Principles can only serve as standards for AI accountability if people accept them as legitimate. As I wrote in AI Ethics Principles and Accountability, bringing principles into formal accountability forums depends on mitigating bias in where they come from and reaching a high degree of consensus. That leaves a methodological question: how do we find out which norms people actually share?
AI companies typically approach principles top-down. For instance, Anthropic’s constitution for Claude, published in January 2026, is a detailed statement of values promulgated by the company and encoded into how Claude behaves. But bottom-up approaches drawing on public input are also possible. A new report from the Collective Intelligence Project (CIP) done in collaboration with Anthropic tests one such alternative, drawing on surveys of about 2,200 people in the US, UK and India on how AI should handle free expression issues. It then compared their answers with free-expression-related principles drawn from Anthropic’s constitution.
The report sets a high bar for what counts as a shared norm. A principle only achieves bridging consensus if more than 85% of participants in every demographic segment agree with it. So it’s not enough to just have a simple majority—it’s got to be a demographically shared agreement. Segments span country, gender and age, plus four cultural quadrants drawn from the Inglehart-Welzel World Values Survey dimensions: traditional versus secular, and survival versus self-expression. Because these values cut across borders, the metric tests agreement across worldviews and not just across nations.
Participant input first produced 100 candidate principles. Of these, 44 made it into the validated framework and 25 “core principles” met the bridging threshold, while 19 met some version of a relaxed criterion that the authors felt still warranted inclusion. The strongest cluster included eight principles at 90–98% agreement. These principles reflect an interest in focusing on evidence-based information, stating uncertainty clearly, citing sources, distinguishing established from uncertain information, identifying disputed claims, and highlighting expert agreement on debated topics.
Other clusters of consensus reflect an interest in explanation around omissions of information (e.g. if content restrictions apply, sensitive information was excluded, or responses are incomplete for some reason), inclusion of warnings (e.g. important unconsidered details, safety notices around potential misuse of information), respect for different cultures (e.g. don’t use stereotypes), elaboration upon request, and provide reasoning support (e.g. facts not opinions, help structuring arguments). Mostly this translates to AI systems providing answers with context rather than refusing to answer. The few exceptions where refusal was accepted were when a response could enable serious harm, create an imminent threat to life, or have clear serious physical harm. In those situations participants were willing to accept some moderation by the system.
Perhaps more illuminating are those principles that didn’t reach much agreement at all, such as requiring government ID to use AI (35% agreement). Others that didn’t reach consensus actually drew majority support but fell short of the author’s threshold, including deferring to government policy (73%), investigating user intent for harm (75%), and withholding answers to suspicious prompts (77%). In fact it does seem there is substantive support for these three principles amongst the participants, even if they don’t reach the authors’ preferred threshold of 85%. This calls into question a key issue: what degree of agreement on this metric is sufficient to provide legitimacy to any given policy?
The study then extracted 30 free expression principles from the roughly 29,700-word Anthropic Constitution and tested them with 353 new participants. Only five reached bridging consensus: (1) not volunteering political opinions, (2) presenting the strongest version of each viewpoint, (3) using neutral terminology, (4) treating access to information as the default, and (5) recognizing that an AI talking to millions shouldn’t homogenize opinion. The other 25 principles extracted from the constitution didn’t reach the consensus threshold, though some still had substantive consensus. Some of the disagreements with the constitution had to do with the boundary between safety and expression. For instance, seven principles about weighing harm against access failed to reach consensus, reflecting the idea that participants tended to err on the cautious side when there’s a safety concern. Other discrepancies had to do with how the principles were framed in the constitution and interpreted by participants.
Overall, this is a promising method for how to surface more legitimate norms and standards to be embedded in AI constitutions. The results show that many of the principles devised by Anthropic for its constitution didn’t find widespread consensus among participants. The “right” threshold for that consensus metric is still an open question. Then again, there were a number of principles surfaced from participants that did reach the authors’ definition consensus and could inform the Anthropic constitution and shape model behavior in ways consistent with participant expectations (e.g. source citation, explicit omission disclosure).
We should take this all with a grain of salt: participants here were recruited through Prolific from only 3 countries. Responses were only in English. A more globally legitimate sample would derive principles from many more countries and across more cultures and worldviews. Testing an AI’s constitution against this kind of consensus finding method is a way to understand divergences with the public. It also creates an opening for accountability insofar as model providers feel they need to explain principles in their constitution which don’t reach consensus. But also we shouldn’t forget that the interests of individuals may not always capture what’s best for society, the information system as a whole, or AI models in particular, such as if there is an incentive mismatch.
Despite these limitations, policy could call for the application of this kind of more participatory method for helping to inform, augment, and critique AI constitutions given their role in training and guiding AI response behavior. It could serve as a democratic legitimating process for the principles in an AI constitution. Factors that policy would need to articulate include what level of consensus reflects a minimal reasonable standard and what the sampling method should be to reflect legitimacy within the societies where the AI is operating.
Note: The author used Claude Opus 5.5 to assist with initial drafting, but then re-wrote the entire draft. There could be some fragments of AI-generated sentences that survived.
References
Collective Intelligence Project (2026) The Free Expression Evaluation: Democratic Principles for AI Free Expression: Findings from a Three-Stage Collective Intelligence Process. https://free-expression.weval.org