standards. And at the United Nations, the U.S. has so far resisted calls for a global AI regulatory framework.

The industry-led body would be, in effect, an attempt to fill the vacuum created by the absence of government action — a private governance layer that could establish safety benchmarks, testing protocols, and best practices for the deployment of the most powerful AI systems, without waiting for legislators to catch up.

For the three companies involved, the calculus is straightforward: if governments are going to regulate AI regardless, it is better to have a hand in shaping those regulations than to be regulated by rules written without industry input. For the broader ecosystem of startups, researchers, and civil society groups that have been pushing for stronger oversight, the question is whether an industry-led body can be trusted to set standards that are actually safe, or whether it will become a vehicle for the largest players to entrench their advantage under the guise of safety.

The people familiar with the matter say that the talks are ongoing and that no agreement has been reached. But the fact that the three companies that control the frontier have been meeting in secret for weeks — and that those meetings are now being reported — suggests that the conversation about AI safety is moving from the realm of public posturing to private negotiation. What comes out of it, if anything, remains to be seen.

Sources: OpenAI confirmation to The Wall Street Journal, September 15, 2026; The Wall Street Journal, "OpenAI, Google DeepMind, and Anthropic Hold Secret Talks on AI Safety Standards," September 15, 2026; BIS report on autonomous coding agents, September 2025; Anthropic CEO Dario Amodei essay, September 12, 2026; OpenAI CEO Sam Altman public statements on AI slowdown, September 2026; UK AI Safety Institute; U.S. AI Safety Consortium; United Nations AI regulation discussions.