Where does AI beat humans in research, and where does the work still need human judgment that's hard to check?
Do humans or AI perform better at different research stages?
This asks how research work splits between people and AI, stage by stage: idea generation, experiments, analysis, writing, review. Where is each side stronger, and why does the line fall where it does?
This asks how research work splits between people and AI, stage by stage. The short answer from the corpus is that the split isn't random. It follows one fairly stable rule: AI is strong wherever something outside the AI can check its output, and weak wherever the work depends on judgment that nothing can easily check. Literature retrieval, coding, drafting and analysis fall on the AI side. Choosing which question is worth asking, and judging whether a result matters, fall on the human side. One synthesis describes this as a sharp boundary that depends on the research stage, and finds that it holds even as the specific tasks on each side shift (Where does AI assistance become unreliable in research?).
Real research venues show the same pattern. At Agents4Science, a conference built for AI-authored work, the accepted papers had more human involvement than the rejected ones. The humans clustered in hypothesis and design work, and the AI ran more independently on analysis and writing (Do accepted papers need more human guidance than rejected ones?). Fully autonomous systems can complete the whole loop: The AI Scientist went from idea to a manuscript that passed a first round of workshop review (Can one AI system complete a full research cycle end-to-end?). But a workshop with a 70% acceptance rate is a low bar, and the hardest capability on the list for autonomous science is self-correction, the ability to notice your own mistakes (What capabilities do AI systems need for autonomous science?).
The surprise is in the idea stage. You might expect AI to be a great brainstorming partner, and on narrow metrics it improves: Co-Scientist's hypotheses get better ratings the more compute the system spends debating and refining them (Does more thinking time improve AI-generated research hypotheses?). But across 219,655 ideas from five agent frameworks, AI ideas cluster more tightly than human papers do and stay close to the papers they started from. Even multi-agent designs don't widen the range (Do AI research agents explore as broadly as human researchers?). The same narrowing appears at the scale of a whole field. Scientists who use AI publish three times as many papers and get almost five times as many citations, yet science as a whole covers fewer topics and has less collaboration, because work drifts toward problems that already have plenty of data (Does AI help individual scientists while narrowing scientific focus?). So AI makes each researcher more productive while making the field explore less.
The deeper bottleneck is checking the work. AI can produce plausible research faster than anyone can confirm it's right. Most failures in agent-run research come from made-up content and failed retrieval, not from misunderstanding, and the gap is widest exactly where novelty matters (Can AI verify research outputs as fast as it generates them?). This is why claims that automating AI research will compress years of progress into months rest on shaky ground: they assume research can be checked at scale, and that skills learned on small tasks carry over to important ones (Could automated AI research compress years of progress into months?). Some systems are pushing into roles that were human, such as building up lessons across experiments and adding domain knowledge as they go (Can AI research itself without losing human oversight?). Others propose a different arrangement. They argue humans and AI improving together is both faster and safer than autonomy, because human intuition covers exactly the gap between generating results and verifying them (Can human-AI research teams improve faster than autonomous AI systems?).
The takeaway: the useful question is less "which stage?" than "can anything check this?" When a stage gains a reliable checker, such as a test suite, a benchmark, or an experiment you can rerun, AI moves into it. Where none exists, humans still decide what counts as an interesting question and what counts as a real result.
Sources 11 notes
AI excels at structured, externally verifiable tasks like literature retrieval and drafting, but fails sharply on novel ideas and scientific judgment. The boundary consistently tracks whether an external oracle can verify the output—a principle that remains stable even as specific task assignments shift.
At Agents4Science, organizers observed that accepted papers carried more human input than rejected ones, with humans concentrated in design and hypothesis work while AI gained autonomy in analysis and writing. The pattern emerged from self-reported disclosure tiers across four research stages.
The AI Scientist performed ideation, coding, experiments, writing, and self-review autonomously, producing a manuscript that passed the first round at a machine learning workshop with 70% acceptance rate. Five ensemble reviewers and an area-chair model judged the output against NeurIPS guidelines.
The Virtuous Machines framework identifies hypothesis generation, experimental design, data analysis, and iterative self-correction as essential for autonomous scientific research, none of which standard LLM benchmarks reliably evaluate. Self-correction poses the deepest challenge due to documented degradation in reasoning accuracy.
Co-Scientist's builders report that Elo ratings of AI-generated hypotheses improve as the system spends more compute in a tournament-based debate loop. Three biomedical cases show independent recapitulation and testable predictions, though replication is limited to the builders' own validation.
Show all 11 sources
Across 219,655 ideas from five agent frameworks, AI-generated concepts cluster 7.5% more tightly than human papers and stay 21% closer to their seed literature. Even multi-agent designs fail to widen the exploration range.
AI-augmented researchers publish 3× more papers and receive 4.8× more citations, but collective science shrinks topic coverage by 4.63% and researcher collaboration by 22%. AI concentrates work on data-rich problems rather than exploring new questions.
AI can produce plausible research outputs faster than it can prove them correct or meaningful, shifting the bottleneck from authorship to verification. Evidence shows 39% of agentic research failures stem from content fabrication and 32% from retrieval failures, not comprehension—and the gap widens precisely where novelty and scientific judgment matter most.
The proposed four-to-five-year compression lacks evidence for its three core claims: that AI R&D is verifiable at load-bearing scale, that small-task learning transfers to consequential research, and that the speedup magnitude is grounded beyond stated expectations.
ASI-Evolve demonstrates that AI systems can systematically accumulate experimental insights and inject domain priors—functions humans typically provide—across data, architecture, and algorithm discovery, achieving results like 105 SOTA designs and +3.96 MMLU gains.
Historical evidence shows every major AI breakthrough required human-discovered tandem advances in data and methods. Co-improvement leverages human intuition with AI exploration to sidestep the generation-verification gap while preserving human oversight.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- AI for Auto-Research: Roadmap & User Guide
- What Does It Take to Be a Good AI Research Agent? Studying the Role of Ideation Diversity
- Evaluating Sakana's AI Scientist: Bold Claims, Mixed Results, and a Promising Future?
- PostTrainBench: Can LLM Agents Automate LLM Post-Training?
- Autonomous Research Agents: A Survey of AI Scientists and the Verification Gap
- The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search
- Predicting Empirical AI Research Outcomes with Language Models
- AI Research Agents Narrow Scientific Exploration