Shadow AI

Starter
AI Security
Managers & front offices
Policy officers

Also called: unapproved AI use, shadow IT.

Detail

In five seconds

What it is

Use of AI tools outside approved organisational processes, especially where data, security, legal or oversight rules are bypassed.

In plain words

People use an AI tool for work without the organisation approving or supervising it.

An everyday example

A team in Addis Ababa and Bogotá needs quick translations during a busy week. One colleague starts using a free phone app to translate draft meeting notes because the approved service feels slow. Soon others send the app short extracts, names and schedule details because the translations come back in seconds. The work gets done faster, but no one in the organisation knows what was uploaded, whether the app stores it, or who can view it later.

Where the story stops working

This analogy focuses on data leaving the organisation, but shadow AI is broader. It can also involve locally installed tools, browser plug-ins or team-built systems that never send data to a public app but still bypass approval, logging, procurement or security review. Also, not every unapproved use involves sensitive data; the core issue is lack of authorised oversight, not only secrecy.

Remember thisShadow AI is not limited to secret chatbot use. It includes unapproved plug-ins, connected services and personal accounts whose outputs quietly enter formal workflows.

This term is contested

The exact boundary varies by organisation. Some use the term only for public generative AI tools, while others include any AI capability adopted without formal approval, including local models, browser extensions and vendor features quietly added to existing software.

Understand it

Why it matters

For non-specialists, this matters because it changes risk without looking like a security incident. A team may simply be trying to save time, yet the organisation can lose sight of what data is being processed, which vendors are involved, what contractual terms apply, and whether outputs are being checked. That affects confidentiality, records management, procurement, legal compliance, auditability and trust. For leaders, shadow AI can create a false picture: the organisation may appear cautious on paper while unapproved tools are already shaping drafts, analysis and decisions in practice. Good decisions require visibility, approved pathways and clear rules for when AI may or may not be used.

How shadow AI appears and why it becomes a risk

After viewing this, the reader should be able to explain how a useful shortcut turns into a governance and security problem.

01
Work need
A person has a task such as summarising notes, translating text or drafting a reply under time pressure.
Show a document icon and a clock inside a rounded box labelled 'Work need'.
02
Quick shortcut
The person uses an AI tool that has not been approved or reviewed because it is easy to access.
Show a laptop and a sparkle icon in a second box labelled 'Unapproved tool' with an arrow from step 1.
03
Hidden data flow
Text, files or metadata move into the tool, but the organisation may not know what was entered, stored or shared.
Show arrows from a document into a cloud-shaped box, with a small question mark badge, labelled 'Unknown handling'.
04
Loss of oversight
Security, legal and managers cannot apply normal checks such as access control, retention rules, monitoring or procurement review.
Show a shield, checklist and eye icon faded or crossed in a fourth box labelled 'Controls bypassed'.

A left-to-right sequence of four labelled boxes. Read from first use to organisational impact. Each box shows one stage: work pressure, unapproved tool use, hidden data flow, and loss of oversight. Arrows show how convenience leads to repeated use and then to unmanaged risk.

On hover or tap, each step expands to show one example: summarising notes, translating text, uploading a spreadsheet, or using a browser plug-in.

Describe this without the diagram

The diagram presents four stages in order. First, a worker has a time-pressured task such as summarising or translating. Second, they choose an AI tool that the organisation has not approved because it is fast and easy. Third, work content or data goes into that tool, but the organisation may not know what data was entered, where it is stored, or who can access it. Fourth, because the tool sits outside approved processes, normal controls such as access restrictions, monitoring, retention rules and legal review are bypassed. The overall message is that convenience can lead to hidden AI use and then to unmanaged risk.

With reduced motion: A static four-box sequence with arrows, showing the same stages without animation: work need, unapproved tool, hidden data flow, and controls bypassed.

This term means staff use an AI tool that their organisation has not approved, reviewed or connected to its normal rules. The tool might be public, privately bought, or added by a team without central oversight. The input can be a prompt, a document, meeting notes, code or spreadsheet data. The output can be a summary, draft, translation, analysis or recommendation.

What happens in the middle is the main risk. Because the tool sits outside approved processes, the organisation may not know what data was entered, where that data was stored, who can access it, or how the tool uses it. It may also be unclear whether the tool keeps prompts, uses them for further training, shares them with other services, or applies proper access control, meaning rules about who can see or do what.

The result may look useful, so the practice can spread quickly. But the organisation loses visibility and control. Security teams may be unable to monitor use. Legal and policy teams may not be able to check contracts, retention rules or human rights implications. Leaders may think AI use is limited when, in practice, it is already affecting real work.

Shadow AI is related to shadow IT, which means unapproved technology use more broadly. It is not the same as an AI attack such as prompt injection or data poisoning, though it can make those problems more likely because safeguards are missing. It is also not the same as confidentiality itself: confidentiality is the requirement to protect information, while shadow AI is one way that requirement can be put at risk.

How “A private translation app used for office notes” maps onto the real thing

In the story	In the AI system
Free phone translation app used without approval	Unapproved external AI tool
Staff paste in draft notes and names	Organisational data entered into the tool
Fast translations encourage more people to use it	Convenience drives hidden adoption
No one knows what the app keeps or shares	Loss of oversight over storage, access and data handling

A worked example

An information management officer at a regional hub must consolidate implementing-partner updates into a monitoring report by close of business. An AI add-on promises to search the shared reporting folder and draft the first version.

Starting from: The officer authorises the add-on through a personal account and grants it access to the folder, rather than submitting the service through the established approval process.

The add-on reads reports, filenames, document metadata and material in connected folders, not only the passages the officer intends to summarise.
It produces a polished draft and citations to source files, making the service appear ready for routine use.
The officer circulates the draft for clearance, but colleagues cannot see that an unapproved service contributed to it.
There is no confirmed record of the add-on’s access scope, processing terms, retention settings, deletion controls or activity logs.

Result: The draft may be useful, but the workflow is shadow AI. An unapproved service has accessed organisational data and supplied text for a formal report without the required security, legal, procurement and oversight checks. The access should be removed and the incident handled under the office’s applicable guidance.

In a United Nations workplace

After a vendor demonstration in a procurement unit, a political affairs officer considers activating the vendor’s public AI transcription plug-in to turn the discussion into a briefing note before the six o’clock clearance round.

The plug-in may sit outside the approved architecture even if the underlying product looks familiar. A personal account, unclear processing terms, unmanaged access and absent logging would leave no reliable organisational record of what the service received or retained. Its summary could nevertheless enter the file as if produced through an approved workflow.

Before you try thisDo not record, upload or process vendor submissions, pricing, evaluation comments or other procurement-sensitive material through an AI service that has not been approved for that use. Procurement and security review must cover the service, plug-ins, account controls, data handling, retention and deletion before it is used.

Where you may hear it

At a vendor demonstration when someone asks whether the AI plug-in is covered by the existing approval
During clearance of a briefing note whose drafting history is unclear
In an ICT governance board reviewing unsanctioned browser extensions and personal accounts
At a procurement meeting examining what data an AI-enabled service can access

Avoid confusion

Often confused with

Common misconceptions

Common misconception: Shadow AI only means staff are secretly misbehaving.
Correction: Not necessarily. It often starts with ordinary work pressure and a useful shortcut. People may not realise a tool is unapproved, may assume free consumer services are harmless, or may not know the organisation has separate rules for public and approved systems.
Common misconception: If no obvious hack happened, there is no security issue.
Correction: Security is not only about someone breaking in. Risk also comes from data going into systems outside approved controls, from unclear storage and retention, from weak access rules, and from outputs being used without review. The harm can come from process gaps, not dramatic intrusion.
Common misconception: Shadow AI is the same as data exfiltration.
Correction: They overlap but are not identical. Data exfiltration means data is taken out without authorisation. Shadow AI is broader: it includes any unapproved AI use, even if no sensitive data was entered. A harmless-seeming draft request can still be shadow AI because oversight was bypassed.

Quick check

Which example best fits shadow AI?

Go deeper

The simple version

People use an AI tool for work without the organisation approving or supervising it.

The technical version

Shadow AI is the use of AI systems, models, plug-ins or AI-enabled services outside established organisational governance. Typical gaps include missing procurement review, absent security assessment, unclear data-processing terms, unmanaged identities and access, no logging or monitoring, and no defined retention or deletion controls. Inputs may include prompts, files, metadata or connected service data. Outputs may then enter formal workflows despite the tool sitting outside approved architecture and policy. The security issue is therefore not only model behaviour; it is the loss of visibility and enforceable controls around data handling, access, audit and accountability. The term is related to shadow IT, but focuses on AI-specific uses and integrations.

Reviewed external resources

Primary sourceOWASP Top 10 for LLM Applications (opens in a new tab)OWASP
OWASP's Top 10 for LLM Applications gives a practical security risk framework for language-model systems and suits practitioners, managers and policy readers who want to place shadow AI among other AI risks.
Intermediatelink checked 2026-07-31
Primary sourceAdversarial Machine Learning: A Taxonomy and Terminology (opens in a new tab)NIST
NIST's taxonomy and terminology resource offers careful definitions for adversarial machine learning terms and suits readers who want to distinguish governance problems like shadow AI from attack categories.
Deeperlink checked 2026-07-31
EncyclopaediaShadow IT (opens in a new tab)Wikipedia
Wikipedia's Shadow IT page gives broad background on unapproved technology use in organisations and suits readers who want the wider context behind this AI-specific term.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Shadow AI (opens in a new tab)YouTube search
The YouTube search results page can help readers find short video explainers on shadow AI from different creators, and it suits people who prefer audiovisual introductions over text.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Shadow AI (opens in a new tab)Reddit search
The Reddit search results page lets readers see community questions and experiences about shadow AI, and it suits readers who want informal discussion rather than an official definition.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

Shadow AI is not limited to secret chatbot use. It includes unapproved plug-ins, connected services and personal accounts whose outputs quietly enter formal workflows.