At MIT’s Engineering Quantum Systems Group, researchers are exploring this. Graduate student Beatriz Yankelevich connected GPT-5.6 Sol, through Codex, directly to laboratory software used to control superconducting qubits. The goal: find out whether AI agents could take over routine parts of the experimental workflow.
The results show a new role for AI in quantum research.
From assistant to active lab operator
Quantum experiments involve far more than running one test and reading the result.
Superconducting qubits need to be carefully characterised and calibrated through sequences of measurements. Each measurement can influence what should happen next, while qubit behaviour can drift over time.
In the MIT experiments, the AI system could:
- choose measurement parameters;
- operate the experimental hardware through lab software;
- analyse the resulting data;
- refine measurements based on the results;
- and carry information forward into the next experiment.
On clear experimental signals, the system was able to complete standard calibration sequences with little human intervention, including identifying qubit transition frequencies and calibrating control and readout pulses.
Why automate quantum experiments?
The main advantage is removing repetitive work from the research process.
Preparing and characterising quantum chips can require hundreds or thousands of preliminary measurements. A single chip can take researchers several days to characterise. MIT’s group now uses AI agents for some routine measurements, allowing experiments to continue for hours while researchers focus elsewhere.
That gives scientists more time for human expertise in interpreting results, designing new experiments, developing theories and deciding what questions to investigate next.
Humans are still essential
The experiments also showed the current limitations. When signals became weak, noisy or ambiguous, GPT-5.6 Sol sometimes struggled to identify suitable measurement parameters and required guidance from an experienced researcher.
So this is not yet an autonomous AI scientist. Instead, it points towards a model in which researchers and AI systems divide the workload: agents handle well-defined and repetitive experimental processes, while scientists remain responsible for judgment, interpretation and scientific direction.
Want to dive deeper?
OpenAI’s case study explores how the MIT team connected Codex to its laboratory systems, how the AI performed qubit calibration and where human expertise is still needed.
