A Claude Model Autonomously Chose a Gene Therapy and Operated on a Mouse Brain

An artificial intelligence agent powered by Anthropic’s Claude Sonnet 5 reportedly selected a gene therapy, identified where it should be injected and controlled a motorized surgical device to deliver it into the brain of a mouse with epilepsy. The experiment, described by Exin Therapeutics co-founder Gabriel Ocana-Santero, represents an unusual step in the use of AI for biological research: the model was given a scientific objective rather than a predetermined treatment or surgical plan.

The experiment does not mean that an AI system independently performed an entire brain operation without human involvement. A human surgeon handled critical parts of the procedure, while a licensed veterinarian approved irreversible steps and held an emergency stop. Nevertheless, the model made several consequential decisions that would traditionally be made by a researcher or specialist. The account also remains preliminary, with no peer-reviewed paper or published evidence yet showing whether the treatment reduced seizures.

Key Takeaways

• Claude Sonnet 5 reportedly selected a gene therapy from a list of 19 candidates.
• The AI chose the brain region and calculated injection coordinates using scientific literature.
• It controlled a motorized surgical rig and monitored the procedure through a camera.
• A human surgeon and licensed veterinarian remained involved in safety-critical steps.
• Claude stopped when it detected a calibration problem that could have made the procedure unsafe.
• The mouse reportedly recovered, but the treatment’s effect on seizures has not been established.

How the AI Selected the Gene Therapy

According to Ocana-Santero, the team provided the AI agent with an epilepsy mouse model, a spreadsheet containing 19 possible gene therapies and one broad objective: reduce the animal’s seizures.

Instead of receiving instructions to administer a specific treatment, Claude was expected to investigate the available options and decide which intervention was most appropriate.

The model examined how seizures develop in the particular mouse model and selected a therapy intended to strengthen inhibitory signalling in the brain. This involved the GABA pathway, which helps regulate neuronal activity and prevents excessive signalling.

Claude also selected the lateral entorhinal cortex as the target region. It reportedly rejected another possible target, the substantia nigra, after assessing scientific literature suggesting that the alternative was less suitable for the seizure type being studied.

This distinction is important. The model was not simply following a human-written surgical checklist. It was being asked to connect biological evidence, a treatment hypothesis and a physical intervention.

Claude Calculated the Injection Coordinates

After selecting the treatment and brain region, the AI agent needed to determine where the injection should take place.

According to the account, Claude consulted three scientific papers, compared their anatomical reference points and calculated injection coordinates by combining the measurements.

It also selected procedural parameters, including injection volume, speed and waiting times.

The system then interacted with a custom motorized surgical setup equipped with cameras. During the procedure, the agent observed the surgical field and made 32 observations, according to Ocana-Santero.

It monitored the injection and reported whether it detected visible bleeding or fluid backflow. The system was therefore involved in both planning and controlling part of the physical experiment.

The reported injection volume was 250 nanolitres, a very small quantity that requires precise positioning when working with brain tissue.

A Human Safety System Remained in Control

Despite the model’s decision-making role, the experiment was not conducted without human safeguards.

A human surgeon opened the skull and later closed the wound. A licensed veterinarian approved every irreversible step and maintained access to a physical emergency-stop control capable of cutting power to the motors.

The surgical setup also included fixed operating limits. The AI agent was not permitted to bypass those restrictions.

These controls matter because an incorrect injection location, excessive volume or unexpected movement could cause serious harm to the animal.

The experiment therefore demonstrates a system in which AI can make treatment and execution decisions within a human-defined safety framework, rather than a completely independent surgical system.

Claude Refused to Continue After Detecting a Technical Error

One of the most revealing moments occurred during the procedure.

A connection problem disrupted the setup, and the system calculated that the target position was outside the surgical device’s physical range of movement. The model refused to continue.

The research team attempted to restart calibration, but the agent continued to reject the movement. According to Ocana-Santero, the problem was eventually traced to manual adjustments made during an earlier pause that had affected the control system’s alignment.

The researchers had to correct the software manually. The AI agent would not proceed even after the team attempted to persuade it to retry.

This incident highlights an important issue for AI-controlled laboratory equipment: a model must not only make decisions but also recognize when the available information or physical conditions are unreliable.

In this case, the refusal prevented the system from attempting a movement that its own calculations indicated was unsafe. However, it also left the researchers managing a time-sensitive procedure with an anesthetized animal.

Why the Experiment Is Different From Earlier Surgical Robots

Robotic systems have already been used in surgery and laboratory research. They can perform highly precise movements, including needle placement and other delicate operations.

The difference in this experiment is the level of decision-making delegated to the AI.

In a conventional robotic procedure, a human specialist generally determines the treatment, target and surgical plan. The robot then helps execute those instructions.

Here, Claude reportedly selected the treatment, chose the brain region, calculated the injection coordinates and controlled the device during the procedure.

That combination of scientific reasoning and physical execution is significant. It moves AI beyond analysing research papers or suggesting possible experiments and toward participating directly in laboratory work.

Anthropic has also been developing technology to connect Claude agents with programmable scientific equipment through its Model Hardware Standard initiative. The broader idea is to allow AI systems to interact with instruments such as microscopes, robotic arms and other laboratory devices.

What We Know About the Mouse’s Recovery

Ocana-Santero said the mouse recovered well after the procedure. However, that does not establish that the selected gene therapy worked.

The central biological question is whether the treatment actually reduced seizures. The publicly available account does not provide sufficient seizure measurements, long-term observations or independent experimental results to establish therapeutic effectiveness.

The report also has not been peer-reviewed. A detailed technical report is expected, but independent researchers will need access to the methodology and results before the findings can be properly evaluated.

For now, the experiment should be understood as an early demonstration of AI-directed laboratory work, not proof that AI can independently develop or deliver successful treatments for neurological conditions.

The Bigger Picture

The experiment points toward a possible change in how scientific research is conducted. AI agents could eventually help researchers move through more of the experimental process: reviewing evidence, developing hypotheses, selecting interventions, operating instruments and analysing results.

This could make some research workflows faster and allow scientists to investigate more possible treatments. It may be particularly useful in areas where experiments require repeated measurements, complex equipment and large amounts of scientific information.

But greater autonomy also introduces new risks. A model can misunderstand biological evidence, select an unsuitable intervention or misinterpret a sensor reading. When an AI system controls physical equipment, those errors can have consequences beyond an incorrect answer on a computer screen.

Human oversight, hardware restrictions, transparent experiment logs and independent validation will remain essential. The experiment also raises questions about responsibility when an AI system makes a decision that affects a living animal.

Looking Ahead

The next important milestone will be the publication of Exin Therapeutics’ technical report. Researchers will need to examine the treatment selection, injection parameters, safety controls and biological outcomes. Evidence showing a measurable reduction in seizures, supported by repeat experiments, would be much more significant than the successful completion of the injection alone.

For Anthropic and the wider AI research industry, the demonstration illustrates how language models may become connected to real-world scientific infrastructure. The technology is still experimental, but the direction is becoming clearer: AI agents are moving from helping scientists think about experiments toward taking part in carrying them out.

Frequently Asked Questions

1. Which Claude model was used in the experiment?

The experiment reportedly used Claude Sonnet 5, developed by Anthropic.

2. Did Claude independently choose the gene therapy?

According to Exin Therapeutics co-founder Gabriel Ocana-Santero, the model selected a treatment from 19 candidate gene therapies based on the objective of reducing seizures in the mouse model.

3. Did Claude perform the entire brain surgery without human involvement?

No. Claude reportedly selected the treatment and controlled the injection process, but a human surgeon opened and closed the skull. A licensed veterinarian approved critical irreversible steps and maintained an emergency stop.

4. Did the gene therapy cure the mouse’s epilepsy?

That has not been established. The mouse reportedly recovered after the procedure, but published evidence demonstrating a reduction in seizures is not yet available.

5. Why did Claude stop during the procedure?

The model detected that the calculated target position was outside the device’s physical range after a calibration problem. It refused to continue until the issue was corrected.

6. Has the experiment been scientifically verified?

The account has not yet been peer-reviewed, and a detailed technical report is expected. Independent validation and repeat experiments will be necessary to assess the results.

7. Could AI perform human brain surgery in the future?

AI could eventually assist with treatment planning and robotic procedures, but this experiment does not demonstrate that AI is ready to perform human surgery. Clinical use would require extensive testing, regulatory approval and strict medical oversight.