Shadow AI in labs: Why 77% use public generative AI tools at the bench

The rise of the AI Lab Notebook blog series (Part 2)

This post is part of the AI lab notebook (AILN) series. In this blog, we define shadow AI in labs and explain why public generative AI tools have become a default workaround at the bench.

You can also read about experiment reuse and ELN data findability, including why labs repeat experiments when prior work is hard to reuse and the lab AI maturity model, a roadmap from passive to active labs.

Get the full research report here.

Shadow AI in labs refers to scientists using public generative AI tools outside governed workflows, often through unmanaged accounts, to get work done faster. An AI lab notebook (AILN) is a third-generation ELN that brings science-aware assistance into governed workflows so reasoning can stay with the experimental record.

The new tab problem and the rise of shadow AI

When the notebook does not make it easy to get to the right context quickly, scientists do what most people do. They open another tool.

Increasingly, that tool is a public generative AI service like ChatGPT, Claude, or Gemini. These tools can help draft protocols, summarize results, restate complex outcomes, and structure next steps in plain language. The appeal is obvious. They are fast, conversational, and available.

The problem is not that scientists want to bypass governance. The problem is that governed workflows are not always where the work of interpretation and planning is easiest to do.

What the research shows

AI use is now close to universal in the lab.

• 97% of professionals use some form of AI to support their work.
• 77% use public generative AI tools alongside their ELN.

At that level of usage, the key question becomes where the work is happening. The research suggests governance has not kept pace. Forty-five percent of those using public AI do so through personal accounts, while only 32% use company-managed logins. That gap matters because it shifts interpretation and decision logic into places without reliable audit trails, retention controls, or consistent oversight.

Helpful is not the same as fit for purpose for scientific workflows

Public AI can be helpful while still being the wrong place for scientists to attempt serious scientific reasoning.

The central limitation is context. A generic model does not know your assay conditions, instrument behavior, validated methods, or which dataset version is authoritative. That gap shows up in the satisfaction results.

• Only 27% say current AI tools meet their scientific needs “very well.”
• 15% say they are poorly suited to scientific workflows.

There is also a long-term risk that can be easy to miss. When interpretation happens inside tools and only final outcomes are pasted back into the notebook, the record becomes thinner over time. The next scientist inherits conclusions without the reasoning that produced them. That makes reuse harder, increases repeat work, and weakens organizational learning.

From crackdowns to governed capability

Blanket bans can reduce visible use, but they rarely remove demand. If no practical alternative exists, behavior tends to continue in less observable ways.

A better approach is to bring high-value use cases into governed workflows and make trust reviewable.

• Start with real use cases. Map what scientists are doing today: summarizing results, drafting methods, interpreting anomalies, comparing runs, and planning experiments.
• Move AI into governed tools. The goal is not “approved chatbot access.” It is context-aware assistance inside the electronic lab notebook (ELN) workflow, tied to the experimental record.
• Make trust reviewable. Eighty-one percent say they would only trust AI suggestions if they could review the underlying evidence. That requires provenance links back to source data, logging of prompts and outputs with role-based access, and explicit human review where regulated records are affected.

This is where third-generation ELNs, including science-aware AI notebooks like Sapio ELaiN, are designed to help. The intent is to keep AI-assisted interpretation and documentation inside the governed environment, rather than split across tabs and unmanaged accounts.

In the final post, we pull this together into a lab AI maturity model and a practical roadmap for moving from passive documentation and shadow workarounds to governed AI inside lab workflows.

Download the full research report here.

In summary

• Shadow AI is a workflow signal, not just a policy issue.
• The biggest risk is not only exposure; it is reasoning happening outside the experimental record.
• The durable fix is governed, context-aware assistance inside the notebook workflow, not a crackdown.