Who oversees artificial intelligence? The transparency challenge at big tech companies.

Artificial intelligence is increasingly present in people's daily lives and in companies' operations. From virtual assistants to the automation of complex processes, the technology is advancing quickly and expanding its range of applications. However, as these systems become more sophisticated, a concern grows along with them: who is responsible for checking whether these technologies are actually safe?

Recently, discussions involving OpenAI, the developer of ChatGPT, and Anthropic, the company behind Claude, brought this question to the center of the technology debate. According to information released by Reuters and reported by Olhar Digital in October 2026, the companies are negotiating with independent evaluators the limits of access to information and the conditions for publishing the safety reports on their artificial intelligence models.

The main point of discussion is the balance between protecting confidential business information and the need to ensure transparent, independent evaluations.

Although the companies defend the importance of preserving strategic data, experts warn that excessive restrictions can undermine the credibility of the analyses and make it harder to identify the risks associated with the technology.

The role of independent evaluations in AI safety

Before making an artificial intelligence system available to the public, it is important to evaluate how it behaves in different situations, identifying possible failures, vulnerabilities and unexpected results.

These evaluations may involve security testing, behavioral analysis, bias identification and simulations of situations in which the technology could pose risks to users or to other systems.

Part of this work is carried out by the developers themselves. However, independent evaluations add an important layer of oversight, allowing outside experts to examine aspects that could go unnoticed during internal testing.

In practice, these professionals may investigate questions such as:

• The possibility of a model generating incorrect or potentially harmful information.
• The existence of vulnerabilities that allow the behavior of the artificial intelligence to be manipulated.
• Risks related to privacy and the exposure of confidential information.
• The ability of autonomous systems to take actions not foreseen by their developers.
• The effectiveness of the safeguards implemented by the companies.

To carry out in-depth analyses, evaluators often need access to technical information, operating logs and specific versions of the models.

This is exactly where one of the main challenges arises: how can an independent investigation be guaranteed when access to information depends on the very company responsible for the system being evaluated?

Transparency and confidential information: where is the line?

Developing advanced artificial intelligence models involves significant investment in research, infrastructure and computing resources. For this reason, companies in the sector seek to protect information they consider strategic to their business.

This concern is legitimate. Unrestricted disclosure of certain technical data can expose trade secrets, make it easier to exploit vulnerabilities or harm the competitiveness of organizations.

On the other hand, when companies have influence over what is disclosed in safety reports, questions arise about the independence of the evaluations.

According to the report published by Olhar Digital, Anthropic states that the evaluators it has already worked with retain autonomy over methods, criteria, conclusions and published content. However, in negotiations with new teams, the company foresees the possibility of commenting on and approving reports while the working relationship is being established.

OpenAI, for its part, denies seeking approval of the full documents and states that confidential information may be omitted, as long as evaluators are able to indicate that significant cuts were made.

The discussion highlights a problem that goes beyond these two companies: how can we ensure that safety reports present reliable results when the developers also take part in deciding what information may be disclosed?

One alternative lies in creating clear procedures for audits, with objective criteria for access to systems, the protection of sensitive information and the disclosure of conclusions.

That way, it would be possible to preserve legitimate business data without compromising the technical autonomy of the professionals responsible for the evaluations.

Why does this discussion matter to other companies too?

Although the debate involves some of the largest artificial intelligence developers in the world, its consequences are not limited to the technology sector.

Companies across different industries are incorporating AI tools into their operations, using the technology to automate customer service, analyze information, produce content and support decisions.

This transformation offers important opportunities to increase productivity and streamline processes. However, it also calls for attention to the risks associated with using these systems.

An organization that uses artificial intelligence to process customer information, for example, needs to consider how that data is handled, what protection mechanisms are available and what consequences may arise if something fails.

Likewise, companies that adopt automated systems to support decisions must understand the limitations of the technology and put oversight mechanisms in place.

In this context, three concepts become especially relevant:

Security: involves adopting measures to prevent failures, protect information and reduce the possibility of systems being misused.

Transparency: relates to clarity about how technology solutions work, their limitations and the conditions under which they are used.

Governance: covers the policies, responsibilities and procedures established to guide the development, implementation and monitoring of these technologies.

These principles should not be considered only by the companies that develop artificial intelligence models. Organizations that hire and use third-party solutions also need to assess the risks involved in adopting them.

After all, using a technology developed by another company does not remove the responsibility for how it is applied to internal processes and to customer information.

The future of artificial intelligence depends on trust

Artificial intelligence will continue to play an important role in digital transformation. With the development of more advanced models and systems capable of performing tasks with greater autonomy, the need for reliable evaluation mechanisms is likely to grow.

In this scenario, trust does not depend only on a technology's ability to deliver efficient results. It is also tied to the ability to understand its risks, identify its limitations and verify that its safety mechanisms work properly.

The discussions involving OpenAI and Anthropic show that technological progress must be accompanied by evaluation practices that match the complexity of the systems being developed.

Independent audits, transparent criteria and protecting the autonomy of evaluators are important measures to strengthen the credibility of these technologies.

At the same time, it is necessary to recognize that transparency does not mean indiscriminately disclosing all of a system's technical information. The challenge is to establish mechanisms that allow reliable evaluations without compromising security and the protection of legitimate business information.

For companies that develop or use artificial intelligence, the key takeaway is that innovation and responsibility must go hand in hand.

More than building ever more advanced systems, the future of artificial intelligence will depend on the ability to show that these technologies can be used in a safe, transparent and trustworthy way.

---

Reference source: OpenAI e Anthropic querem controlar relatórios de segurança da IA (in Portuguese), Olhar Digital, published on October 13, 2026, with information from Reuters.