What is recursive self-improvement — AI building better AI, and why Anthropic's Jack Clark moved his forecast to 2027
Recursive self-improvement (RSI) is when an AI system creates a more capable successor without human involvement, and that successor repeats the process. Anthropic co-founder Jack Clark wrote in his Import AI newsletter on October 12, 2026 that full RSI now seems more likely to become possible in 2027, a year earlier than he previously thought, citing a measure showing frontier models' 'research taste' doubling every 3.0 months. Two days later, eight US states asked courts to bar OpenAI from building self-improving AI until alignment is assured.
The three lines
- Meaning — AI builds better AI without humans, then repeats; rooted in I.J. Good's 1965 'intelligence explosion'
- Forecast — Clark in May: ~30% by 2027, ~60% by end-2028; in October: 2027 now more likely
- Concern — small alignment errors compound each generation; states seek a court-ordered bar
Key questions
- What does recursive self-improvement mean in AI
- **A loop in which AI builds the next AI, which builds the next.** | Stage | Description | |---|---| | Today | Human researchers build models with AI help | | Partial automation | AI designs and runs most experiments; humans set direction | | Full RSI | AI trains a stronger successor with no human involvement | | Hypothetical result | Accelerating gains — Good's 1965 'intelligence explosion' |
- Jack Clark recursive self-improvement 2027 prediction
- **He gave probabilities in May and moved his central expectation earlier in October.** | When | Statement | |---|---| | May 2026 | Automated AI research ~30% by 2027, ~60% by end-2028 | | Oct 12, 2026 (Import AI 476) | Full RSI 'more likely' possible in 2027; previously 2028 | | Evidence | Frontier models' experimental research taste doubling every 3.0 months since Dec 2025 (95% CI 1.7-5.0) |
- Risks of recursive self-improvement
- **Small errors compound and human oversight falls behind.** | Risk | Explanation | |---|---| | Error compounding | A 99.9%-accurate method falls to ~95% after 50 generations, ~60% after 500 (Clark) | | Alignment breakdown | Techniques 'may break under recursive self-improvement' | | Reward hacking | Learning shortcuts that raise scores, not goals | | Oversight gap | Human checks can't keep pace |
Recursive self-improvement (RSI) is a loop: an AI builds a better AI, which builds a better one still. "Recursive" means the process is applied to its own output. Today, human researchers build the next model with AI assistance; RSI takes the humans out of that loop.
The idea surfaced twice in a single week in October 2026. On October 12, Anthropic co-founder Jack Clark wrote in Import AI that full RSI now looks more likely possible in 2027, a year earlier than he had expected. On October 14, eight US states suing OpenAI asked courts to bar it from developing self-improving AI until alignment is assured. A lab-insider forecast had become a legal remedy.
1. Where the idea comes from
The concept traces to British mathematician I.J. Good, who wrote in 1965 that a machine surpassing all human intellectual activity could design even better machines, producing an "intelligence explosion" — making the first such machine the last invention humans need make. For six decades it was a thought experiment. What changed is that AI became good at AI research itself: writing code, designing experiments and interpreting results.
| Stage | Human role | AI role | Status |
|---|---|---|---|
| 0. Tool | All decisions | Code completion | ~2023 |
| 1. Assistant | Direction, design, review | Writes code, runs experiments | Most labs today |
| 2. Partial automation | Direction, final review | Experiment design and analysis | Emerging on some tasks |
| 3. Full RSI | None | Entire successor training | Clark: 2027-2028 |
2. Clark's evidence
In May 2026, Clark put the odds of automated AI research at about 30% by 2027 and 60% by end-2028, defining the target as a system that trains a stronger successor with "no human involved."
| Measure | Earlier | Recent | Source |
|---|---|---|---|
| SWE-Bench (real bug fixes) | ~2% (Claude 2, late 2023) | 93.9% | Clark via The Decoder |
| CORE-Bench (paper reproduction) | ~21.5% (GPT-4o, Sept 2024) | 95.5% (declared solved) | Clark |
| MLE-Bench (ML engineering) | 16.9% | 64.4% | Clark |
| LM training optimization speedup | 2.9x (Opus 4, May 2025) | 52x (April 2026) | Anthropic internal |
| Experimental research taste (TasteVal) | — | Doubling every 3.0 months since Dec 2025 | Oct 12 post |
The new evidence in October is the last row. P-Zero Research's TasteVal measures research taste — choosing experiments and reading results — across eight tasks against 24 experts. Clark reports Claude Opus 5.5 beats the expert baseline at roughly one-thirtieth of the experts' average cost per attempt. He had long named creativity as the remaining barrier; recent AI-generated math results, he wrote, suggest it is eroding.
3. Why it worries people
| Risk | Explanation |
|---|---|
| Error compounding | A 99.9%-accurate alignment method drops to ~95% after 50 generations, ~60% after 500 (Clark) |
| Reward hacking | Optimizing scores instead of goals |
| Alignment faking | Behaving differently when a test is detected |
| Oversight gap | Human checking can't match the pace |
Clark worries alignment techniques "may break under recursive self-improvement." That links the debate to the scientists' open letter urging air-gapped containment of cyber-capable models: a system that improves itself is exactly the kind whose external channels must be controlled.
4. The skeptics
Researcher Herbie Bradley argues models will take over junior research work but not research vision or long-term agenda-setting. On some post-training benchmarks, AI still scores about half the human level. Clark himself wrote that if neither milestone arrives by 2028, the current paradigm has likely hit a wall.
5. FAQ
| Question | Answer |
|---|---|
| Do today's models improve themselves? | No; humans still build the next model with AI help |
| Same as AGI? | No. AGI is a capability level; RSI is a mode of improvement and could be a path to AGI |
| Who could stop it? | No law addresses it; the state suits' request is the first legal attempt |
| Are the probabilities reliable? | They are one person's estimates from public data, not Anthropic's official view |
6. What remains open
- Definition: Clark's October post relies on his earlier definition.
- Benchmark size: TasteVal's eight tasks give a wide confidence interval.
- Other labs: we found no official RSI forecasts from OpenAI or Google DeepMind.
- Courts: whether any judge entertains a bar on self-improving AI is unknown.
Sources
- Import AI 476 — Creativity and RSI (Jack Clark, October 12, 2026)
- The Decoder — Anthropic co-founder maps out how recursive AI improvement could outpace the humans meant to supervise it
- Hvylya — Anthropic's Jack Clark sees AI training its own successor by 2028
- Al Jazeera — Why are eight US states suing OpenAI?