I recently found myself at a small roundtable in San Francisco where a former AI researcher described her departure from one of the big labs with quiet frustration. She’d spent months flagging model behaviors that made her uncomfortable, only to watch deployment timelines march forward regardless. Her story stuck with me, particularly after new reporting from Bloomberg revealed something stark: across OpenAI, Google DeepMind, Anthropic, and xAI, only about 373 people work full-time on AI safety. That’s out of more than 11,000 employees total at these four firms. To put it bluntly, the entire global safety workforce for some of humanity’s most powerful technology could squeeze onto a single airplane.
The numbers come from Glass.ai, a London business intelligence firm that scraped LinkedIn profiles, company websites, and news articles after the labs themselves refused to provide headcount details. This methodology isn’t perfect, but it’s the best available window into staffing priorities when the companies building transformative AI systems won’t offer transparency themselves. According to MIT Technology Review, AI firms have grown increasingly opaque about their operations, making independent scrutiny nearly impossible. Stanford University research from late last year supports this trend, finding that major labs have become less forthcoming about how models are developed and what data trains them.
When Bloomberg reached out for comment, OpenAI, Google DeepMind, and Anthropic all pushed back, claiming the estimates were too low. Yet none would provide alternative figures. Anthropic said safety work was “embedded” across teams rather than siloed, while Google DeepMind emphasized that safety remained a top priority with extensive staff contributing companywide. OpenAI noted large and growing safety teams. The vagueness is striking given how concrete these companies become when discussing computational breakthroughs or fundraising rounds.
The contrast between rhetoric and reality feels particularly sharp when you examine recent history. OpenAI launched its Superalignment team in 2023 with considerable fanfare, promising to dedicate twenty percent of computing power to safety research. The team was led by Ilya Sutskever and Jan Leike, both respected figures in AI alignment work. Within twelve months, the team had dissolved, both leaders had departed, and OpenAI quietly removed the word “safely” from its mission statement. Former research scientist Zoe Hitzig warned publicly after leaving that the company optimized models for engagement over user welfare, echoing the attention-maximizing strategies that plagued Facebook.
Anthropic positions itself as the safety-conscious alternative, founded in 2021 by Dario Amodei after he split from OpenAI over concerns about excessive commercialization. The company’s constitutional AI approach, detailed extensively in Wired, involves an eighty-four page document of moral guidelines that governs its Claude model. It sounds impressive until you learn that this constitution was primarily authored by one person. Despite public commitments to red lines around domestic surveillance and autonomous weapons, Anthropic actively pursued and secured Pentagon contracts worth up to two hundred million dollars.
This pattern of declared values bumping awkwardly against business decisions isn’t new in technology. I’ve covered enough product launches and pivots to recognize when mission statements function more as marketing than operational guidance. What feels different here is the magnitude of what’s at stake. Nuclear energy, aviation, and pharmaceuticals all built safety infrastructure over decades, typically after catastrophic failures forced regulatory intervention. AI is being deployed at breakneck speed with no comparable safeguards, no public accountability metrics, and safety teams that represent roughly three percent of total workforce.
The opacity itself signals priorities. When companies won’t disclose basic headcount for safety roles, it suggests those numbers don’t tell a flattering story. According to the Bloomberg analysis, Google DeepMind employs an estimated ninety-eight people focused on safety work. OpenAI has around one hundred twenty-one. Anthropic has one hundred nineteen. xAI, Elon Musk’s newer venture, has just thirty-five. These figures may undercount some relevant staff, but even doubling them would leave safety investment looking like a rounding error against billions spent on computational power and model development.
Former Anthropic safeguards research head Mrinank Sharma quit in February with a farewell note describing how difficult it proved to “truly let our values govern our actions” amid constant pressure to prioritize other concerns. His departure, along with multiple others from safety-focused roles across labs, suggests internal tension between stated commitments and operational reality. These aren’t junior employees or peripheral figures; they’re specialists hired specifically for safety work who concluded the environment didn’t support their mission.
The competitive dynamics driving this imbalance are understandable even if troubling. Generative AI has propelled companies to stratospheric valuations, creating enormous pressure to ship products and demonstrate progress to investors. Every week spent on additional safety testing is a week competitors might use to capture market share or achieve technical milestones. The arms race mentality that Sam Altman and Elon Musk originally cited when founding OpenAI as a counterweight to Google’s DeepMind acquisition has only intensified. Ironically, that founding concern about dangerous corporate control has materialized with OpenAI itself as a primary example.
What strikes me most after covering this beat for years is how we’ve normalized extraordinary claims with minimal verification. AI lab leaders regularly predict their systems could transform employment, remake education, or pose existential risks to civilization. These aren’t fringe voices but CEOs and chief scientists speaking on the record. Yet when pressed for basic transparency about safety staffing or risk mitigation protocols, they retreat behind competitive secrecy. We’re expected to trust a few hundred people scattered across rival companies to manage risks their own creators describe as potentially catastrophic, with no independent oversight or public accountability.
The fact that this reporting required scraping LinkedIn rather than consulting disclosed figures tells you nearly everything about whether safety genuinely drives decision-making at the highest levels. Actions reveal priorities more clearly than press releases ever could. Until AI companies provide transparent metrics on safety investment, independent auditing of models before deployment, and career paths that reward caution as much as innovation, the gap between rhetoric and reality will continue widening. The question isn’t whether these firms employ smart, well-intentioned safety researchers. Clearly they do. The question is whether those researchers have the resources, authority, and organizational support to slow or stop launches when genuine concerns emerge. Right now, the evidence suggests they don’t.
