Shadow AI

What it means

Shadow AI is AI that employees use at work without their organization's approval or knowledge. IBM puts it this way: "Shadow AI is the unsanctioned use of any artificial intelligence (AI) tool or application by employees or end users without the formal approval or oversight of the information technology (IT) department."¹

Here's an everyday example. A manager has a long report due and pastes the notes into a free public chatbot to get a summary. IBM calls this the common case: "A common example of shadow AI is the unauthorized use of generative AI (gen AI) applications such as OpenAI's ChatGPT."¹ The intent is usually good. IBM says "Employees often turn to these tools to enhance productivity and expedite processes."¹ The trouble is that whatever gets pasted in leaves the organization's control.

It happens a lot. In a Microsoft and LinkedIn survey "of 31,000 people across 31 countries," Microsoft reported that "78% of AI users are bringing their own tools to work — Bring Your Own AI (BYOAI)."²

Where the word comes from

We could not find a reliable source for who first used the phrase "shadow AI" or when, so we don't name one. The term builds on an older one, "shadow IT," which means technology used at work without the IT department's approval. IBM draws the line between them: "While shadow IT focuses on any unauthorized application or service, shadow AI zeros in on AI-specific tools, platforms and use cases."¹ IBM also ties the rise of the problem to generative AI's growth at work: "Alongside this growth came a rise in shadow AI."¹

How others define it

• IBM (explainer, 2024): "Shadow AI is the unsanctioned use of any artificial intelligence (AI) tool or application by employees or end users without the formal approval or oversight of the information technology (IT) department."¹
• IBM and Ponemon Institute (breach research, 2025): "shadow AI (unregulated, unauthorized use of AI)."³
• TDWI (glossary): "Shadow AI refers to artificial intelligence tools and systems adopted and used within an organization without the knowledge," and the sentence goes on to name the people it bypasses: central IT or data leadership.⁴
• Chhetri and colleagues (research preprint, 2026): "the unsanctioned use of frontier AI outside established organisational controls," which the authors go on to describe as "a systemic socio-technical condition of assurance erosion."⁵

Where they agree and differ

All four definitions agree on the core idea: AI in use that the organization has not approved. They disagree on two points. The first is who is being bypassed. IBM names the IT department, while TDWI also names data leadership. The second is how big the problem is. The IBM definitions treat it as people using the wrong tools. The 2026 preprint, which has not yet been peer reviewed, treats it as something larger: a slow wearing-away of the checks an organization relies on.⁵

There is also a fair objection to the term itself. If most people using AI bring their own tools,² and they do it to get work done,¹ then "shadow" may say less about the employees than about tools the organization failed to provide. That view is strong. It doesn't make the risks go away, but it does suggest that bans alone won't fix the problem.

In the news

In October 2026, a Baltimore hospital turned the term into a real case. WMAR reported that "Johns Hopkins Bayview Medical Center has begun notifying about 3,182 patients" after "a staff member uploaded a spreadsheet of patient information to an outside AI platform that was not cleared for that use."⁶ The hospital "said it discovered the problem on Sept. 21, 2026," and "began mailing letters to affected patients on Oct. 22, 2026."⁶ According to WMAR, "The exposed details may have included patient names, dates of admission and treatment, medical record numbers," among other details.⁶ The trade site healthsystemCIO headlined its account "Johns Hopkins Shadow AI Breach: PHI Uploaded Despite Policy and Approved Tools."⁷ (PHI means protected health information.) The headline makes the point that a policy and approved tools were already in place.

The wider numbers tell the same story. IBM's 2025 breach research, which Ponemon Institute conducted and IBM sponsored and analyzed, studied breaches at 600 organizations. It found that "One in five organizations reported a breach due to shadow AI, and only 37% have policies to manage AI or detect shadow AI."³ Organizations with high levels of shadow AI "observed an average of $670,000 in higher breach costs than those with a low level or no shadow AI."³

What it means for your organization

This is our perspective so far, and we treat it as a hypothesis, not a verdict. Shadow AI is a question about who an AI tool answers to. When an employee uses an unapproved tool, the organization isn't in the chain at all: it can't see what the tool was given or what it did. (See [principal](/definitions/principal) and [a person in the middle](/definitions/a-person-in-the-middle).) The Hopkins case also suggests that a written policy may not be enough on its own.⁷ We are watching to see whether organizations that hand out good approved tools see less shadow AI. Evidence that bans or approved tools make no difference would change our minds.

One question to ask your team: "Which AI tools did you use for work this week, and which of them did we give you?"