The AI Industry's Transparency Conundrum: Balancing Safety with Secrecy

Last month, both Anthropic and OpenAI wanted independent scrutiny, provided independence does not extend to deciding what the public gets to know. They want the public to believe that their systems are being tested rigorously, that risks are taken seriously, and that the industry is capable of policing itself. Yet the moment an outside auditor steps in, things get murkier. Who gets the final say over what the public is allowed to know? Shouldn’t the AI safety assessment be an independent exercise or should it be a self-certification by the companies? Anthropic CEO Dario Amodei has publicly argued for evaluators to publish key findings without editorial control, while OpenAI has called for editorial independence, which is why this question becomes fundamental. Their technical expertise is considerable and their commercial interest is served only when their products are seen as safe. Don’t assume that the first cancels out the second. Imagine asking a pharmaceutical company to fund a drug trial, select the investigators, decide which findings may be published, and then assure patients that the process was independent. In AI, the boundaries between funding , testing, and controlling the message are still being negotiated. There are legitimate reasons for confidentiality. Publishing the exact weaknesses of a powerful model, the steps needed to exploit them, or the details of a dangerous capability could do real damage. An evaluator should not be able to release a technical instruction manual for misuse and defend it as transparency. But that is an argument for carefully limiting what is disclosed. A distinction needs to be made between protecting technical details and protecting a company from embarrassment. If an evaluator can be prevented from publishing a material finding because the company disagrees with its wording, the public is entitled to ask what independence means in practice. You cannot accept a test that satisfies the subject of the test. The answer cannot be to publish everything, regardless of consequence. Nor can it be to trust the companies to decide, behind closed doors, which findings deserve to see daylight. The rules need to be settled before the company and its evaluator start arguing over what the public gets to hear. Independent evaluators should have control over their technical conclusions. Firms should be entitled to challenge errors, supply additional evidence, and identify information that would create a credible security risk if released. Correcting facts is not editorial control. Also, a dispute over wording shouldn’t bury a finding. Where disclosure could enable harm, an independent reviewer with appropriate security expertise could determine what should be withheld from the public version. The evaluator’s conclusion, the reasons for any redactions, and the existence of a material disagreement should still be recorded. I am not arguing against the AI industry but for taking its own safety claims seriously. The same principle applies to robots and other AI-enabled systems. On October 10, in Cologne, Germany, a humanoid robot allegedly groped a 16-year-old girl while two young men filmed the incident. This incident should not automatically be treated as proof that AI systems are inherently dangerous. One incident rarely establishes a general conclusion. But neither should an incident be dismissed as an isolated case before anyone has examined what happened, what failed, and whether the same weakness might exist elsewhere. India should pay attention before it imports the industry’s habits along with its technology. The country is encouraging AI adoption across business and public services, while trying to work out how much oversight is necessary without making innovation prohibitively expensive. A workable Indian framework could begin with a modest requirement for high-impact AI systems: independent risk assessments, clear rules on publication, mandatory reporting of serious incidents, and protection for evaluators or employees who raise concerns in good faith. The framework should distinguish between a system recommending a restaurant and one influencing a medical decision, a credit assessment or a public-service entitlement. Each of these require different levels of scrutiny. Smaller firms should not be buried under compliance costs designed for technology giants. Equally, a company shouldn’t escape scrutiny simply because its product is technically complicated or commercially important. India also needs to decide who qualifies as an independent evaluator, how conflicts of interest are disclosed, and what happens when a company refuses to cooperate. Without answers to these questions, an audit requirement ends up as one more box to tick. None of this requires the government to publish sensitive model details or demand access to every commercial secret. It requires the rules to separate legitimate secrecy from the suppression of evidence. Independent audits are worthwhile only if the audited companies cannot decide what the auditor is allowed to say. AI businesses will understandably worry about reputational damage, litigation, and the possibility that rivals might exploit disclosed weaknesses. Those concerns deserve a hearing. The public is being asked to trust systems that are becoming embedded in decisions it may not understand and cannot easily challenge. Trust cannot rest solely on the assurances of the people selling those systems. AI leaders have an opportunity to set standards before a scandal, accident, or preventable injury forces governments to impose them. Herein lies the opportunity. If they insist on independent scrutiny, they must also accept that independence means the evaluator may publish something the company would rather not read. India must not let AI companies write their own report cards, grade their own homework, and then demand a standing ovation for transparency. Trade secrets must be protected; public safety needs independent audit. As a senior bureaucrat puts it, “When the accused controls the inquiry, the verdict is ready before another Nilekani committee is constituted.” The writer is Fortune-500 advisor, start-up investor, and co-founder of Medici Institute for Innovation Disclaimer: The views expressed are the author’s own and do not reflect the official policy or position of Financial Express.