AI is getting so advanced it can now help build the next generation of AI. Prominent lab Anthropic reports its models are dramatically speeding up research, but it's also clear: AI is a powerful tool, not yet an autonomous scientist.
What is AI-Assisted Research?
Imagine a team of
brilliant but tireless research assistants who can write code, run experiments, and analyze data 24/7. That's essentially the role Anthropic's own AI models, like Claude, are beginning to play inside the company. This isn't about an AI independently having a 'eureka' moment. Instead, it's about AI handling the grunt work of research and development at an unprecedented scale. Human researchers still provide the ideas and direction, but they can now test more hypotheses and explore more avenues than ever before. Anthropic reports that its engineers are merging significantly more code than in previous years, not because they are typing faster, but because AI is doing much of the writing, leaving humans to direct and review the work.
The Promise of Rapid Progress
The impact of this new approach is significant. A recent independent assessment by the research organization METR found that Anthropic's rate of progress on AI capabilities has roughly doubled since October 2025. This means the company may now be achieving in one year what would have previously taken two. This acceleration is not primarily due to using more powerful computers, but rather from the efficiency gains of using AI in the R&D process itself. This creates a feedback loop: better models help build even better models, faster. This concept, known as recursive self-improvement, has long been a theoretical milestone in AI, and while Anthropic says we aren't there yet, these developments suggest it could be approaching faster than many expected.
The All-Important Human Check
Despite the startling speed, Anthropic is drawing a firm line in the sand. The company has been clear that its current models are not autonomous and that it does not expect them to be able to design and build future AI generations without 'meaningful human supervision or understanding'. This is a crucial distinction. While an AI can execute tasks, it still lacks the critical judgment and scientific self-criticism that are hallmarks of true research. Studies show that even advanced models can struggle to judge their own work or innovate beyond known techniques. Anthropic's public statements emphasize that human oversight is, and will remain, essential to guide development safely.
Navigating a New Frontier of AI Safety
This acceleration presents both a huge opportunity and a profound challenge. The faster AI capabilities advance, the harder it is for safety measures to keep pace. Anthropic has openly acknowledged this tension in filings, warning that progress could outstrip the ability to validate safeguards. The company is a vocal proponent of responsible scaling and has even established internal tripwires, such as a 2x acceleration in research speed, to trigger enhanced safety reviews. Recent reports suggest its progress is hovering just below that threshold. The situation highlights a central conflict in the AI industry: the race for commercial and technological advantage versus the need for cautious, deliberate development to ensure these powerful systems remain aligned with human values.
