Anthropic and OpenAI both want “independent” safety evaluators embedded in their operations. Here is the problem: when the companies being watched fund the watchers, pick the watchers, and set the criteria, you don’t have oversight. You have a PR strategy worth billions.
Why This Is Happening Now
This push didn’t start from genuine concern. It started from regulatory pressure. The EU AI Act, which began enforcement in 2024, requires conformity assessments for high-risk AI systems, according to the European Commission. In the United States, executive guidance from two administrations created reporting requirements for frontier models, with the Trump-era rules softening enforcement while keeping most of the structure, according to the National Institute of Standards and Technology.
Both Anthropic and OpenAI responded by proposing “independent” safety boards and external red team evaluators. OpenAI launched its Safety Advisory Group in 2024. Anthropic has pushed for third-party audits of its Constitutional AI systems. On paper, this looks like accountability. In practice, the structure tells a different story.
The companies being evaluated often pick the evaluators, set the scope, and control what gets published. That’s not independence. That’s a veto dressed up as a review.
Who Pays the Watchdog Wins
I’ve seen this movie before. In finance, credit rating agencies like Moody’s and S&P were paid by the very banks whose mortgage products they were rating. According to the Financial Crisis Inquiry Commission, those agencies handed out AAA ratings to securities that later collapsed and contributed to an estimated $22 trillion in global economic losses. The companies being rated funded the raters. The raters needed the business. Nobody wanted to fail the golden goose.
AI safety evaluations are structured the same way. Anthropic raised approximately $7.3 billion through 2024, according to Crunchbase. OpenAI’s valuation crossed $150 billion in late 2024, according to reporting by The Wall Street Journal. These aren’t scrappy startups seeking legitimacy. These are massively funded companies with investors who expect returns. An independent evaluator who shuts down a key product line isn’t getting invited back.
The conflict runs deeper than money. Most people qualified to evaluate frontier AI systems work at the labs being evaluated or consult for them. The talent pool is tiny. According to a 2025 survey by the AI Now Institute, fewer than 400 researchers globally have the technical background to audit large language models at the frontier level. Nearly half of those researchers have financial ties to Anthropic, OpenAI, Google DeepMind, or Meta. So when these companies say they want truly independent evaluators, they’re drawing from a pool that’s already compromised.
The rich mindset sees this clearly: control the evaluation framework and you control the regulatory outcome. If Anthropic gets to define what “safe” means and pick who measures it, they’ve bought a compliance certificate, not a safety guarantee. Their investors get protection from liability. Their competitors face a higher barrier to entry. And the public gets a press release.
The average person reads “AI safety board” and thinks someone is watching out for them. They’re not. They’re watching out for the board’s contract renewal.
If you’re already thinking about what these AI companies know about you and how that data exposure could affect your financial life, it’s worth setting up IdentityIQ credit monitoring. It won’t stop AI from training on your public data, but it will alert you when your personal financial information surfaces in breaches that tend to follow large tech companies around.
What This Means For You
If you’re building on top of OpenAI or Anthropic APIs, a captured safety board produces no real risk disclosure. You won’t get early warning when these systems fail in ways that expose your business to liability. That risk is priced into nothing right now.
Here is what I would do. First, don’t treat any AI company’s safety claims as ground truth. Read the actual audit reports when they’re published. Look at who funded the audit, who set the scope, and what was explicitly excluded. If those details aren’t public, assume the answer is unflattering.
Second, if you’re making financial decisions based on AI company valuations or betting your business infrastructure on one provider, think about diversification now. A company that controls its own safety narrative can also hide problems until they become public and expensive. If you need capital to build redundancy into your AI stack, comparing your options through SuperMoney loan comparison will surface better rates than going straight to your bank with no.
Third, pay attention to who’s actually pushing for stronger independent oversight. The most credible voices tend to be researchers who have left these labs without golden parachutes and academics with no funding ties to frontier AI companies. When they say something is dangerous, that’s worth real weight. When the labs say they evaluated themselves and everything is fine, weight that accordingly.
The regulatory environment will keep tightening. The EU AI Act is already generating compliance costs. The US is moving toward similar requirements. Companies that built genuine internal safety culture will adapt. Companies that built safety theater will scramble. Picking the right vendors now is cheaper than unwinding a bad dependency later.
The Bottom Line
Anthropic and OpenAI don’t want safety evaluators. They want safety credentials. One protects the public. The other protects the company from the public. Until evaluators are funded by governments or neutral third parties, set their own scope, and can publish findings without company approval, you’re not looking at oversight. You’re looking at a very expensive press release. I’ll believe it’s independent when the evaluators can say “shut it down” and make it stick.
Frequently Asked Questions
What are AI safety evaluators and why do they matter?
AI safety evaluators are researchers or organizations hired to test AI systems for harmful outputs, bias, and security vulnerabilities. They matter because the findings can determine whether a product gets deployed to millions of users or gets held back. When the company being evaluated controls the process, those findings tend to be optimistic.
Are Anthropic and OpenAI legally required to have safety evaluations?
In the European Union, the AI Act requires conformity assessments for high-risk AI systems, according to the European Commission. In the United States, there is no equivalent legal requirement yet. The evaluations both companies are running are largely voluntary and structured on their own terms.
Can AI companies self-certify their own safety reviews?
Right now, yes. In the United States, no law prevents companies from designing and publishing their own safety evaluations. The voluntary frameworks encourage third-party involvement, but companies still control access, scope, and publication. Until regulators mandate truly independent audits with public reporting requirements, self-certification stays the norm at frontier AI labs.
How is this similar to financial conflicts of interest I should know about?
The closest comparison is the credit rating agency problem before the 2008 financial crisis. Agencies were paid by the banks whose products they rated, which created pressure to give favorable ratings. According to the Financial Crisis Inquiry Commission, this conflict was a direct contributor to the housing collapse. AI safety evaluations face the same structural incentive problem today.
What can I do to protect myself financially from AI company risks?
Diversify your vendor dependencies, read third-party reporting instead of relying on company press releases, and monitor your personal data exposure through tools like IdentityIQ credit monitoring. If you’re investing in or building on top of these companies, price in the regulatory and reputational risk that comes with self-certified safety records.
