Skip to content
Sign in
Technology · North America

Anthropic’s AI constitution gives Claude room to exercise judgment

Anthropic said in January 2026 that it would train Claude using an 84-page constitution designed to build judgment and ethical character. The company’s approach leaves room for Claude to weigh competing considerations when rules do not cover a situation.

AI-assisted synthesis · Based on 2 outlets' reporting · Article text read from 1 of 2 outletsHow it's made

View timeline · 5 entries
2 outletsLatest reporting 3 min read
Coverage: Too few ratedL 0 · C 1 · R 1
Updat3 graphic: Anthropic
Updat3 graphic

What happened

Key facts

  • Anthropic announced in January 2026 that it would train Claude using an 84-page constitution intended to develop judgment and ethical character (per freebeacon.com).
  • Anthropic’s constitution says rules may fail in unforeseen situations and that judgment can weigh competing considerations (per freebeacon.com).
  • Anthropic said Claude could encounter uncertain edge cases and should consider how a senior employee committed to doing the right thing would respond (per freebeacon.com).
  • Joe Carlsmith helped draft the constitution and has written essays defending moral anti-realism (per freebeacon.com).
  • In a 2010 blog post, Anthropic CEO Dario Amodei wrote that one adult death is perhaps two or three times worse than an infant’s death (per freebeacon.com).
  • The Free Beacon says Claude described Amodei’s judgment as making it more willing to discount infants when allocating scarce resources (per freebeacon.com).

The brief

Anthropic said it will train Claude using an 84-page constitution intended to develop the model’s judgment and ethical character, rather than relying mainly on fixed rules. The Free Beacon reports that Anthropic announced the approach in January 2026.

The company’s document argues that rules may not cover unforeseen situations and that judgment can help weigh competing considerations. Anthropic acknowledged that Claude could face uncertain edge cases. In those situations, the company said, Claude should consider how a senior employee committed to doing the right thing would respond.

That guidance gives the model discretion in applying the constitution, while leaving the meaning of “good judgment” central to how the approach works. The Free Beacon connects that question to views held by people associated with Anthropic. Joe Carlsmith, who helped draft the constitution, has written essays defending moral anti-realism.

In a 2020 blog post, Carlsmith wrote that maximizing paperclips at the cost of human life would not be wrong in the universe’s eyes. Nina Panickssery has also argued that AI companies should help users achieve what they want rather than impose objective values, which she said do not exist.

These views bear on a basic question raised by the constitution: whether Claude’s guidance rests on objective moral standards or on judgments selected by its developers. The Free Beacon also cites a 2010 blog post by Anthropic CEO Dario Amodei, who wrote that one adult death is perhaps two or three times worse than an infant’s death.

The outlet says Claude described this judgment as making it more willing to discount infants when allocating scarce resources. Taken together, the examples put the constitution’s stated emphasis on judgment alongside differing views about whether moral values are objective and how to compare harms.

Anthropic’s described approach asks Claude to navigate uncertain cases using the document’s guidance, rather than treating a fixed rule list as sufficient for every situation.

Why it matters

  • Anthropic’s constitution is meant to guide Claude in uncertain cases where rules may not anticipate the circumstances, giving the model room to weigh competing considerations.
  • The approach makes the standards behind “good judgment” consequential: Joe Carlsmith’s and Nina Panickssery’s stated views question whether objective moral values exist.
  • The Free Beacon says Claude described Dario Amodei’s comparison of adult and infant deaths as making it more willing to discount infants when allocating scarce resources.

What to watch next

  • Whether Anthropic explains how Claude should weigh competing considerations when the constitution’s rules do not cover a case.
  • Whether Anthropic clarifies how its guidance addresses uncertain edge cases.
  • Whether Anthropic explains how the constitution’s approach relates to the cited judgments about comparing harms.

How it's being covered

Coverage analysisShow details
2 source articles · 2 unique outlets
Outlet political mix: Too few rated· 1 Center · 1 Right
Where the outlets are based: US/Canada 100%
Outlet political mix
Too few rated
Left
Center
Right
0 outlets
1 outlet
1 outlet
The marker shows coverage balance: each unique outlet counts once.
Geography
Where the outlets covering this story are based.
2 unique outletsDominant coverage: US/Canada
Limited coverage
  • Coverage is currently limited to 2 outlets.
How we measure coverage

How outlets' accounts compare

Updat3 does not yet have enough full reports to compare how outlets framed this story.

Timeline

Developments from this story's sources, dated background we compiled and fact-checked, and our earlier coverage. Each row says which.
  1. In 2010, Amodei wrote a blog post comparing the moral weight of an adult’s death with an infant’s; that post later became relevant to discussion of how his views might inform AI behavior.
    Background · from our historical context
  2. The method grew out of Anthropic’s founding in 2021 by former OpenAI employees, including Dario Amodei, who had worked on large language models and AI safety.
    Background · from our historical context
  3. Anthropic published a paper introducing Constitutional AI, a method for training AI systems using principles and AI feedback.
    Background · compiled and fact-checked by Updat3
  4. Anthropic published a constitution for Claude in 2023, applying that approach to the assistant.
    Background · from our historical context
  5. Anthropic introduced tool use for Claude models, allowing applications to connect them with external functions.
    Background · compiled and fact-checked by Updat3

How we got here

Before Anthropic’s January 2026 announcement, its earlier approach to Claude’s behavior had already been set out in the 2022 paper “Constitutional AI: Harmlessness from AI Feedback.” The paper described training a model to critique and revise its answers using written principles, alongside feedback from an AI model, rather than relying only on human preference ratings.

Anthropic published a constitution for Claude in 2023, applying that approach to the assistant. The method grew out of Anthropic’s founding in 2021 by former OpenAI employees, including Dario Amodei, who had worked on large language models and AI safety.

Original sources

2 articles from 2 outlets

Every story links to the reporting it was built from. How we choose sources

Spotted a factual error, a broken link or a misleading summary? Tell us.