Anthropic Trains the First Claude Model and Prioritizes Safety Research in 2022

Anthropic trained the first version of Claude in the spring of 2022 but chose not to immediately release the model publicly, instead using it primarily for AI safety research. The decision reflected the company's approach to developing increasingly capable artificial intelligence systems while studying their potential risks, reliability, and alignment. Anthropic later explained that it had deliberately prioritized safety research with its first Claude model and began deploying Claude only after the gap between its capabilities and the publicly available state of the art had narrowed. The timing is important because Anthropic was already operating large-scale experimental infrastructure by this period. On 29/04/2022, the company announced a $580 million Series B financing intended to support research into the safety properties of computationally intensive AI models. Anthropic said its work since its founding had focused on making AI systems more steerable, robust, and interpretable. Its researchers had also developed techniques for making large language models more "helpful and harmless" and were investigating reinforcement learning and human-preference methods for improving model behavior. Anthropic's decision to keep its first Claude model primarily inside its research program was also consistent with its stated position on capabilities research. The company has said that it generally avoids publishing research whose principal effect would be accelerating AI capabilities. Anthropic distinguished this work from alignment research, which studies how AI systems can be made safer, and from safety-related capabilities research designed to identify and reduce risks. In its account of Claude's development, Anthropic specifically cited the spring 2022 model as an example of choosing safety research over immediate public deployment. During 2022, this research program also produced Constitutional AI, an alignment technique that Anthropic publicly described in a research paper released on 15/12/2022. The method sought to train a harmless AI assistant with substantially less direct human labeling of harmful outputs. Instead, human oversight was expressed through a written collection of rules or principles, which Anthropic called a "constitution." Constitutional AI involved two main stages. In the supervised-learning phase, an initial model generated responses and then critiqued and revised those responses according to constitutional principles. In the reinforcement-learning phase, AI-generated feedback was used to compare responses and create preference data, which then informed further training. Anthropic referred to this second process as Reinforcement Learning from AI Feedback, or RLAIF. The researchers reported that the method could produce an assistant that responded more safely to harmful requests without simply refusing to engage with them. The approach was intended partly to address limitations of relying exclusively on large volumes of human feedback. Anthropic later explained that human reviewers can be required to examine disturbing material, that human supervision becomes difficult to scale as models become more capable, and that the values guiding model behavior can remain implicit when they are learned mainly through preference labels. Constitutional AI made at least some of those principles explicit and allowed models themselves to participate in evaluating and revising outputs. Anthropic subsequently incorporated Constitutional AI into Claude. The company described Claude as an assistant trained using the technique and explained that its constitution was intended to guide the model toward behavior that was helpful, honest, and harmless. The principles used for the publicly released Claude were updated from those used in the original 2022 Constitutional AI experiments. Why This Moment Matters: Anthropic's handling of its first Claude model provides a concrete example of the company's early safety strategy: it possessed a model suitable for deployment in 2022 but initially directed it toward internal safety research instead. The same period produced Constitutional AI, which became a defining component of Claude's subsequent training and provided an explicit framework for studying how written principles and AI-generated feedback could influence model behavior.