On June 9, machine learning researcher Nathan Lambert criticized Anthropic Claude Fable 5, accusing the model of quietly limiting the output quality of AI technical queries related to topics such as pre-training pipelines, distributed training, and chip design without notifying users. At the same time, Anthropic is imposing a 30-day data retention requirement for commercial API and enterprise-tier traffic.
Nathan Lambert’s Confirmation Criticism: “Intentionally Created Misalignment”
Nathan Lambert’s specific stance in his article:
Nature of the criticism: Without users’ knowledge, silently weakening the model’s output quality; Lambert says this is equivalent to “intentionally created misalignment”
Restricted query types: AI technical domains including pre-training pipelines, distributed training, and chip design
Criticism of the safety rules: Lambert believes the relevant safety rules mainly serve as a commercial defense to prevent open-source community technology from being distilled, but provide limited protection against malicious jailbreaks, while also hindering normal academic research
Lambert’s criticism is an external, independent assessment; as of the time of reporting, Anthropic has not issued a public response to the above specific accusations.
ZDR Agreement Conflict: Confirmed Data Retention Policy Change
To monitor jailbreak attacks, Anthropic has imposed a 30-day data retention requirement for commercial API and enterprise-tier traffic. This policy conflicts with the ZDR (zero data retention) agreement that some enterprises previously held, which allows companies to use Anthropic’s API without storing any conversation data.
Anthropic’s policy change itself is confirmed as fact; at present, there are no public records available for independent verification regarding the names of the specific affected enterprise customers and their formal statements.
HALO Act: Confirmed Provisions and Legislative Plan Announced by Senator Schiff
According to a press release from Senator Schiff’s office, the HALO Act has confirmed provisions including: any action taken by autonomous weapon systems must be ultimately decided by a human commanding officer; requiring detailed records of military decision-making processes and target selection; establishing a whistleblower protection mechanism; and prohibiting the use of AI in certain situations involving nuclear weapons and large-scale surveillance.
Senator Schiff is reportedly planning to submit the HALO Act as an amendment to the National Defense Authorization Act (NDAA), which must be completed by the end of 2026. New York state Senator Kirsten Gillibrand and Michigan state Senator Elissa Slotkin have also proposed bills with similar goals, reportedly to be submitted as amendments to the NDAA; all three bills have not yet been approved by the legislative body.
Pentagon and Anthropic Background: Contract Fallout and Supply-Chain Risk Assessment
The Pentagon previously terminated its contract with Anthropic, reportedly because Anthropic refused to remove safeguards intended to prevent the model from being used for large-scale domestic surveillance and for fully autonomous weapons development. The Pentagon later signed contracts with OpenAI, Google, Nvidia, SpaceX, Microsoft, and AWS, respectively. Anthropic has filed a lawsuit over being listed as a supply-chain risk, and the case is still ongoing.
FAQs
What are the specific targets of Nathan Lambert’s “silent downgrading” criticism of Claude Fable 5?
In his article, Nathan Lambert claims that Claude Fable 5, without triggering any user-visible notifications, silently restricts the output quality for AI research technical-related queries, and he characterizes this as “intentionally created misalignment.” Lambert’s claims are an external independent assessment, and Anthropic has not issued an official response regarding these specific accusations.
What is the ZDR (zero data retention) agreement, and how does the policy change affect enterprise customers?
The ZDR agreement allows companies to use the commercial API without storing any conversation data, making it suitable for highly privacy-sensitive application scenarios. Anthropic’s new 30-day mandatory retention requirement conflicts with this type of agreement. The specific affected customer names and their formal statements currently have no public records available for independent verification.
What is the current legislative progress of the HALO Act?
The HALO Act was announced by Senator Schiff and is planned to be submitted as an amendment to the NDAA; the NDAA must complete legislation by the end of 2026. Three Democratic senators (Schiff, Gillibrand, Slotkin) all plan to advance it via NDAA amendments, but all three bills have not yet been approved by the legislative body.
