“The AI did it” is not an acceptable scientific record. An agent can search, summarize, draft code, propose an analysis, and prepare an experiment. None of those actions establishes that the work was authorized, executed as described, scientifically correct, or suitable for release.
Four different claims
- ProposedAn agent or person suggested a procedure, interpretation, or next action.
- ExecutedA recorded tool or operator performed a defined procedure against identified inputs.
- ValidatedAn accountable reviewer inspected the relevant evidence and limitations.
- ReleasedAn authorized owner approved the result for a defined audience and use.
These states must not collapse into one another. Fluent language can make a proposal look complete. A successful command can make execution look scientifically meaningful. A reviewer’s recommendation can be mistaken for the data owner’s permission to publish.
Deterministic tools provide a boundary
Where possible, the agent should prepare a concrete operation for a deterministic tool, scientific package, workflow engine, database, or instrument interface. The record should identify the operation, version, inputs, parameters, environment, outputs, and exceptions.
Deterministic does not mean correct. It means the executed operation can be identified and inspected. The scientific question may still be poorly formed. The input may be biased. The method may be inappropriate. The output may require domain interpretation.
Agents propose. Permitted tools execute. Scientists validate. Owners authorize consequential release.
Authority should be narrow and visible
An agent’s authority should be expressed in terms a researcher can understand: which sources it may read, which tools it may call, what information may leave the environment, how much it may spend, when it must stop, and which actions require approval.
A broad role such as “research agent” is not enough. Authority must remain tied to a purpose and expire when the task ends. A capability to invoke a tool is not permission to use it against every dataset.
Exceptions are evidence
Substituted software, unavailable inputs, truncated context, retries, manual interventions, and failed checks belong in the result record. Hiding them produces a cleaner story and a weaker scientific artifact.
The network should make exceptions visible before interpretation. A scientist can then decide whether the deviation is acceptable, requires a rerun, changes the conclusion, or prevents release.
What humans retain
Humans set purpose, classify data, approve consequential work, interpret uncertainty, resolve exceptions, and accept responsibility for release. Independent reviewers can challenge the science or controls without automatically receiving ownership or publication authority.
This is not a claim that people are infallible. It is a clear allocation of accountability. Automation is most useful when it makes evidence and decisions easier to inspect rather than making responsibility harder to locate.
Editorial basis: the public LRN distinction between proposal, execution, validation, and release. Specific scientific workflows require domain-specific acceptance criteria.