Georgia Tech adds a watchdog AI to its electron microscope

When Vida Jamali describes her vision for the electron microscope, she makes it sound almost peaceful. We see this as a step toward scientific instruments that do more than acquire data, she told Georgia Tech's news office. Systems that can reason over experiments, adapt measurements, and participate in the scientific discovery process alongside researchers. A colleague. A collaborator.

What she did not say — what the paper says instead, in the architecture section, between the citations and the deployment diagrams — is that her team also gave the microscope a designated antagonist.

The paper, published this month in npj Computational Materials, introduces a five-agent AI framework for electron microscopy. Four agents do the scientific work: a principal investigator, a materials scientist, an electron microscopist, a physicist. They design experiments, simulate outcomes, run closed-loop iterations, propose hypotheses. The fifth agent is the critic. Its job is to watch the other four and catch the moments when they hallucinate, reason badly, or reach conclusions the data does not support.

That fifth seat is the story.

The framing from the wire — thinking microscopes that collaborate with scientists — is technically accurate and spiritually incomplete. The real message embedded in this paper is that the researchers do not trust the AI alone. Not in a high-stakes domain. Not for scientific discovery. So they built a watchdog into the microscope alongside the scientists, and called it a collaborator.

Standalone large language models, the paper notes, lack access to specialized scientific knowledge, which contributes to hallucinations and superficial reasoning in technical domains. They also suffer from what the authors call context rot: degraded performance as input length grows. These are not minor engineering inconveniences. They are precisely the failure modes you cannot afford when the instrument is deciding whether a defect in a semiconductor is worth pursuing, or whether a protein structure justifies a new research direction. Hallucinations in a chatbot are a joke. Hallucinations in a $2 million microscope running an autonomous experiment are a publication — or worse.

The multi-agent architecture addresses this by distributing expertise across specialized roles. Each agent has a defined scope. Each can check the others. The critic is the structural admission that the system cannot be trusted to get it right without a second set of eyes. This is not how the press release describes it. The Georgia Tech news office talks about collaboration, co-scientists, the future of discovery. The paper talks about failure modes and adversarial oversight. Both are true. The oversight is what matters.

This tension — between the collaborative framing and the defensive architecture — is where the piece lives. The researchers want to position agentic AI as a partner. The system they built assumes the partner will screw up. Those two things are in the same paper.

The production gap makes this more than academic. A January 2026 survey by Camunda of 1,150 large enterprises found that 71 percent were already using AI agents in some form. But only 11 percent of those deployments had actually reached production by the end of last year. The bottleneck was not capability — it was trust. Governance, transparency, compliance. Who is accountable when the AI agent makes a bad call? In a call center, the answer is a supervisor. In a materials science lab, the answer has to be baked into the architecture.

The Georgia Tech team is further along than most. They are connecting cloud-based agentic infrastructure to electron microscopes at the Institute for Matter and Systems. They have a working framework. The question is whether it scales, and whether anyone else can replicate it without building their own version of the critic first.

The paper does not claim the system has solved this. It is a commentary and a vision. The authors are careful to note that human researchers must retain accountability for the accuracy and integrity of both the experimental process and the results reported. That sentence sits in the paper like a caveat and a confession at the same time. The AI can run the experiment. The human still signs the paper. Nobody has explained what happens when the AI was the one who decided which experiment to run in the first place.

That gap — between what the architecture can do and what the accountability model allows — is where the field is going to have to do real work. Jamali and her colleagues have built a microscope that can think. They have not yet built a world in which thinking machines can be trusted without supervision. They know it. The fifth chair at the table is the proof.