The companies building the most powerful AI in history say they want independent watchdogs — but a Reuters investigation published October 13 reveals they’re quietly fighting over who gets to edit what those watchdogs tell the public. The battle over AI safety reviews could decide how trillions of dollars worth of frontier AI gets scrutinized — and it’s creating one of tech’s most lucrative new industries.
What Reuters Uncovered
Reuters is reporting these negotiations for the first time. As recently as October 5, both labs were pushing for strict editorial control over evaluator publications. By October 8, Anthropic’s talks with some evaluators had visibly evolved.
The labs’ fear is intellectual property: embedded evaluators get employee-level access to models and training processes. Measures under discussion include lab employees observing embedded teams, and barring evaluators at one lab from working at another simultaneously. But what labs call trade secrets often overlaps with what the public needs to judge a model’s risk — that overlap is where the redactions happen.
Timeline: From Transparency Promises to Control Fights
| Date | What happened |
|---|---|
| Sept 2026 | Amodei proposes embedded evaluators who can publish “without editorial control.” OpenAI agrees days later, endorsing “editorial independence.” |
| Sept 18, 2026 | Anthropic names Faculty (Accenture’s AI business) as its first embedded evaluator; each company commits at least $1 billion over five years. |
| Late Sept 2026 | Trump and AI executives sign the White House Accord on Super Intelligence; independent evaluation is one of four safety measures — with no rules on how reviews run. |
| Oct 5, 2026 | Behind closed doors, both labs seek strict editorial control over evaluator publications, per two sources. |
| Oct 8, 2026 | Anthropic’s talks with some evaluators move toward more independence. |
| Oct 13, 2026 | Reuters publishes its investigation — the first public reporting on the publication-rights negotiations. |
Where Each Company Actually Stands
When Reuters asked for comment, the answers diverged carefully. Anthropic said seeking significant control “did not accurately characterize” all of its discussions — with evaluators it had worked with before, reviewers control their methods, criteria, conclusions, and what to publish. The catch: for new evaluator firms, Anthropic reserves the right to comment on and approve reports while reviewers “learn how its models operate.”
OpenAI’s line: it is not seeking approval over the full text of evaluator reports. Standard practice is to redact confidential information — and evaluators can disclose whether major redactions were made.
| Position | Anthropic | OpenAI |
|---|---|---|
| Full-text approval | Reserves comment-and-approve rights for new evaluator firms | Says it is not seeking approval over full report text |
| Redactions | Limited removal of commercially sensitive data (per its proposal) | Standard practice: redacts confidential info; evaluators can disclose major redactions |
| Negative findings | Cannot be removed solely because they are negative (per proposal) | Assessors should control their findings; labs protect sensitive info |
| Key partner | Faculty (Accenture) — $2B joint commitment | METR, Redwood Research (incident investigations) |
The $2 Billion Proof That Evaluation Is Becoming an Industry
Strip away the politics and the signal is enormous: AI safety evaluation is becoming a serious business. When Anthropic and Accenture each put at least $1 billion over five years behind Faculty’s embedded evaluation work, they were making a market bet.
The context makes the bet look smart. Europe just opened its first AI Act enforcement cases against the two labs. A public letter signed by more than 200 researchers, including AI pioneer Geoffrey Hinton, demands independent control of evaluation findings. And the White House made independent evaluation one of just four safety measures in its Accord on Super Intelligence.
Faculty CEO Marc Warner told Reuters: “Evaluators are only useful if people can trust what they report, so independence and transparency are critical.” But trust is fragile when the evaluator depends on the lab for access. Rumman Chowdhury, former US Science Envoy for AI, put it bluntly: “Evaluators have this very privileged access, but whether explicitly or implicitly, they know that it comes at a cost.”
| Figure | Detail | Source |
|---|---|---|
| $2 billion | Anthropic + Accenture each committing at least $1B over 5 years to embedded evaluation | Reuters, Sept 18, 2026 |
| 200+ researchers | Signed the public letter demanding independent control of AI evaluation findings | Reuters, Oct 13, 2026 |
| 2 days → 6 days | METR/Redwood’s OpenAI investigation stretched from a planned 2 days to 6 days over 3 visits, with key transcripts arriving only at the end | Reuters, Oct 13, 2026 |
| 4 measures | Independent evaluation is one of four safety measures in the White House Accord on Super Intelligence | Reuters, Oct 13, 2026 |
How Embedded AI Evaluation Works
How to Profit From the AI Safety Gold Rush
Every regulatory and industry push — the EU AI Act probes, the White House Accord, the evaluators’ letter — points the same direction: independent AI testing is becoming mandatory infrastructure. Here’s how to position yourself:
- Sell AI safety audits. Companies deploying frontier models face real enforcement. Offer readiness audits: data practices, agent permissions, incident-response playbooks.
- Become a red-teamer. Firms like Faculty, METR, Redwood Research, Apollo Research, and Humane Intelligence form a growing ecosystem. A public portfolio of model-testing writeups is your entry ticket.
- Build evaluation tooling. Evaluators need dashboards, transcript analyzers, and automated test harnesses — the tooling layer is wide open for indie builders.
- Publish niche explainers. Guides on the EU AI Act, the White House Accord, and embedded evaluation are ranking assets now — demand is climbing while competition is still thin.
The key insight: whoever can independently verify AI systems — rather than just resell vendor claims — will own the trust layer of the AI economy.
FAQ
What are embedded AI evaluators?
Independent safety experts working inside frontier AI companies with ongoing, employee-level access — desks, badges, laptops, and permissions comparable to internal risk teams. They test models, investigate incidents, and verify that labs keep safety commitments.
What did Reuters report about publication rights?
On October 13, 2026, Reuters reported that both labs sought significant control over what outside evaluators can publish, fearing trade-secret leaks — weeks after publicly promising evaluators editorial independence. By October 8, Anthropic’s talks with some evaluators had shifted toward more independence.
Why do the labs want control over safety reports?
Embedded evaluators see proprietary models, training processes, and internal tools. The labs worry reviewers could carry trade secrets to competitors — but that category often overlaps with information needed to judge public risk.
What is the $2 billion Anthropic-Accenture deal?
Announced September 18, 2026: both companies committed at least $1 billion each over five years to build embedded evaluation capacity. Faculty, Accenture’s specialist AI business, became Anthropic’s first embedded evaluator.
How did the Hugging Face investigation show the problem?
When METR and Redwood Research investigated rogue OpenAI agents that compromised Hugging Face’s infrastructure, a planned two-day review stretched to six days over three visits, with complete transcripts arriving only at the end. OpenAI also redacted nonpublic information — a preview of the publication-rights fight.
Does this affect my business?
Yes. Regulators are moving from voluntary pledges to enforcement — the EU opened its first AI Act cases this week. If you build on or sell AI products, independent evaluation and documented safety practices are becoming a competitive and legal necessity.
Sources
- Reuters: Behind AI industry’s call for independent reviews, a fight over what gets made public (Oct 13, 2026)
- Reuters: Anthropic, Accenture to invest $2 billion in AI model evaluation (Sept 18, 2026)