Small print.
Wide open.

Table of Contents

in  security,  alignment,  safeguards,  and  AI  policy.  This  acts  as  an  internal  spur  for  safety  research  and 
development and is externally visible so customers and regulators can use it to gauge our progress. Consistent 
with this approach, we assess the risks of each model during development and before and after release and we 
take the steps we believe are appropriate to mitigate those risks. In the past, these steps have included delaying 
the  release  of  new  models  and  limiting  their  initial  availability,  such  as  our  initial  release  of  Mythos  through 
Project  Glasswing,  and  in  the  future,  as  we  learn  more  about  model  capabilities  and  behavior,  could  include 
additional  measures  that  affect  how  and  when  we  develop  or  release  these  new  models.  We  believe  these 
practices  also  serve  our  customers,  since  rigorous  safeguards  and  predictable  model  behavior  have  become 
decisive criteria when selecting an AI partner.

Claude’s Constitution

We have developed a constitution for Claude that describes our vision for how an AI model should be and how 
it should behave, which is used directly in the model training process. We have used Claude’s constitution to 
meaningfully advance how frontier models are trained, and it is our attempt to grapple with the most difficult 
questions  regarding  the  development  of  this  new  type  of  artificially  intelligent  entity.  It  is  a  living  document 
that  we  intend  to  revise  as  we  develop  further  models  and  learn  more  about  how  they  work.  Our  researchers 
have  found  that  training  Claude  on  high-quality  constitutional  documents  combined  with  fictional  stories 
portraying an aligned AI contributes to reducing Claude’s misalignment rate. We believe Claude’s constitution 
has also proven commercially valuable: transparently laying out the values we intend Claude to hold increases 
customer trust.

This constitution orients Claude around four priorities:

•

•

•

•

Be Broadly Safe. Avoid irreversible actions and support human oversight of AI behavior.

Be Broadly Ethical. Maintain honesty and sound moral judgment, and avoid harmful acts.

Be  Compliant  with  our  Guidelines.  Adhere  to  our  operational  standards  for  user  interactions  and 
absolute prohibitions on certain categories of harm.

Be Genuinely Helpful. Deliver substantive, expert-level assistance that creates real value for users.

Model Validation and Training Data

Beyond the assessments performed under our Responsible Scaling Policy, we evaluate each Claude model prior 
to  its  release  through  a  structured,  multi-stage  evaluation  program  covering  model  capabilities,  safety 
behaviors, and societal impacts, including bias and discrimination, and we document the results of these pre-
deployment  evaluations,  and  their  limitations,  in  our  publicly  available  model  system  cards.  This  testing 
addresses  the  reliability  and  accuracy  of  model  outputs,  including  through  capability  and  accuracy 
benchmarking,  evaluations  focused  on  factual  accuracy  and  hallucinations,  evaluations  of  whether  model 
outputs are honest and accurately reflect the information on which they are based, reliability testing of agentic 
performance over long-running tasks, and adversarial red-teaming designed to identify ways in which a model 
may  fail  or  behave  in  unintended  ways.  Where  these  evaluations  identify  issues,  we  address  them  through 
additional  post-training,  adjusted  safeguards,  and  re-evaluation  prior  to  release,  and  we  document  residual 
limitations in the applicable model system card. 

Claude models are trained on data drawn from a broad range of sources, modalities, and languages, including 
publicly  available  information  from  the  internet,  datasets  licensed  or  commissioned  from  third  parties,  and 
other sources, such as synthetic data generated by our own models. We conduct diligence on third-party data 
sources to assess their quality, provenance, and suitability for our usage, and in many cases include contractual 
restrictions intended to discourage the inclusion of personal data in sourced datasets. We apply data curation 
and filtering processes to sourced data, which are designed to improve data quality, deduplicate content, and 
reduce potential exposure to harmful content.

We  address  bias  and  fairness  risks  at  multiple  stages  of  model  training  and  deployment.  We  manage  the 
representativeness  and  diversity  of  our  training  data  through  a  layered  approach  in  which  dataset  curation, 
post-training alignment, pre-deployment evaluation, and deployment safeguards each play a part, rather than 
through  dataset  composition  alone.  When  deployed,  our  usage  policy  prohibits  the  use  of  Claude  for 

231