AI’s recursive self-improvement might not come so quickly after all

Details

The AI industry's most bullish bet right now is that models will soon automate AI research itself, enabling recursive self-improvement. A new study from a multi-institution group led by Peter Kirgis and Sayash Kapoor at Princeton throws cold water on the near-term version of that claim. The researchers introduced a rigorous new evaluation framework called 'shadow evaluation,' designed to test whether AI agents can do open-ended AI research — the kind with no checkable right answer, requiring hypothesis selection, experimental judgment, and creative decision-making. They chose two unpublished papers submitted to NeurIPS 2026 as test cases, so the model couldn't have memorized answers. Claude Opus 4.8, running on the open-source OpenClaw harness, was given 6 days, ,000 in Anthropic API credits, a GPU budget, and internet access. It was asked to produce research papers worthy of a top conference. The original authors evaluated the results using the same standards they'd apply to conference submissions. Both papers were rejected. The agents were capable of all the engineering involved — running experiments, reviewing literature, processing results. What they lacked was the judgment to make good research choices: which hypotheses to pursue, when to abandon a direction, how to produce original insights. The finding suggests that benchmarks measuring narrow engineering tasks overstate progress toward genuine AI research automation.