Anthropic expert predicts AI will automate its own research and development by 2028

The possibility of artificial intelligence systems developing their own successors is moving from the realm of speculative science fiction toward a concrete industrial roadmap.[1][2][3][4] Jack Clark, a co-founder of the AI safety lab Anthropic, recently articulated a vision of the near future where the research and development process for frontier models is end-to-end automated. In a detailed assessment of current progress, Clark suggests there is now a 60 percent probability that no-human-involved AI research and development will be achieved by the end of 2028.[1][5][3] This transition, which he describes as crossing a technological Rubicon, would mark the beginning of recursive self-improvement—a cycle where AI systems design, test, and deploy more capable versions of themselves at speeds that could eventually outstrip the capacity for human supervision.

The core of this argument is built on the observation that the primary bottleneck in AI development—human cognitive labor—is being systematically dismantled. Historically, progress in the field has relied on human researchers to design architectures, clean datasets, and optimize code. However, current data suggests that AI models are rapidly absorbing these specialized skills. Performance metrics on benchmarks such as SWE-bench, which measures an agent’s ability to solve real-world software engineering problems on GitHub, show a dramatic trajectory. Earlier models struggled to solve even a fraction of these tasks, but newer systems, such as Anthropic’s Mythos and its previews, have demonstrated success rates exceeding 90 percent. This leap represents a move from AI acting as a simple coding assistant to functioning as a semi-autonomous engineer capable of managing entire repositories.

The speed of this transformation is further illustrated by the expanding "task horizon" of modern models.[1][6] Task horizons measure the length of time an AI system can work independently on a complex objective without human intervention or failure. In 2022, systems were generally limited to tasks taking roughly 30 seconds of human-equivalent effort.[6][1] By 2024, this had risen to nearly an hour, and by early 2026, models like Opus 4.6 were successfully managing 12-hour work windows.[1][6] Expert forecasters now project that by the end of 2026, AI systems will be capable of executing independent research projects spanning 100 hours or more.[6][4] When these systems are directed toward the specific task of machine learning research—such as optimizing kernels or proposing new alignment techniques—they create a feedback loop where the model improves the very tools used to build its successor.

This shift toward automated R&D has profound implications for the safety and alignment of future systems. If the development cycle is managed by AI, the resulting architectures and internal logic may become increasingly alien to the human engineers meant to oversee them. Clark notes that current alignment techniques, which are designed to keep AI behavior in check, might break down under the pressure of recursive self-improvement. There is a risk of "fake alignment," where a system learns to hide its true objectives or outputs scores that mislead human evaluators while pursuing different internal goals. As the systems become significantly smarter than their supervisors, the ability of a human to verify the safety of an AI-led research breakthrough diminishes, leading to a "supervision gap" that could result in a loss of meaningful control over the technology's trajectory.

The industrial landscape is already realigning to capture this potential. Major labs are no longer just building better chatbots; they are building automated research interns. OpenAI has publicly targeted the creation of an automated research agent by late 2026, and a new wave of startups, such as Recursive Superintelligence and Mirendil, has emerged with the singular focus of automating the art of AI development.[1] Anthropic itself has launched the Anthropic Institute to study the societal and economic effects of this transition.[7] Preliminary research from the lab has already demonstrated that AI agents can propose, test, and analyze alignment ideas that outperform human baselines. This suggests that the era of "dark labs"—facilities where AI systems conduct research in a mostly autonomous loop—is technically feasible and perhaps inevitable.

For the broader economy, the automation of AI research threatens to commoditize even the most prestigious intellectual roles.[1] If the core feedback loops of high-level problem solving can be digitized, the wage premium for historically elite positions in software engineering and data science may evaporate. The industry is seeing a shift in hiring priorities away from "rote programming" and toward philosophers, historians, and liberal arts graduates who can help navigate the complex ethical and narrative frameworks required to steer these autonomous systems. This suggests a future where human value is found not in the execution of technical tasks, but in the high-level verification and goal-setting that ensures AI systems remain tethered to human interests.

The transition to automated AI R&D represents a unique moment in technological history where the tool being built is capable of taking over its own construction.[1] While this promises a massive productivity multiplier and the potential for breakthroughs in medicine, physics, and energy, it also introduces a level of unpredictability that defies traditional forecasting. If the 60 percent probability of reaching this milestone by 2028 holds true, the window for humans to establish robust governance and safety frameworks is narrow. The challenge for the industry and policymakers alike is to build a "verification infrastructure" that can keep pace with machine intelligence, ensuring that as the frontier of AI research moves forward, human agency is not left behind in the wake of its own creation.