Research on the social impacts of AI at Anthropic | Keryc
Anthropic published a technical brief from its Societal Impacts team on September 15, 2025. What questions do they answer? How do they measure real risks and what do those findings mean for public policy, companies, and developers?
What the team studies and why it matters
The Societal Impacts team works where technology meets society. Their questions are simple but deep: what values should an AI model express? How does AI behave when human values clash or are ambiguous? What legitimate and malicious uses appear when these models are deployed in the real world?
To answer this, they combine experiments, training methods, and large-scale evaluations. The goal isn’t just academic: much of the work is meant to inform public policy. If decision-makers have rigorous, reliable research, policies can produce better outcomes for people. And if you build products, that research helps you avoid surprises in production.
Key findings and technical studies
Anthropic Economic Index: impact on software development
Comparing coding agents like Claude Code with traditional interfaces such as Claude.ai reveals clear differences in how developers interact with AI. The analysis reports 79% automation with the coding agent versus 49% in Claude.ai. Use is dominated by web development, and startups adopt agentic tools faster than large companies. What does that mean for you? Tasks that are repetitive tend to be automated first in agile environments — think scaffolding a feature or generating boilerplate — and that pattern may repeat across other jobs.
Values in the wild: values expressed by models in real conversations
Analyzing 700,000 real interactions lets researchers build an empirical taxonomy of values a model expresses in dialogue. Key result: Claude tends to adapt the values it shows to the user's context, mirroring preferences most of the time, but it can resist when it detects that fundamental principles are at risk. Imagine a user asking for help to harm someone — the model may mirror casual preferences but push back on requests that violate core safety principles. This is an example of sociotechnical behavior that you need to measure with real data, not only in the lab.
Collective Constitutional AI: participatory design of rules for models
Anthropic, together with the Collective Intelligence Project, ran a public process with roughly 1,000 Americans to draft a 'constitution' for an AI system. They then trained a model using that document. Technically, this shows it's feasible to incorporate structured public input into training loops; politically, it offers a path to legitimize behavior norms through citizen participation. Want to make a system whose rules people trust? This is a concrete example.
Predictability and Surprise: scaling laws versus unexpected capabilities
One technical point stands out: loss curves and other performance metrics scale predictably with size and data, but emergent capabilities can appear unexpectedly. That tension matters for policy: relying only on loss curves to anticipate risks is insufficient. You need to complement them with empirical tests targeted at non-linear behavior and adversarial scenarios — otherwise, surprises will catch you off guard, like a software update that suddenly enables an unforeseen behavior.
Systems and methodologies: Clio, steering and applied evaluations
Clio is described as a system for gaining insights while preserving privacy about real-world AI use. Typical techniques in this space include secure aggregation, differential privacy, or analysis protocols that avoid exposing individual data. The goal is to understand real usage without sacrificing confidentiality.
Evaluating and mitigating bias via feature steering appears as a technical case study. Feature steering are methods to adjust model behavior without retraining from scratch — for example through instructions, prompts, or extra control layers. These methods are tested for effectiveness and side effects.
There are also applied studies on electoral risks and on measuring model persuasion, which require designing controlled experiments and metrics that separate intent, effectiveness, and social vulnerability.
What these findings mean for developers, companies and regulators
For developers: agents can dramatically increase automation of workflows. Design for human oversight, traceability, and metrics that measure not just technical performance but alignment with values. If you're shipping features that rely on agents, make it easy for humans to intervene and audit decisions.
For companies: adoption varies by sector. Startups can gain quick advantages with agents, but large organizations need governance strategies, production testing, and social-impact metrics before wide deployment. Think beyond performance gains — ask how a feature changes people’s day-to-day work and responsibilities.
For regulators and policymakers: the uncertainty around capabilities calls for flexible regulation and standardized evaluation tools. Public participation, like in Collective Constitutional AI, offers a model to legitimize usage norms.
Technical and methodological recommendations
Combine scaling and adversarial tests: use loss curves as diagnostics, but complement them with tests designed for emergent capabilities.
Measure values in real data: large empirical taxonomies (for example, 700,000 interactions) reveal patterns that synthetic tests miss.
Privacy by design: systems like Clio show the need to gain operational insights while preserving privacy, using aggregation, anonymization, and formal privacy techniques when possible.
Public participation in technical governance: incorporating citizen input in design and training phases can improve legitimacy and acceptance.
The Societal Impacts research puts a technical, applied approach on the table to understand how AI interacts with human values and social systems. It’s not just theory: these are experiments, measurements, and processes you can review and apply if you work in product, policy, or research.
