What Is an AI Scientist?

Working definition

The name should describe authority, not personality

An AI scientist is a research system with delegated scientific authority. That authority can be narrow, such as selecting the next simulation parameter, or broad, such as generating a paper draft after running code experiments. The important question is not whether the model sounds scientific. It is what decisions the system can make, what tools it can invoke, and what evidence remains after each decision.

A useful autonomy ladder has five rungs. The first is a literature assistant that summarizes papers. The second is an analyst that writes code or notebooks under direct human instruction. The third is a planning agent that proposes hypotheses and experiments but waits for approval. The fourth is a closed-loop agent that can run tools, inspect results, and choose the next step inside a bounded sandbox. The fifth is a lab or field system that can interact with instruments, robots, purchases, collaborators, or live biological and chemical workflows.

Claude Scientist focuses on rungs three through five. General literature review belongs to Claude Researcher; general Claude tutorials belong elsewhere. This site asks whether an agent can move evidence through a scientific loop without hiding the steps.

"fully automated open-ended scientific discovery"

The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery, arXiv

Minimum viable evidence

Four gates separate an assistant from an autonomous scientist

The first gate is hypothesis formation. The system must produce a claim that could be wrong, not just a topic list. The second gate is protocol formation. The system must specify the method, variables, controls, stopping rules, and success criteria. The third gate is execution. The system must operate a tool, simulation, notebook, instrument, or robotic workflow in a way that generates new evidence. The fourth gate is audit. The system must preserve prompts, tool calls, data artifacts, code, errors, parameters, and human approvals.

A system may be valuable before it passes all four gates. Claude Science, for example, is best understood as a workbench that can make research work more legible and productive. It becomes part of an autonomous-science system only when connected to bounded tools and governed by explicit review policy.

• Hypothesis: a falsifiable claim or search target.
• Protocol: a method with variables, controls, and stop conditions.
• Execution: tool use, code, simulation, robotic work, or instrument control.
• Audit: traceable artifacts that another scientist can inspect or reproduce.

Misleading labels

Do not confuse autonomy with speed

Fast paper summarization is useful, but it is not autonomous science. A model that drafts a grant section from a human outline is not acting as a scientist. A model that proposes a docking screen, writes the analysis code, runs the screen in a sandbox, detects an error, revises the protocol, and sends a reproducible report for approval is much closer.

The boundary matters because the risk changes with authority. A bad summary wastes time. A bad closed-loop agent can waste compute, contaminate data, damage equipment, or make a misleading discovery claim look more complete than it is.

"illusions of understanding"

Artificial intelligence and illusions of understanding in scientific research, Nature