Anthropic unveils Claude, a GPT-4 rival designed to be safer

Anthropic unveiled Claude on Tuesday, its first powerful conversational assistant. The company, founded by former OpenAI employees, aims to compete in the language-model market with a key difference: the system was trained to be helpful, honest and harmless using a proprietary technique called constitutional AI.

The announcement came the same day OpenAI launched GPT-4. The timing positions Anthropic as one of the most significant alternatives to the company that popularized ChatGPT, although Claude is beginning its rollout on a more limited basis: the company is initially offering it to a small group of businesses through a chat interface and a developer API.

An assistant that learns from a “constitution”

Claude belongs to the family of large language models: systems trained on enormous amounts of text to predict and generate words, answer questions, summarize documents or draft text. But the outcome of that training depends as much on the data as on the rules and corrections the model receives afterward.

Anthropic’s proposal is to incorporate those rules more explicitly. Constitutional AI starts with a set of principles that the model must use to review its own responses. When it receives a problematic instruction, the system generates a critique of its first answer and tries to reformulate it in line with those principles.

It is an alternative to traditional human supervision. With current models, it is common to hire people to compare responses and flag which ones are more useful or safer. That process is known as reinforcement learning from human feedback. Anthropic combines that supervision with feedback generated by another AI, a method it calls reinforcement learning from AI feedback.

The idea does not eliminate human involvement or make the system infallible. Instead, it aims to make the rules applied to the model more consistent and make it possible to examine where they come from. In a chatbot open to the public, the challenge is clear: it must refuse to provide harmful content without responding evasively to perfectly legitimate requests.

Safety as a product and a strategy

Anthropic was founded in 2021 by a group of researchers and executives who had worked at OpenAI, including Dario Amodei and Daniela Amodei. From the outset, the company has placed the safety of advanced systems at the center of its research, a position it is now bringing to a commercial product.

This is not just a technical issue. Chatbots capable of writing fluently are already used for customer service, information search, drafting and programming assistance. They can also make up facts, reproduce biases present in their training materials or follow inappropriate instructions if their safeguards fail.

Claude is therefore arriving at a time when technology companies are competing not only on capabilities, but also on trust. Google announced a $300 million investment in Anthropic in February, while accelerating its own generative AI projects. Microsoft, meanwhile, has strengthened its partnership with OpenAI and integrated the company’s technology into Bing.

Safety could become a commercial advantage if corporate customers prefer models with more predictable responses. But it also means testing whether those barriers hold up across long conversations, ambiguous instructions and unexpected uses. A controlled demonstration is not enough to measure how an assistant behaves once thousands or millions of people begin testing its limits.

The race is no longer just for the most powerful model

Claude’s launch confirms that the market will not be limited to ChatGPT and Google’s products. Companies with the ability to train large models are pursuing different ways to solve the same problem: how to offer highly capable assistants without opening the door to serious errors or easily scalable abuse.

For users, the practical difference will be seen in the quality of responses, the restrictions imposed by each service and the transparency with which it explains its failures. For businesses, the question will be whether Claude can be integrated reliably into real-world products and whether its safety approach offers assurances that other models do not.

Anthropic still has to prove that point beyond its initial customer group. But Claude’s arrival introduces a necessary debate into the race to build conversational assistants: it is not enough for a model to respond well; what also matters is the rules under which it decides to respond and how it reacts when it should not.

SEO Title: Claude: Anthropic launches a safer GPT-4 rival SEO Description: Anthropic unveils Claude, an assistant based on constitutional AI that aims to compete with GPT-4 by prioritizing helpful, safe responses.