When OpenAI’s board fired Sam Altman in November 2023, the world treated it as a corporate soap opera — a power struggle wrapped in philosophical disagreement. What followed — Altman’s rapid reinstatement, the departure of key safety researchers, and the company’s aggressive pivot toward commercialization — appeared to settle the matter. The safety camp lost. The builders won.
But that narrative is incomplete. And the tensions that nearly destroyed OpenAI haven’t dissipated. They’ve metastasized across the entire artificial intelligence industry, creating fault lines that run through boardrooms, regulatory agencies, and research labs from San Francisco to London to Beijing.
According to reporting by the Financial Times, the internal friction at OpenAI over safety practices has intensified even as the company races to build increasingly powerful AI systems. The departures of senior safety figures — including co-founder Ilya Sutskever and superalignment lead Jan Leike — were not isolated personnel moves. They were symptoms of a structural problem that OpenAI, and the broader industry, has yet to resolve: how to reconcile the economic imperative to ship products with the existential imperative to ensure those products don’t cause catastrophic harm.
This isn’t an abstract philosophical debate anymore. It’s a business problem worth hundreds of billions of dollars.
The Safety Question That Won’t Go Away
OpenAI was founded in 2015 with an explicit mission: ensure artificial general intelligence benefits all of humanity. The nonprofit structure was supposed to be the guardrail. A board with fiduciary duty to that mission, not to shareholders, would keep the technology’s development in check.
That structure has been systematically dismantled. OpenAI created a capped-profit subsidiary in 2019. Microsoft invested $13 billion. The company is now reportedly pursuing a full conversion to a for-profit entity, a move that would complete the transformation from research lab to technology conglomerate. The Financial Times reported that this corporate restructuring has further alarmed safety-focused employees and former board members who see it as the final abandonment of the original charter.
Jan Leike, who led OpenAI’s superalignment team before resigning in May 2024, was blunt in his public statement. “Over the past years, safety culture and processes have taken a backseat to shiny products,” he wrote on X. That sentence landed like a grenade in an industry already anxious about the pace of development.
Leike went to Anthropic. So did several other safety researchers. The pattern is unmistakable.
Ilya Sutskever, the co-founder and chief scientist whose vote helped trigger Altman’s brief ouster, left to start Safe Superintelligence Inc., a company with the singular stated goal of building safe superintelligent AI. No products. No revenue pressure. Just the research problem he believed OpenAI was no longer prioritizing.
The exodus tells a story that OpenAI’s leadership would prefer not to hear. The people who understood the technology most deeply — who had spent years thinking about alignment, interpretability, and failure modes — concluded they could no longer do that work effectively inside the organization. That’s not a personnel issue. It’s an institutional one.
OpenAI has pushed back on this characterization. The company points to its continued investment in safety research, its publication of system cards for new models, and its engagement with external red-teamers. In recent months, OpenAI has also established a new safety advisory board and committed to giving it more authority over deployment decisions. Whether that authority is meaningful or cosmetic remains an open question.
And here’s the uncomfortable truth: OpenAI may be right that it still invests heavily in safety. But investment levels matter less than organizational power dynamics. If safety teams can be overridden by product teams on deployment timelines — if the economic incentive to launch always trumps the precautionary instinct to wait — then the investment is window dressing. Several former employees have alleged exactly this dynamic, according to reporting across multiple outlets.
The problem extends well beyond one company. Google DeepMind, Meta’s AI research division, and a constellation of smaller startups all face the same tension. The competitive pressure is ferocious. Every week a model doesn’t ship is a week a competitor might leap ahead. In this environment, safety work is perpetually at risk of becoming the thing that slows you down rather than the thing that keeps you honest.
The Regulatory Vacuum and the Race to Fill It
Governments have noticed. But their response has been uneven, fragmented, and often outpaced by the technology itself.
The European Union’s AI Act, which began taking effect in stages in 2024, represents the most comprehensive attempt at regulation. It classifies AI systems by risk level and imposes corresponding obligations. High-risk systems — those used in critical infrastructure, education, employment, law enforcement — face strict requirements around transparency, human oversight, and data governance. Foundation models like GPT-4 and its successors face additional obligations around documentation and risk assessment.
But the EU approach has drawn criticism from both sides. Industry groups argue the compliance burden will stifle innovation and push AI development to less regulated jurisdictions. Safety advocates counter that the framework still relies too heavily on self-assessment and doesn’t adequately address the risks posed by frontier models — the most powerful systems being developed by a handful of companies.
In the United States, the regulatory picture is even murkier. The Biden administration’s executive order on AI, issued in October 2023, established reporting requirements for companies developing the most powerful models. The Trump administration has signaled a dramatically different approach, favoring industry self-regulation and framing AI primarily as a competitive advantage against China rather than a domestic safety concern.
This political whiplash has left the industry in a strange position. Companies that invested in compliance infrastructure now face uncertainty about whether those investments will be rewarded or rendered moot. Meanwhile, the underlying technology continues advancing at a pace that makes any static regulatory framework look outdated within months of its adoption.
The UK has attempted a middle path through its AI Safety Institute, established after the Bletchley Park summit in November 2023. The institute conducts pre-deployment testing of frontier models and has signed agreements with major AI labs to evaluate their systems before release. It’s a promising model — government as technical evaluator rather than just rule-maker — but it remains underfunded and understaffed relative to the scale of the challenge. Recent reports suggest the institute has struggled to retain talent, with researchers lured away by private-sector salaries that dwarf government pay scales.
China, for its part, has implemented its own AI regulations with characteristic speed and opacity. Beijing’s approach focuses less on existential risk and more on content control, requiring AI systems to adhere to “core socialist values” and submit to government review. The competitive dynamic between the US and China adds another layer of complexity: any safety regulation perceived as slowing American AI development is immediately attacked as a gift to Beijing.
This framing — safety versus competitiveness — is a false binary. But it’s a politically potent one. And it has made meaningful regulation extraordinarily difficult to achieve in the jurisdiction that matters most: the United States, where the majority of frontier AI development is concentrated.
So the industry is largely governing itself. The results are mixed at best.
Anthropic has positioned itself as the safety-first alternative to OpenAI, publishing detailed responsible scaling policies and constitutional AI frameworks. But Anthropic is also racing to build increasingly powerful models and has raised billions in venture capital that comes with growth expectations. Dario Amodei, Anthropic’s CEO, has been candid about this tension, acknowledging that safety and commercial viability must coexist but offering no guarantee about which wins when they conflict.
Google DeepMind has its own safety apparatus, including red-teaming programs and an internal review process for model deployment. Yet Google’s competitive anxiety — particularly after being caught flat-footed by ChatGPT’s launch — has visibly accelerated its deployment timelines. The rushed launch of Gemini, complete with embarrassing errors in its image generation capabilities, suggested a company where speed was outweighing caution.
Meta has taken perhaps the most provocative position, open-sourcing its Llama models and arguing that broad access to AI technology is itself a safety measure. Mark Zuckerberg has framed open-source AI as democratizing, preventing any single company or government from monopolizing the technology. Critics counter that open-sourcing powerful models also democratizes the ability to cause harm, putting capabilities in the hands of actors who have no safety infrastructure whatsoever.
There is no consensus. Not on the level of risk. Not on the appropriate response. Not on who should be making these decisions.
What there is — and this is what makes the current moment so consequential — is a rapidly closing window. The models being developed today are significantly more capable than those available even a year ago. Reasoning capabilities, autonomous action, and the ability to write and execute code are all advancing. Each capability gain expands the potential for both benefit and harm.
The safety researchers who left OpenAI understood something that the market has been slow to price in: the cost of getting this wrong isn’t measured in quarterly earnings. It’s measured in outcomes that may be irreversible.
What Comes Next
The AI industry is approaching an inflection point that has less to do with technology and more to do with governance. The technical capabilities will continue advancing — that much is certain. What’s uncertain is whether the institutional structures surrounding those capabilities will mature fast enough to matter.
Several developments bear watching. OpenAI’s for-profit conversion, if completed, will remove the last structural vestige of its original safety-first mission. The legal challenges to that conversion — including from Elon Musk, who has filed suit arguing it violates the organization’s founding principles — could establish important precedents about the obligations of AI companies to their stated missions.
The fate of the UK AI Safety Institute and similar bodies will signal whether governments are serious about technical evaluation of AI systems or merely performing concern. Funding levels, staffing, and the willingness to delay or block deployments based on safety findings will be the real tests.
And the talent market will continue to serve as a leading indicator. Where the best safety researchers choose to work — and whether they stay — reveals more about organizational priorities than any corporate statement or policy document ever could.
The departures from OpenAI weren’t just about one company’s internal politics. They were an early warning. The question now is whether anyone with the power to act is paying attention — or whether the competitive frenzy will drown out the signal until it’s too late to matter.
For an industry that prides itself on building intelligence, the inability to solve its own governance problem is a striking irony. And an increasingly dangerous one.
WebProNews is an iEntry Publication