Knowledge

Shadow AI: what it is and how to control it without a ban

Shadow AI is the use by employees of artificial intelligence tools that the organisation has not approved and does not oversee - most often public chatbots on private accounts. Data, knowledge and decisions then end up beyond the organisation’s reach, while responsibility for the consequences stays with it

46%civil servants (NASK)
59%employees
5steps
3.10.2026legal status
01 / 10

What shadow AI is

The term echoes shadow IT, i.e. software used without the IT department’s knowledge. The literature describes shadow AI as the unsanctioned use of AI tools outside the organisation’s oversight. Its most common forms:

Private account

an employee pastes documents into a public chatbot on a private account; five questions worth asking before pasting are on the “Before you paste” cheat sheet

Shared account

a team shares an account in an AI tool with no contract with the provider

AI feature in software

an AI feature switched on in software the organisation already uses, with no assessment of where the data goes

Institutional shadow AI

whole units work with AI without procedures - it is the absence of an organisational decision that fails, not an individual

02 / 10

Scale

Two studies · 2025-2026
46%

civil servants use generative AI tools at work

NASK 2026 · 6,093 civil servants
87%

civil servants want training; 70% see internal guidelines as a condition for successful AI deployment

NASK 2026
59%

employees use unapproved AI tools; 75% of them share sensitive information

Cybernews, August 2025 · 1,003 US employees
In a public office of 100 people, around 46 use AI - most without a company toolFull sources in section 09
03 / 10

Risks

Personal data

Disclosure to a provider with no legal basis or data processing agreement (GDPR Art. 28), breach of security requirements (Art. 32)

LiableThe organisation as controller

Trade secrets and professional secrecy

Information is no longer confidential and its legal protection weakens

LiableThe organisation and, under criminal law, the person who disclosed it

Organisational knowledge

Know-how settles in chat histories on private accounts and leaves with the employee

LiableThe organisation loses an asset

AI Act compliance

The deployer of an AI system used at work is the organisation - without a tool register and training (Art. 4), compliance is hard to prove

LiableThe organisation

ISO certification

An external service processes information outside the management system (ISO/IEC 27001, ISO/IEC 42001)

LiableThe organisation - risk of a nonconformity at audit

Common denominator

The organisation is liable in all five areas - not the employee who “used AI at their own risk”

That is whya ban does not remove liability - section 05
04 / 10

Personal data protection and artificial intelligence

The AI Act does not replace the GDPR but adds to it: when an AI system processes personal data, all GDPR requirements still apply (Art. 2(7) AI Act). In practice, four things matter:

Legal basis and contract

Personal data may be entered into an AI tool only on a legal basis under Art. 6 GDPR (for special categories, also Art. 9) and with a data processing agreement with the provider (Art. 28). An employee’s private account has no such agreement

Minimisation

Only what the task needs goes to the model (Art. 5(1)(c) GDPR). Summarising a letter or drafting a reply usually does not require names, PESEL (national ID) numbers or addresses

Masking before sending

PROXY:AI detects and masks personal data before a query reaches the model, and the WORM log records who worked with which model and when

Impact assessment

Where processing is likely to result in a high risk, a data protection impact assessment is required (Art. 35 GDPR). For high-risk systems, the AI Act requires deployers to use the provider’s information in it (Art. 26(9)), and the fundamental rights impact assessment complements it (Art. 27(4))

61%

of surveyed civil servants believe AI may threaten the privacy of public office clients through unauthorised data processing (NASK 2026). You can check which data should not be pasted into a chatbot in the Data Guard game

05 / 10

Why a ban does not work

OFTEN CONFUSED

A ban is not the same as control

A ban

It does not remove the need, it only hides it. Employees move their use to private phones and accounts. The organisation no longer sees what is being used - and its liability does not change

vs
A safe alternative

A company tool at least as convenient as a public chatbot, with data under organisational control, an AI gateway and a log. Usage returns to the organisation’s view

0

A ban does not change who is liable - the organisation is equally responsible for data pasted from a private phone. It only changes whether it can see it

06 / 10

Five steps without a ban

1Inventory

Which AI tools employees really use: a no-blame survey, network traffic, AI features in software

2Policy

Approved tools, data that must not be entered, responsibility for outputs

3A safe tool

A company AI tool at least as convenient as a public chatbot

4AI gateway

A single control point: model inventory, personal data masking, limits, a request log

5Training

By role (Art. 4 AI Act): what is allowed, what is not and why Course 09 Official → Course 11 IT and DPO

07 / 10

How much shadow AI is in your organisation

Three questions, an estimate based on NASK and Cybernews studies and the first step - no form

Tool · estimate

Check your own organisation

Instant result · no e-mail
1 · How many people work with documents and e-mail
2 · Is there a company AI tool
3 · Is there an AI use policy
?

Please answer three questions - the estimate appears here immediately, nothing is sent anywhere. The figures come from the NASK 2026 and Cybernews 2025 studies

An estimate, not a measurement - the exact number comes from an inventory (a no-blame survey + network traffic)
09 / 10

Sources

For information only, not legal advice. As of 3.10.2026

FAQ

Questions about shadow AI

What is shadow AI?

The use of AI tools that the organisation has not approved and does not oversee, most often public chatbots on employees’ private accounts

Is using ChatGPT at work safe?

It depends on the version and the contract. On a private account the organisation has no contract with the provider, no say over data retention and no access to chat history. A tool approved by the organisation, with a contract, data controls and logging, is safe

Is it enough to block public chatbots?

A block with no company alternative moves usage to private devices. It is more effective to approve a safe tool and route traffic through an AI gateway

Who is liable when an employee pastes client data into a chatbot?

Towards the client and the supervisory authority - the organisation: it is the data controller under GDPR and the deployer under the AI Act. An employee’s statement that they use AI “at their own risk” transfers nothing

Where should a public office start?

With an inventory and a policy. Then a safe tool and training for staff - before a ban turns into a workaround

How does the AI Act relate to the GDPR?

The AI Act does not replace the GDPR: when an AI system processes personal data, both apply at the same time (Art. 2(7) AI Act). The GDPR determines whether and on what legal basis data may be processed, and the AI Act how to use an AI system safely and transparently. For high-risk systems, information from the provider is used for the impact assessment under Art. 35 GDPR

All questions →

We describe what a company and an employee risk when a contract or a client list ends up in a public chat in Pasting company documents into ChatGPT. Is it already a trade secret leak?