Anthropic has released an updated version of what they call «Claude Constitutional AI» – a set of principles the model uses to decide how to behave in complex situations. In short: previously, these rules were drafted by a small team inside the company; now, for the first time, ordinary people have been included in the process.
What Is an AI «Constitution» and Why Is It Needed 📜
When you chat with Claude, the model is constantly making decisions: whether to answer or refuse, how to phrase the response, and which topics to consider acceptable. These decisions rely on a built-in set of rules – a sort of «constitution».
Previously, this constitution was compiled by the Anthropic team. They relied on principles from UN human rights documents, Apple's approach to privacy, artificial intelligence ethics, and other sources. The result was a text that defined the model's behavior, but it was written by a few dozen people within the company.
The problem is obvious: AI is used by millions of people in different countries, with diverse cultures and expectations. What seems reasonable to a development team in San Francisco might not align with what users in Brazil, Japan, or Germany want to see.
Что такое «конституция» ИИ и зачем она нужна
How the User Participation Experiment Worked
Anthropic launched a process called «Collective Constitutional AI». The essence is simple: they invited thousands of people from all over the world to have their say on what rules should determine Claude's behavior.
Participants were presented with specific situations – for example, how the model should respond to a controversial question or what to do if a request touches on a sensitive topic. People voted, discussed, and suggested phrasing. They gathered about a million responses.
Then, these responses were processed and converted into an updated constitution. Now it reflects not only the views of the Anthropic team but also the opinions of real users from different regions.
Как организовано участие пользователей в формировании AI Constitution
What Changed in the New Version
Anthropic does not publish the full text of the constitution – it is an internal document. But they did share the key changes:
- The model has started to take greater account of cultural differences. For example, what is considered polite in one country might seem strange in another – now Claude tries to consider this.
- Clearer rules have appeared for situations where there is no unambiguous answer. Previously, the model might have been too cautious and refused where it could have been helpful. Now the balance has shifted slightly towards helpfulness.
- Attention to transparency has been strengthened. If the model isn't sure or the topic requires caution, it should say so rather than just silently refusing.
An important point: the changes do not mean that Claude has become less safe. It is about the model better understanding context and being able to adapt to different requests without losing its core limitations.
Что изменилось в новой версии Claude Constitutional AI
Why This Matters for the Industry
Anthropic's approach shows one of the possible trajectories for AI development. Most companies still decide how models should behave within their own teams. This is faster and simpler, but creates an obvious problem: a small group of people defines the rules for technology used by millions.
«Collective Constitutional AI» is an attempt to make the process more open. Of course, this isn't direct democracy: Anthropic still controls exactly how user opinions turn into rules. But the very fact that the company is ready to ask and consider answers is already a step toward greater transparency.
Other companies are in no hurry to replicate this approach yet. OpenAI and Google use internal processes to tune models, sometimes bringing in external experts, but they don't conduct mass user polls. Perhaps Anthropic is testing a model that could become the standard – or conversely, will show why such an approach is too difficult to scale.
Значение пользовательского участия в развитии ИИ для индустрии
What Remains in Question
Despite the openness of the experiment, details remain undisclosed. We don't know exactly how participants were selected – was the sample random, or did the company specifically seek a balance by country, age, and profession? It is unclear how conflicting opinions were weighted: if one group wants more freedom and another wants more restrictions, who wins?
It is also unclear how often Anthropic plans to update the constitution. If this is a one-off experiment, the new version will quickly become outdated. If it is a regular process, that represents a serious load on the team and participants.
And finally, the main question: how much will the new constitution change Claude's real behavior? Will users notice the difference, or will it remain an internal improvement that only manifests in rare borderline cases?
Вопросы без ответов об эксперименте Anthropic с "конституцией" ИИ
What's Next
Anthropic says it will continue working in this direction. Perhaps the process will become regular, and every major Claude update will include a new round of opinion gathering.
For the rest of the industry, this is a signal: the question of who defines AI behavior rules is becoming increasingly important. Models are being integrated into critical processes – education, medicine, information handling. And if now these rules are written by a few people in a company office, in a couple of years, this might look like a glaring problem.
It is not yet clear if Anthropic's approach will become the standard. But the very fact that the company is trying is interesting enough to keep an eye on.