# Study finds AI agents lack creativity needed for recursive self-improvement

Digest AI · Research · published 2026-09-11T17:57:42Z · updated 2026-09-14T16:04:16Z

Canonical: https://digestai.news/story/vinyals-says-ai-selfimprovement-is-inevitable-but-not-an-explosion

## Summary

A multi-institution study led by Princeton University researchers challenges the prevailing narrative that AI recursive self-improvement is imminent. While current Large Language Models excel at narrow engineering tasks, the team found they lack the judgment and creativity required for open-ended scientific research. Using a new "shadow evaluation" method, the researchers tested Anthropic’s Claude Opus 4.8 on unpublished NeurIPS 2026 papers. Despite having access to compute resources and web tools, the agents produced work that was rejected by the original authors for failing to meet top-tier conference standards.

The agents demonstrated strong technical execution, such as running experiments and compiling data, but struggled with hypothesis generation, backtracking from failed approaches, and integrating feedback. They often committed to unpromising methods too quickly and failed to rethink their strategies fundamentally. This gap suggests that while AI can automate the mechanical aspects of research, it cannot yet replicate the intuitive leaps necessary for transformative breakthroughs.

Anthropic co-founder Jack Clark echoed these findings, noting that current systems exhibit "rote, formulaic thinking" that hinders their effectiveness as researchers. The study implies that timelines for fully automated AI development may be overly optimistic, particularly if open-ended creativity remains a bottleneck. Future tests are underway using Anthropic’s more advanced Mythos model to see if newer architectures can bridge this creative divide.

## Key points

- Princeton-led study shows AI agents fail open-ended research tasks despite strong engineering capabilities.
- Claude Opus 4.8 outputs were rejected by NeurIPS authors for lacking novelty and scientific judgment.
- Researchers argue current AI lacks the creative intuition required for true recursive self-improvement.

## Why it matters

This research tempers expectations for rapid AI self-improvement, suggesting that creative judgment remains a human bottleneck. It highlights a critical gap between automated engineering and genuine scientific discovery, potentially slowing the timeline for autonomous AI development cycles.

## Sources

1. [AI recursive self-improvement might not come so quickly after all](https://technologyreview.com/2026/08/18/1142188/ai-recursive-self-improvement) (technologyreview.com, 2026-09-13)
2. [Why you should work on AI for AI Research](https://latent.space/p/recursive) (Latent Space, 2026-09-14)
3. [AI researchers debate how close we are to recursive self-improvement](https://dwarkesh.com/p/john-beren-charlie) (dwarkesh.com, 2026-09-11)
4. [Ex-Deepmind VP Vinyals says AI self-improvement is coming but won't trigger an intelligence explosion](https://the-decoder.com/ex-deepmind-vp-vinyals-says-ai-self-improvement-is-coming-but-wont-trigger-an-intelligence-explosion) (The Decoder, 2026-09-11)

Part of the developing story: [The AI Automation and Self Improvement Debate](https://digestai.news/thread/vinyals-says-ai-selfimprovement-is-inevitable-but-not-an-explosion) (2 stories)

## Cite

Digest AI, "Study finds AI agents lack creativity needed for recursive self-improvement", 11 September 2026, https://digestai.news/story/vinyals-says-ai-selfimprovement-is-inevitable-but-not-an-explosion

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/vinyals-says-ai-selfimprovement-is-inevitable-but-not-an-explosion.json