Researchers Let AI Do Science – It Still Can’t Publish

A collaborative effort spanning several universities and AI labs set out to answer a bold question: can today’s most advanced AI agents replace human scientists? The answer, according to a new study published on Decrypt, is a resounding “not yet.”

The researchers gave state‑of‑the‑art language models—augmented with tool‑use capabilities and autonomous planning—full control over the entire research pipeline. The agents drafted hypotheses, designed and executed simulations, and even wrote draft papers. On the mechanics side they performed admirably, automating data collection and model tuning with a speed that would make any lab manager jealous.

However, when the output was evaluated by a panel of top AI conference reviewers, none of the AI‑generated manuscripts met the novelty threshold required for acceptance at venues like NeurIPS or ICML. The agents tended to remix existing ideas, produce incremental improvements, and, crucially, lack the intuitive leaps that human researchers bring to the table. The study’s authors note that the agents “stumbled on the creative axis of scientific discovery,” a gap that current reinforcement‑learning‑from‑human‑feedback pipelines cannot bridge.

For the broader AI ecosystem, the findings are both sobering and instructive. Autonomous agents have already proven valuable in low‑risk, high‑throughput tasks—think DeFi arbitrage bots or on‑chain monitoring scripts. This research underscores that the same agents are not yet ready to shoulder the intellectual heavy lifting of frontier science. Expect a continued hybrid model where AI handles repetitive, data‑intensive chores while humans retain strategic direction and hypothesis generation.

Investors and developers should take note: hype around “AI‑only research labs” is premature. Allocating capital to AI agents that claim to replace entire research teams without a clear path to genuine innovation is a high‑risk bet. Instead, funding should target incremental integration—tools that augment human scientists, improve reproducibility, and accelerate the iterative loop.

The study also raises governance questions. As AI agents become more autonomous in sensitive domains—biotech, finance, or security—regulators will need to define accountability frameworks. Until AI can consistently produce peer‑review‑worthy breakthroughs, human oversight remains the safety net.

In short, the experiment proves that while AI agents can manage the logistics of science, they still lack the spark that turns data into discovery. The frontier remains open for researchers to build the next generation of truly creative machines, but for now, the human mind still holds the pen.