Jack Clark’s 2028 AI Prediction: Why AI Building Its Own Successor Matters

In plain English: Clark is warning that AI may move from helping humans build better AI to taking over enough of the AI-development process that the next model can be produced with little or no human participation ^[7].

What “AI building its own successor” means

A “successor” is the next, more capable system in a model-development sequence. Clark’s concern is that a sufficiently capable AI system could participate in, or automate, the critical work needed to create that next system ^[7].

This is different from today’s familiar idea of AI as a coding assistant. Coding help is one task inside a broader research process. Clark’s scenario is about automating the research-and-development pipeline itself: the process by which frontier models are designed, trained, evaluated, and improved ^[4][5][7].

The important nuance is that Clark is not claiming this outcome is already here. He is making a probabilistic forecast about what could happen by the end of 2028 ^[7].

Why this is called recursive self-improvement

The scenario is often described as recursive self-improvement: an AI system helps create a more capable AI system, which can then help create an even more capable one ^[5].

The high-risk version of that loop is not just “AI improves software.” It is a compounding process in which the system that improves AI becomes more capable each generation. Reports discussing Clark’s warning connect this possibility to an “intelligence explosion,” where AI capabilities accelerate once AI systems can improve the systems that come after them ^[10].

That is why the prediction is treated as more than a technical milestone. If the bottleneck in AI progress shifts from human researchers to AI systems improving AI systems, the pace of capability growth could become harder for humans to monitor and govern ^[4][10].

Why Clark sees it as a major risk

The central risk is loss of oversight over the AI-development pipeline. In an end-to-end automated R&D process, humans may have fewer meaningful checkpoints before a more capable successor system exists ^[4][7].

Three concerns follow from that:

• Fewer human review points. If AI systems can carry out enough research, training, and improvement work on their own, companies and researchers may have less time to evaluate safety before the next system is created ^[4][7].
• Faster capability acceleration. Clark’s warning is linked to the possibility that the speed of AI progress increases as AI begins improving AI, a dynamic described in coverage as an “intelligence explosion” risk ^[10].
• Governance pressure. Clark wrote that the implications are large and that he is not sure society is ready for the changes implied by automated AI R&D ^[7].

In other words, the risk is not a science-fiction image of a robot building another robot. The risk is a faster, more automated frontier-AI production loop that may move beyond the normal pace of safety evaluation, regulation, and public understanding ^[4][7][10].

What remains uncertain

Clark’s 60%+ estimate is a probability judgment, not an established fact ^[7]. The timeline is disputed.

One critique of the forecast argues that human-free, end-to-end recursive self-improvement by 2028 is unlikely, putting the odds below 10%, while still allowing that something like it could become possible over a longer period through 2036 ^[12].

There is also a technical disagreement about whether recursive self-improvement would actually produce accelerating gains. A report citing computer scientist Pedro Domingos notes that the key issue is not merely whether AI systems can generate or modify software, but whether that leads to reliable increasing returns; Domingos argues that this has not been clearly shown ^[6].

These disagreements matter because “AI can assist with AI research,” “AI can automate most of AI R&D,” and “AI can recursively improve itself fast enough to cause an intelligence explosion” are related but not identical claims ^[6][12]. Clark’s warning is about the most consequential version of that progression ^[7].

Bottom line

Clark predicts that by the end of 2028, it is more likely than not that AI systems could perform no-human-involved AI R&D and plausibly build their own successors ^[7].

If he is right, the main danger is not just faster innovation. It is that the creation of more powerful AI systems could begin moving faster than human oversight, safety testing, and governance can adapt ^[4][10]. The strongest caveat is that this remains a contested forecast, with critics arguing that the 2028 timeline may be too aggressive ^[12].