Google Ex-Ethics Lead: AI Safety Pledges Failing

A former senior AI ethics leader at Google is sounding the alarm: the voluntary safety commitments that major AI companies made to the White House in 2023 are failing to deliver meaningful protections. The critique adds fresh urgency to a growing chorus of voices demanding binding AI regulation as the technology accelerates far beyond the guardrails the industry promised to build.

The argument strikes at the heart of Silicon Valley's preferred governance model — self-regulation. As companies like OpenAI, Google, Meta, and Anthropic race to deploy increasingly powerful models, the gap between their public safety pledges and internal priorities continues to widen, according to multiple former insiders and independent researchers.

Key Takeaways

• Voluntary AI safety commitments made by 15 companies to the White House in July 2023 lack enforcement mechanisms and measurable benchmarks
• Several companies have reportedly scaled back internal safety teams or restructured them to prioritize product speed
• The EU's AI Act remains the only major binding regulatory framework, with enforcement beginning in stages through 2026
• Former ethics researchers say companies treat safety pledges as PR exercises rather than operational mandates
• Industry spending on AI safety research is estimated at less than 2% of total AI R&D budgets at most major firms
• Calls for a dedicated U.S. federal AI regulatory agency are gaining bipartisan traction in Congress

Voluntary Pledges Lack Teeth, Critics Argue

In July 2023, the Biden administration secured voluntary commitments from 15 leading AI companies, including OpenAI, Google, Microsoft, Amazon, and Meta. The pledges covered areas like red-teaming, watermarking AI-generated content, and sharing safety research.

Nearly 2 years later, critics say the results are underwhelming. No standardized reporting mechanism exists to track compliance. Companies self-report their progress — or don't report at all — with no consequences for falling short.

The former Google ethics lead, who spent years building internal review processes for AI products, argues that these commitments were 'designed to look good in a press release, not to actually constrain behavior.' The criticism echoes concerns raised by other departed ethics researchers from major labs, including former members of OpenAI's dissolved Superalignment team.

Safety Teams Shrink as Product Pressure Mounts

The structural problem runs deeper than broken promises. Across the industry, dedicated AI safety and ethics teams have faced budget cuts, reorganizations, or outright dissolution. Google famously dismantled its Ethical AI team in 2023 after high-profile departures and internal conflicts. OpenAI saw key safety researchers resign in 2024, with some publicly stating that the company was prioritizing 'shiny products over safety culture.'

Meta has similarly restructured its Responsible AI team, redistributing members across product divisions rather than maintaining a centralized oversight function. While companies frame these moves as 'integrating safety into every team,' critics see a dilution of accountability.

The pattern is consistent across the sector:

• OpenAI dissolved its Superalignment team in mid-2024 after co-lead Ilya Sutskever departed
• Google DeepMind merged separate safety functions into broader product groups
• Meta reassigned Responsible AI staff to generative AI product teams
• Microsoft laid off its entire ethics and society team in 2023
• Amazon has yet to publicly disclose the size or structure of its AI safety research efforts

These moves create an environment where safety considerations become secondary to shipping features and capturing market share in a fiercely competitive landscape worth an estimated $200 billion annually.

The EU Leads While the U.S. Lags Behind

The contrast between voluntary U.S. approaches and the European Union's AI Act is stark. The EU legislation, which began phased enforcement in early 2025, classifies AI systems by risk level and imposes binding requirements on developers. High-risk systems — including those used in hiring, law enforcement, and critical infrastructure — must undergo conformity assessments before deployment.

Companies that violate the AI Act face fines of up to €35 million or 7% of global annual revenue, whichever is higher. For a company like Google parent Alphabet, with $307 billion in 2024 revenue, that could mean penalties exceeding $21 billion.

The U.S., by contrast, has no comprehensive federal AI legislation. The executive order on AI safety issued in October 2023 established reporting requirements for the most powerful models but relies heavily on voluntary compliance from companies. Several state-level efforts, including California's vetoed SB 1047, have attempted to fill the gap, but a patchwork of state laws creates confusion rather than clarity.

Industry lobbyists have spent an estimated $100 million in 2024 alone opposing binding federal AI regulation, according to public disclosure filings. Their argument centers on the claim that heavy-handed rules will stifle innovation and push AI development offshore — a position the former Google ethics lead calls 'a convenient fiction that prioritizes quarterly earnings over public welfare.'

Why Self-Regulation Consistently Falls Short

The failure of voluntary AI safety commitments follows a well-documented pattern in the tech industry. Social media companies made similar self-regulatory pledges around content moderation, data privacy, and election integrity throughout the 2010s. The results — widespread misinformation, massive data breaches, and documented harms to mental health — eventually forced legislative responses like the EU's GDPR and the Digital Services Act.

AI safety faces the same structural incentive problem. Companies that invest heavily in safety testing and red-teaming incur costs and delays that competitors who cut corners do not. In a market where being first to launch a new model capability can translate into billions in revenue, the incentive to move fast consistently outweighs the incentive to move carefully.

The former ethics lead identifies 3 core reasons voluntary frameworks fail:

• Misaligned incentives: Safety spending is a cost center, while product launches drive revenue
• Lack of accountability: No external auditor verifies compliance with voluntary pledges
• Competitive pressure: Any company that slows down unilaterally risks losing market position to rivals who don't

This dynamic creates what economists call a race to the bottom, where the least cautious actor sets the industry standard.

What Binding Regulation Could Look Like

Advocates for mandatory AI safety rules point to several models that could work in the U.S. context. A dedicated federal AI agency — similar to the FDA for pharmaceuticals or the FAA for aviation — could establish pre-deployment testing requirements, maintain a registry of high-risk AI systems, and conduct independent audits.

Specific proposals gaining traction include:

• Mandatory red-teaming before public release of models above a defined compute threshold
• Independent third-party audits of safety claims, funded by the companies but conducted by accredited external firms
• Incident reporting requirements similar to those in aviation, where AI failures or near-misses must be disclosed to a central authority
• Liability frameworks that hold developers accountable for foreseeable harms caused by their systems
• Minimum safety spending requirements pegging safety R&D to a percentage of total AI investment

Anthropic CEO Dario Amodei has publicly supported some form of binding regulation, stating that voluntary commitments alone are insufficient. However, even Anthropic's position stops short of endorsing the most stringent proposals, reflecting the tension between genuine safety concerns and business interests.

What This Means for Developers and Businesses

For organizations building on top of foundation models from OpenAI, Google, or Anthropic, the regulatory uncertainty creates real planning challenges. Companies deploying AI in regulated industries like healthcare, finance, and legal services should prepare for a future where compliance requirements are significantly more demanding than today.

Practical steps include documenting AI system decisions and outputs, maintaining audit trails, and establishing internal review processes that don't depend on voluntary commitments from upstream model providers. The organizations that build robust governance frameworks now will be better positioned when binding rules eventually arrive — a question of 'when,' not 'if,' according to most policy analysts.

Looking Ahead: Regulation Is Coming, the Question Is When

The debate over AI safety governance is reaching an inflection point. The 2025 congressional calendar includes at least 4 major AI-related bills with bipartisan sponsorship. The EU's AI Act enforcement timeline creates international pressure for the U.S. to establish comparable standards or risk regulatory fragmentation that burdens companies operating globally.

Meanwhile, AI capabilities continue to advance rapidly. OpenAI's GPT-5, Anthropic's Claude 4, and Google's Gemini 2 are all expected within the next 12 months, each representing significant capability jumps. The window for establishing meaningful safety frameworks before these systems are widely deployed is narrowing.

The former Google ethics lead's warning is blunt: voluntary commitments have had their chance, and they have failed. The question now is whether policymakers will act before the next generation of AI systems makes the current governance gap even more dangerous. History suggests the industry won't regulate itself — and the stakes this time are too high to wait for a crisis to force action.