AI: The battle for control over external evaluation

The AI safety paradox: Between promised transparency and corporate control

For years, cutting-edge labs have positioned safety as the cornerstone of their public narrative. OpenAI and Anthropic, in particular, have advocated for the imperative need for external audits to mitigate existential risks and social biases. However, a recent journalistic revelation—corroborated by reports from Reuters and The Next Web—suggests a concerning strategic shift: these companies are pushing to exert direct control over what information independent researchers can publish after auditing their systems. This move is not merely a technical dispute, but a paradigm shift in AI governance that threatens to turn safety into a marketing tool rather than a real safeguard.

What has really happened? The clash between trade secrets and public safety

Tension has escalated following negotiations between the labs and external evaluators. According to sources close to these conversations, both OpenAI and Anthropic have sought to impose pre-publication review clauses that grant them veto power over published findings. The official argument is the protection of intellectual property and trade secrets; a justification that, while legitimate in other software sectors, is problematic in the context of generative AI, where safety risks are intrinsically linked to the model's architecture.

This stance contradicts previous promises of openness. Just a month ago, Anthropic executives, including CEO Dario Amodei, had publicly reaffirmed the commitment to grant researchers almost total operational freedom. The change in tactics suggests that as models become more powerful and commercially valuable, the labs' tolerance for critical scrutiny has diminished drastically. Historically, this behavior is reminiscent of the tobacco or pharmaceutical industries, where control over clinical trial data was systematically used to hide adverse effects under the guise of 'confidential research'.

The conflict of interest: Judge and jury

We are facing a structural conflict of interest. AI safety requires radical transparency to identify vulnerabilities before they are exploited. However, by restricting the right to publish, the labs act as both judge and jury: they design the model, they set the audit parameters, and now, they decide which parts of the audit are 'safe' for public consumption. This dynamic creates an ecosystem of 'security theater', where audits can be filtered to highlight successes and hide systemic failures, thus protecting the company's market valuation over the end user's safety.

Implications for the future of the sector: A systemic erosion

The impact of this trend is profound and multifaceted, affecting not only the scientific community but public trust in technology:

• Erosion of trust: If external evaluation is mediated by the companies, it loses its value as an independent oversight mechanism. An audit that cannot be critical ceases to be an audit and becomes a public relations exercise.
• Risk of corporate bias: Researchers, who often depend on the infrastructure, data, and funding of the labs themselves, could be incentivized (or coerced) to soften their conclusions to maintain access to future versions of the models. It is a risk of private regulatory capture.
• Regulatory obstacles: This corporate closed-mindedness is, ironically, the greatest accelerator of harsh state regulation. If labs refuse to be audited with full transparency, governments (such as those in the EU under the AI Act) will have no choice but to impose forced audits, eliminating the possibility of self-regulation and creating a rigid bureaucracy that could stifle innovation.

Comparing this event to previous technological crises, we observe a parallel with the rise of social media in the 2010s. Just as then, the industry seeks to avoid external oversight by appealing to technical complexity, a mistake that resulted in a decade of misinformation, algorithmic bias, and privacy issues that we are still trying to solve today.

What should TheVortiq readers know?

We are at a critical moment for the future of work and society. The AI industry is maturing and, as often happens in any capitalization phase, priorities have shifted: the protection of assets over the common good is now the norm. We cannot allow 'AI safety' to become an opaque black box.

As users, professionals, and citizens, we must demand evaluation mechanisms that do not depend on the labs' approval. Researcher autonomy is the only real check against potential technological catastrophes or systemic biases. Transparency is not a corporate luxury or an obstacle to progress; it is the foundation upon which the massive and ethical adoption of artificial intelligence is built. Without it, AI will remain a powerful, but dangerously uncontrolled, tool.