TL;DR
Socio-technical alignment could give AI policy a more concrete way to govern frontier AI. Current governance already shapes AI development through risk management, testing, and accountability. The harder challenge is the layer where model behavior is actually specified, evaluated, and adjusted before release. Constitutional AI, Model Spec, Sparrow, and Collective Constitutional AI show that values can enter model development through principles, feedback, reward models, and evaluation criteria. This makes model behavior a public policy issue. If democratic institutions do not help shape this layer, it may be shaped mainly by private company incentives and geopolitical competition.
I believe Socio-technical approaches to AI alignment can offer a different way of AI policy making.
Let me explain!
The starting point to this thinking relates to two issues: the pace of AI capabilities and the challenges faced by policymakers. On one hand, AI systems are developing rapidly, with capabilities advancing exponentially. On the other hand, policymakers are being asked to respond to these changes within institutional frameworks that are often slower and less adaptable. This mismatch has created repeated calls for innovation in how policy is done, because the normal tools of regulation often arrive after the technology has already moved forward. The Fable 5 case shows this reactive pattern. I believe the Fable 5 case It shows the gap between model development and public governance because the government response came after release, when the available tools were suspension, safeguard review, and redeployment. The US government issued an export-control directive after release and required Anthropic to suspend access to Fable 5 and Mythos 5 for foreign nationals. Anthropic disabled access for all customers to comply with the order. The reported concern was a method for bypassing safeguards. Access later returned after new safeguards were implemented and the restrictions were lifted. The sequence matters because policy acted after release. A stronger policy agenda would ask where governance can act earlier in the AI lifecycle, especially around the processes that shape model behaviour before deployment.
But before expanding on how to make AI policy agendas more proactive, the pace argument that policymaking is simply too slow to keep up with rapidly evolving AI systems deserves more scrutiny, since it shapes how the policy problem itself is understood. As Alondra Nelson argues, the narrative of inevitable regulatory lag is often overstated and can serve as a strategic framing that benefits industry actors. Rather than treating delay as unavoidable, her perspective emphasizes the need for policy innovation, for example through approaches like the NIST AI Risk Management Framework 1.0, which introduces versioning into government standard-setting. By designing frameworks that can be iteratively updated, such as moving from a “1.0” to a “2.0,” governance can evolve alongside technological change rather than remaining fixed, showing how existing governance capacity can be adapted, recombined, and extended to respond more effectively to emerging technologies.
With this framing in mind, the smarter question to ask is how policy can become more innovative and technically specific without turning policymakers into model developers?
Much of current AI governance already gives a partial answer. It works through rules, standards, principles, and legal obligations. It shapes how safety, accountability, liability, ethics, and risk are defined. Frameworks such as the NIST AI Risk Management Framework, the OECD AI Principles, and the EU AI Act matter because they articulate what responsible AI should look like. They also influence AI system development practices. Governance already reaches both upstream and downstream parts of the AI lifecycle. It feeds back into design choices through assurance requirements, risk management, testing, and accountability.
The gap is that this engagement often remains indirect. It has less influence over the mechanisms that produce model behavior. These mechanisms are harder for policy to reach because they sit closer to technical development. High-level guidance is not enough. A stronger policy agenda would focus on how model behavior is specified, evaluated, and adjusted before release.
A more socio-technical approaches to AI alignment help reframe how AI policymaking can engage with this layer. Work such as Anthropic’s Constitutional AI and OpenAI’s Model Spec make the technical possibility of this agenda easier to see. Constitutional AI uses a written constitution to guide model behavior. The Model Spec uses behavioral rules and instruction hierarchy to define how the model should respond. Both approaches show that model behavior can1 be specified and evaluated inside development before release.
They also show that this layer is already being governed, largely through private technical choices. A constitution, a reward model, a feedback rubric, or a model specification may look like a technical artifact. Each one carries judgments about authority, values, and acceptable behavior. Each one shapes what the system should do and what kind of relationship it should have with users.
This matters because helpfulness is too narrow as a public standard. A more helpful model may be better as a commercial product. It may still fail to support human judgment, democratic values, or public accountability. For systems expected to shape education, work, politics, and knowledge production, the question cannot stop at whether the model is only useful and hopefully unharmful to the user. It must also ask what kinds of behavior are being reinforced and who has the authority to decide.
Socio-technical alignment focuses on how values and policy considerations are translated into the mechanisms that produce model behavior. Policymakers do not need to design models directly. Public governance can still shape the standards, evidence, and accountability processes around the parts of model development where behavior is being steered.
At a high level, the technical point is straightforward. Foundation models go through several stages after pre-training before release. Their behavior is shaped through post-training methods such as reinforcement learning from human feedback and related approaches such as reinforcement learning from AI feedback. These steps influence which answers are rewarded, which answers are discouraged, and how the model handles user requests, safety concerns, and value conflicts. Deepmind’s Sparrow is one of the earliest expermints that shows one version of the process. The researchers used dialogue rules, human ratings, reward models, and evidence checks to steer a dialogue agent toward behavior judged as more helpful, correct, and harmless. Collective Constitutional AI shows another version. Public input was used to generate principles for model behavior, and those principles were then used in training and evaluation.
The important point for policy is that values can enter model development and we have potential technical behaviour-shaping processes that we can build on. That makes the model-behaviour layer a legitimate object of policy. Policy can therefore focus on:
- How these inputs are formalized into training objectives and constraints.
- How they shape model outputs in practice.
- How developers demonstrate that resulting behaviors align with stated goals before deployment.
Policy should care about this granular layer because model behavior is becoming a form of concentrated power. A small number of firms can set defaults for what models treat as helpful. They can set defaults for what models treat as safe. They can shape what models present as authoritative. These choices affect how people search, learn, work, and make decisions.
Commercial incentives push models toward behavior that feels useful and frictionless. Public governance needs wider standard. It needs to protect human judgment. It needs to protect democratic accountability, and preserve the ability to challenge the system.
The same concern appears in the Sino-American AI race. Model behavior is becoming a site of strategic competition over values and acceptable outputs. Widely deployed systems can carry the value frameworks of the actors that build and govern them. The China’s approach reflects this dynamic more intentionally, with regulatory frameworks that require alignment with state-defined norm. Democratic systems need to put effort in building public mechanisms for defining and testing which values should be encoded. Otherwise, the model-behavior layer may be shaped mainly by private incentives and geopolitical competition.
Empirical results reported in the literature suggest that these approaches can achieve measurable improvements in alignment-related benchmarks, such as reductions in harmful or policy-violating outputs and increases in adherence to specified guidelines. However, reported gains vary by task and evaluation method, and performance remains imperfect, with ongoing challenges in generalization, robustness, and consistency across contexts.