Shadow AI is employees using AI tools — ChatGPT, Claude, Gemini, translation and summarisation apps — outside any sanctioned, visible channel. It is the AI-era version of shadow IT, with a sharper edge: the input is often confidential text pasted in by the very people producing the most work. Bans reliably make it worse by pushing usage onto personal devices where no control can see it. The practical response is to measure it (discover what is actually being used), then make the sanctioned path safer than the shadow path with a control that masks sensitive data at the point of use.
Why it happens — and why bans backfire
People use shadow AI for the same reason they used shadow IT: the tools work and the official alternative is slower or absent. Samsung’s 2023 episode is the textbook case — engineers pasted source code into ChatGPT to fix bugs faster, three leaks in about twenty days, followed by a company-wide ban and an internal survey showing most staff considered generative AI a security risk anyway.
A ban removes the visible usage, not the usage. It relocates the risk to phones and home laptops, where neither policy nor telemetry reaches. The firms that handle this well treat shadow AI as a demand signal: staff are telling you which capability they need.
Measuring your real exposure
Industry measurements consistently find that a meaningful share of what employees paste into AI tools is sensitive — DLP vendor Cyberhaven’s analyses put the sensitive share of AI inputs in double-digit percentages, concentrated in a small group of heavy users. Your firm’s number is unknowable until you look.
A practical first step is a time-boxed risk scan: run a point-of-use detector across a sample of real usage for two weeks and count what would have leaked — categories, volumes, tools — without storing the content itself. That turns an abstract worry into a concrete, prioritisable list.
The short version
- Shadow AI = unsanctioned AI use, invisible to IT and compliance.
- Bans relocate the risk to personal devices; they do not remove it.
- The heaviest AI users are usually the most productive staff — and the biggest paste risk.
- Measure first: a two-week risk scan turns the unknown into a prioritised list.
Related questions
Is shadow AI actually common in professional firms?
The SRA reported as far back as November 2023 that, reportedly, three quarters of the largest law firms were using AI — and informal use runs ahead of sanctioned deployment everywhere. If your firm has no approved AI path, assume the shadow path exists.
What should a firm do first?
Discovery before policy: find out which tools are in use and what kinds of data reach them. The AI Confidentiality Checklist on our compliance hub structures that review in about 20 minutes; a two-week risk scan quantifies it.
Does making one tool official solve shadow AI?
It helps — if the official tool is genuinely good. But coverage still matters: a point-of-use control that masks sensitive data on whatever tool staff open protects you during the long tail of partial adoption.
This page is general information, not legal advice. Where third-party guidance or law is summarised, read the originals via the source links before relying on them. Last reviewed: 17 July 2026.
This is the control in action
Identifiable data is masked in the browser before the prompt ever leaves — try it yourself with a sample prompt, or bring one low-risk workflow to a 20-minute review.
Advise Sarah Thompson, NI AB123456C, on the settlement offer.
Advise <PERSON_7K9X>, NI <NI_8W1R>, on the settlement offer.