Definition of shadow AI
Shadow AI (in French, IA fantôme) refers to the use of artificial intelligence tools by members of an organisation without its framework or approval. It is a term from common usage, formed by analogy with shadow IT: the hardware, software and online services used without the IT department having approved them.
The consolidated text of Regulation (EU) 2024/1689 on artificial intelligence, read in its version of 27 July 2026, does not use this term. The criterion is not the tool but the situation: the same assistant counts as shadow AI in an organisation that has not examined it, and as ordinary use in another that has authorised and framed it.
What the term makes it possible to name
The term refers to a gap between actual uses and what the organisation knows about them. Three questions then go unanswered.
- What data is leaving? Text entered into a tool whose terms of use have not been read may contain personal data, customer information or elements covered by trade secrecy.
- Who checks the results? Content produced by a model may make its way into a deliverable or a decision without any identified review.
- What obligations apply? The General Data Protection Regulation (GDPR, Regulation (EU) 2016/679) applies as soon as personal data is processed. The AI Regulation attaches its obligations to the role held and the risk level of each system, notions presented in our entry on the AI Act.
Assessing an undeclared use against these texts is a matter for the organisation’s lawyers. The application dates of the AI Regulation are covered in our article on AI Act obligations for businesses.
The forms shadow AI takes
| Form | What the organisation does not know |
|---|---|
| Consumer assistant used with a personal account | The data entered and the terms on which the provider retains or reuses it |
| AI feature switched on in software already in service | The existence of the feature, which an inventory kept by application does not show |
| Extension added to a browser or an office suite | The content the extension can access |
| Script or automation calling a model with a personal key | The data flows and the spending |
| AI agent connected to internal tools by an employee | The actions the agent can carry out and the rights it holds |
Some arise from individual initiative, others from a vendor or a team.
Shadow AI, shadow IT and unlawful use: not to be confused
Shadow AI is a particular case of shadow IT, with two features of its own. It can arise inside authorised software, when a feature is switched on, without any new installation. It produces content (text, summaries, code) that is then reused in people’s work, so that the question also concerns what comes back from the tool.
The term says nothing about the lawfulness of the use. An undeclared use may be of no consequence; a declared and approved system may, conversely, breach an obligation. Nor does it refer to the practices prohibited by Article 5 of the AI Regulation: those are prohibited whether or not the organisation knows about them.
Example: a contract summarised by a consumer assistant
Take the legal department of a manufacturing company. A lawyer pastes the clauses of a customer contract into a consumer assistant, from a personal account, and obtains a summary that is circulated to colleagues.
None of this appears in the company’s tools. It does not know that confidential clauses have left its information system, on what terms the provider retains them or who checked the summary. The use becomes visible on the day the lawyer declares it, or when a survey brings it to light. The company can then examine it, and authorise it, frame it or offer another tool.
How to identify these uses, classify them and open an authorisation channel for them is described on our page on AI governance; the order of the steps, in our article on implementing AI governance.