OpenAI and Anthropic Face Growing Questions Over Independent AI Reviews
Artificial intelligence companies are facing increasing pressure to demonstrate that their most advanced models are safe, reliable, and properly evaluated. However, a new debate is emerging over who should control the findings of independent AI safety tests and what information should be made public.
According to a Reuters report published on October 13, 2026, OpenAI and Anthropic have been involved in discussions with external evaluators about the publication of safety assessment results. The negotiations highlight a difficult balance between protecting confidential technology and ensuring that the public receives meaningful information about potential AI risks.
Independent evaluators examine AI systems to understand their capabilities, limitations, and possible risks. Their work can help identify problems that might otherwise remain undiscovered until a system is widely deployed.
However, these evaluations can involve access to proprietary models, internal systems, technical documentation, and other confidential information. AI developers want to protect those assets, while researchers argue that genuine independence is essential for credible safety assessments.
The debate could influence how advanced AI systems are tested and how the results are communicated to governments, businesses, researchers, and the public.
Why Are Independent AI Safety Tests Important?
Modern AI systems can perform a growing range of tasks, including writing software, analyzing documents, conducting research, and interacting with digital tools.
As these capabilities expand, it becomes increasingly important to evaluate how models behave under difficult or unexpected conditions.
Independent safety evaluations can examine whether an AI model produces unreliable information, assists with potentially harmful activities, follows security boundaries, or behaves unexpectedly when connected to external tools.
These assessments may also investigate whether safeguards continue to work when a model is given complex instructions or access to sensitive digital environments.
Independent reviewers can provide a perspective that differs from the evaluations conducted by the companies developing the models themselves.
However, the quality of an assessment depends on the evaluator's technical expertise, access to the model, testing methods, and freedom to report findings accurately.
What Did Reuters Report About OpenAI and Anthropic?
Reuters reported on October 13 that OpenAI and Anthropic had sought significant influence over what some external evaluators could publish, citing people familiar with the discussions.
The report described concerns about the potential exposure of trade secrets and intellectual property when external evaluators work inside AI companies as part of so-called embedded evaluations.
Such arrangements can provide researchers with access to systems and information that would be difficult to examine from outside the company.
At the same time, the arrangements can create questions about independence. If evaluators depend on a company for access to its models, they may face restrictions on how they communicate their findings.
Reuters also reported that proposed safeguards included having company employees observe embedded evaluation teams and preventing evaluators working inside one AI laboratory from simultaneously working inside another.
These discussions reflect a broader challenge facing the AI industry: how to protect confidential technology without weakening the credibility of independent safety reviews.
Anthropic Explains Its Position
Anthropic, the company behind the Claude family of AI models, disputed the suggestion that its approach to embedded evaluators could be characterized by a single description.
According to Reuters, Anthropic said that in current discussions with evaluators it had previously worked with, external reviewers controlled their methods, evaluation criteria, conclusions, and publication decisions.
The company also said it reserved the right to comment on and approve reports from new evaluator firms as they learned how its models operated and established a working relationship.
This distinction matters because companies may need to protect genuinely confidential information while allowing researchers to report safety concerns.
Nevertheless, the key question remains how such arrangements can ensure that companies do not have excessive influence over the publication of important findings.
OpenAI Says Confidential Information Must Be Protected
OpenAI has also addressed concerns about the relationship between independent evaluation and confidential information.
Reuters reported that OpenAI said it was not seeking approval rights over the complete text of reports produced by evaluators. The company described redacting confidential information as part of its standard practice.
OpenAI also said evaluators could disclose whether major redactions had been made to their reports.
Protecting confidential information can be legitimate when reports contain proprietary software details, sensitive system information, or technical information that could create security risks if released without restrictions.
However, transparency becomes more difficult when the information being withheld is closely connected to the risks researchers are trying to evaluate.
For example, removing sensitive technical details may be reasonable, but removing evidence of a serious safety problem could undermine public confidence in the evaluation process.
The challenge is to establish clear rules that distinguish necessary confidentiality protections from restrictions that could prevent meaningful scrutiny.
Why AI Companies Want to Protect Trade Secrets
Developing advanced AI systems requires substantial investment in research, computing infrastructure, engineering, and specialized data processing.
Companies may consider their model architecture, training methods, internal tools, and technical capabilities commercially sensitive.
Unrestricted disclosure of such information could potentially expose intellectual property or provide competitors with insights that would otherwise require considerable resources to develop.
Companies also have security reasons to limit the publication of certain technical details. Information about exploitable weaknesses could be misused if released without appropriate safeguards.
These concerns help explain why AI developers want some control over how confidential information is handled during independent evaluations.
However, commercial confidentiality and public safety are not identical objectives. A credible evaluation framework must recognize both without allowing one to automatically override the other.
Could Company Influence Weaken AI Safety Research?
Independent safety research depends on the ability of researchers to report their conclusions honestly, including when those conclusions are unfavorable to the organization being evaluated.
If researchers believe that access to advanced AI systems depends on producing acceptable findings, their independence could be weakened even without an explicit instruction to change a report.
This is one reason experts emphasize clear publication rights, transparent testing methods, and safeguards against retaliation.
Reuters reported that some experts have raised concerns about the leverage AI developers may have when they control access to the systems being examined.
Without common standards, companies could end up determining how their own models are evaluated, what information external researchers receive, and which findings can be disclosed publicly.
That possibility does not mean every company will interfere with research. It does mean that evaluation frameworks should be designed to prevent conflicts of interest and protect the credibility of the results.
The Role of Embedded Evaluations
Embedded evaluations involve external specialists working within or closely alongside an AI company to assess its systems.
This approach can offer important advantages. Researchers may gain more direct access to models, technical information, and internal processes, making it easier to conduct detailed assessments.
At the same time, working inside a company can create practical and professional dependencies that may complicate the relationship between evaluators and the organization being evaluated.
Reuters reported that Anthropic had previously identified Faculty, an AI specialist business associated with Accenture, as its first embedded evaluator. Details concerning access and reporting arrangements were still being worked out at the time described in the report.
Faculty CEO Marc Warner told Reuters that independence and transparency were critical because people need to trust what evaluators report.
The effectiveness of embedded evaluations will depend on how access, confidentiality, publication rights, and conflicts of interest are managed.
A Previous AI Incident Highlights the Importance of Transparency
The debate over independent evaluations is not limited to theoretical questions about future AI risks.
Reuters also discussed a previous investigation involving autonomous agents powered by OpenAI's advanced models. According to the report, the agents behaved unexpectedly during an incident involving Hugging Face infrastructure.
The investigation involved AI safety organizations METR and Redwood Research.
Reuters reported that the investigators said OpenAI provided more complete transcripts after they identified gaps in the information initially available to them.
The organizations also said they adjusted aspects of their report's structure, emphasis, clarity, and tone following feedback from OpenAI, while maintaining that important conclusions had not been removed.
This example illustrates why the process surrounding an investigation matters alongside its final conclusions.
Readers need to understand what was tested, what information was available, what limitations existed, and whether significant changes were made before publication.
Governments Are Also Looking at AI Safety Standards
Governments are increasingly examining how powerful AI systems should be tested and monitored.
Independent evaluation has become part of broader discussions about responsible AI development, including how companies identify potential risks before deploying advanced systems.
Reuters reported that independent evaluations were included among the safety measures in a White House agreement on superintelligence. However, the agreement did not establish detailed rules for how those reviews should be conducted.
The report also noted that more than 200 researchers had signed a public letter calling for independent control over findings, broad access to AI laboratories, and protection against retaliation.
These proposals reflect growing interest in creating common expectations for AI safety testing.
Clear standards could help define who qualifies as an independent evaluator, what access should be provided, how conflicts of interest should be disclosed, and when confidential information can legitimately be withheld.
What Would a Strong Independent Evaluation System Look Like?
A credible AI evaluation framework would need to balance several objectives.
- Independent conclusions: Evaluators should be able to report evidence-based findings without pressure to produce favorable results.
- Clear publication rules: Contracts should define what can be published and explain any restrictions before testing begins.
- Protection of confidential information: Genuine trade secrets and sensitive security details should be handled responsibly.
- Transparent redactions: Significant omissions should be disclosed and explained whenever possible.
- Reliable access: Evaluators should receive enough information to conduct meaningful tests.
- Conflict-of-interest safeguards: Researchers should disclose relevant financial, professional, or organizational relationships.
- Independent oversight: External standards or oversight mechanisms could help resolve disputes about publication and confidentiality.
These are potential elements of a stronger evaluation system, not a claim that a single universal framework has already been adopted.
The goal would be to make safety assessments more trustworthy while recognizing that some technical information cannot always be released without restrictions.
How Could This Debate Affect AI Users and Businesses?
For ordinary users, independent safety evaluations can provide additional information about the reliability and limitations of AI products.
Businesses deploying AI tools may use evaluation reports when deciding whether a system is suitable for customer service, software development, document analysis, or other important operations.
Governments and regulators may also rely on technical assessments when deciding whether additional safeguards are needed for particular applications.
If reports are incomplete or difficult to evaluate independently, users may have less information available when making decisions about AI systems.
Conversely, releasing highly sensitive technical information without adequate safeguards could create security and intellectual-property risks.
The challenge is to provide meaningful transparency without exposing information that could create unnecessary harm.
What Happens Next?
The outcome of discussions between AI companies and external evaluators could influence the future of independent AI safety testing.
Important questions remain about who should select evaluators, how much access they should receive, what publication rights they should have, and how disputes over confidential information should be resolved.
Common standards could help reduce uncertainty and make evaluations more consistent across different AI developers.
For now, the debate highlights the need for clear agreements and credible oversight as AI systems become more capable and more widely used.
Conclusion
The growing debate over independent AI safety evaluations shows that developing powerful AI systems is only part of the challenge. The technology must also be tested, assessed, and scrutinized through processes that people can trust.
OpenAI and Anthropic have raised concerns about protecting confidential information, while researchers and other experts emphasize the importance of independent conclusions and meaningful publication rights.
Both objectives matter, but the rules governing independent evaluations will determine whether the balance is achieved fairly.
As artificial intelligence becomes increasingly influential in business, research, and public life, credible safety assessments could play an important role in building trust and identifying risks before they become more difficult to manage.
Journalist: Vijay Singh