Updated 2026-10-22
from the news desk
How does Anthropic test Claude's values? Anthropic says routine sample review ended
In a 19 October 2026 Frontier Safety Roadmap update, Anthropic said it no longer routinely runs a representative-sample oversight process over post-training data and rewards because repeat checks found nothing beyond existing review. The…
Anthropic says it now tests Claude’s values mainly through automated alignment assessments that score models against Claude’s Constitution, alongside interpretability-based review and red-teaming of its alignment arguments. In a 19 October 2026 Frontier Safety Roadmap update, the company said it no longer routinely runs a representative-sample oversight check over post-training data and rewards because repeat runs found nothing beyond existing review processes.
What Anthropic says changed
On 19 October 2026, Anthropic updated the goal it had called Upholding Claude’s Constitution and marked it retired. In the update, Anthropic states: “We initially ran an oversight process over a representative sample of our post-training data and rewards to check alignment with the Constitution, but found that it surfaced nothing beyond what our existing review processes were already catching.”
The same update adds: “A recent re-run on a sample of training data again flagged nothing.” Anthropic then says: “We are retiring this goal in its current form while we spend more time reflecting on which goals are most effective for alignment.”
That is a change from the earlier public goal text on Anthropic’s updates page. There, Anthropic had said it would run an oversight process over a representative sample of production-relevant post-training data and rewards, review it for egregious inconsistencies with the Constitution that Claude itself could detect, and describe the process in Risk Reports. One reading is that a distinct, named audit layer has been dropped from routine use, even though Anthropic says the broader review work it relied on is still in place.
How Anthropic says it still tests Claude's values
Anthropic’s Constitution page says: “Claude's constitution is a detailed description of Anthropic’s intentions for Claude’s values and behavior.” The same page states that the document plays a crucial role in training, directly shapes Claude’s behavior, and is the final authority on Anthropic’s vision for Claude.
Anthropic’s 19 October update says routine sample review has ended, but Constitution-based testing has not. The update states: “Our automated alignment assessment evaluates alignment with the Constitution as a core metric, and this is reported for all assessed models.” Anthropic also says interpretability methods have become an important part of its overall alignment assessment methodology. In Anthropic’s earlier goal text, interpretability referred to methods meant to produce signal beyond behavior alone.
The same 19 October update says Anthropic has red-teamed its alignment arguments using model organisms and published findings in Risk Reports, though not as extensively as originally intended. Anthropic also says it will keep the public Constitution synced with the most recent trained-on Constitution within 90 days of relevant internal or external deployments. Readers who want the public outputs of that work can compare this change with Anthropic’s discussion of system cards and this site’s page on what Anthropic publishes about Claude.
Why the change cuts both ways
For outside readers, the policy change cuts both ways. The case for concern rests on Anthropic’s own earlier framing. Anthropic had presented representative-sample oversight of post-training data and rewards as a way to evaluate whether training materials matched the Constitution. If that check is no longer routine, then one public-facing mechanism for showing how Claude’s stated values are operationalized has narrowed.
The case for Anthropic’s move rests on the same documents. Anthropic says the initial oversight process found nothing beyond existing review, and that a later re-run also flagged nothing. On that view, retiring a low-yield check may be a pruning decision rather than a retreat from Constitution-based alignment work, because Anthropic says automated assessments, interpretability evidence, and model-organism red-teaming continue.
The development also matters on the AI rights and moral-status beat because Anthropic’s Constitution is not only a safety checklist. Anthropic says it sets out the kind of entity it wants Claude to be, and the Constitution includes a section on Claude’s nature that discusses uncertainty about whether Claude might have some kind of consciousness or moral status. That makes the testing question partly about governance and partly about how a lab publicly ties behavioral evidence to claims about an AI system’s character. Readers following debates over independent evaluation or over whether Claude is conscious may see this as a small but concrete shift in what Anthropic is willing to show, and what it says can be dropped without losing signal.
Frequently asked questions
Did Anthropic stop all Constitution-based testing for Claude?
No. Anthropic says its automated alignment assessment evaluates alignment with the Constitution as a core metric for all assessed models, and the 19 October 2026 update says interpretability evidence is featured in recent system cards.
What process did Anthropic stop running routinely?
Anthropic says it initially ran an oversight process over a representative sample of post-training data and rewards to check alignment with the Constitution. The company says it stopped running that process routinely after it found nothing beyond existing review, and it adds that a later re-run on a sample of training data again flagged nothing.
What is Claude’s Constitution?
Anthropic says Claude’s Constitution is a detailed description of its intentions for Claude’s values and behavior. The Constitution page also states that the document plays a crucial role in training and directly shapes Claude’s behavior.
Will Anthropic keep updating the public Constitution?
Anthropic says that, at a minimum, it will continue to update the public Constitution so it matches the most recent trained-on Constitution within 90 days of relevant internal or external deployments of models trained on updated Constitutions.
Caveats: The central claims here come from Anthropic’s own roadmap and Constitution pages. Anthropic says it plans to publish more about its thinking in the future, so this briefing relies mainly on the company’s account of why the routine oversight process ended.