Last reviewed:
What is shadow AI? Definition and pitfall to avoid
Shadow AI refers to the use of generative AI models by employees without validation or framing from their company. Widespread in 2025-2026, it creates legal, financial (leakage of business data to public models), and quality (unaudited outputs integrated into client deliverables) risks.
Shadow AI is the AI equivalent of shadow IT, extended in 2024-2025 to nearly all companies. The figures converge: 45.2% of US workers aged 18 to 64 used generative AI for work in the second quarter of 2026 (Federal Reserve Bank of St. Louis, FRED series), up from 37.4% in August 2025. This measure does not separate sanctioned use from shadow use. A PagerDuty/Wakefield Research survey (fielded April 2026, 1,250 office professionals in non-IT roles at companies with $500M+ revenue) quantifies the scale on the corporate side: 66% of those who use AI for work have done so with a tool they believed was not permitted. 88% have shared work information with a public AI tool: 34% customer data, 31% financial information, 31% confidential documents or strategies. Four mechanisms feed shadow AI. The free nature of consumer tools (ChatGPT, Claude.ai): an employee only needs a personal email address. Perceived performance: public models produce useful results immediately. Slowness of official deployment: between identifying a need and providing a validated tool, several months often pass. Lack of explicit policy: without a written AI policy, employees interpret silence as tacit authorisation. These mechanisms show up in the data: 89% of users first adopted the tool in their personal life, 44% use it to work around the limits of company-approved tools, and 38% have delivered AI-assisted work without disclosing it. Frequent consequences: transmission of confidential documents to external models, use of outputs in deliverables without legal review.
Concrete example
An 80-employee HR consulting firm detects in 2025, during a GDPR audit, that 45 consultants use ChatGPT from their personal account to prepare client assessments. Estimated cumulative volume: 8,000 conversations over 9 months, containing personal data of about 1,200 individuals (candidates, evaluated employees). Company response: formal information to the persons concerned, legal audit, implementation of a written AI policy, deployment of a validated internal tool (Claude Enterprise with European hosting), mandatory training. Total cost of the incident: 95,000 euros, including the audit, client communication, training, and three departures of employees not wanting the new framework.
A publicly documented case, on a larger scale: Samsung Semiconductor (Korea), March-April 2023. Three separate incidents in 20 days saw engineers submit proprietary source code, internal meeting notes and chip test data to ChatGPT (free account, data re-usable for training under the terms of service at the time). Samsung responded in April 2023 with a company-wide ban on ChatGPT across corporate devices and networks, then announced work on an in-house LLM. The incident served as a sector-wide wake-up call: Amazon, Apple, JPMorgan and Verizon adopted similar restrictions in the following months.
Comparison
| Shadow IT | Shadow AI | BYOD | |
|---|---|---|---|
| Scope | Unsanctioned apps or SaaS | Generative models (ChatGPT, Claude.ai) via personal accounts | Personal devices used at work |
| Primary risk | Security, data governance, uncontrolled spend | Data leakage + reuse by public models for training | Endpoint security, work/personal boundary |
| Detection | MDM / proxy audits, card statements | DNS audits + surveys + deliverable review | Mandatory enrolment or network blocking |
| Typical framing | Approved SaaS catalogue | Written AI policy + official channel of equivalent performance | BYOD policy + MDM on enrolled devices |
| Emergence | ≈ 2010 (consumer SaaS) | 2023 (public release of ChatGPT) | ≈ 2010 (personal smartphones at work) |
| GDPR impact | Varies by processed data | High as soon as personal data is sent to the LLM | Varies by access and enrolled apps |
FAQ
What's the difference between shadow AI and shadow IT?
Shadow IT refers to unsanctioned digital tools (SaaS, apps). Shadow AI is its 2024-2026 extension: use of generative models (ChatGPT, Claude.ai) via personal accounts, with an added risk: transmitted data is often used to train or improve the public models.
Is using ChatGPT in shadow AI illegal?
Not per se, but it often breaches GDPR (or equivalent laws) when employees send personal data (candidates, clients, staff) to a public model without a legal basis or informing the individuals. The controller (the company) remains liable for processing carried out through these tools.
How can I detect shadow AI in my organisation?
Three signals: proxy or DNS audits for the relevant domains (chatgpt.com, claude.ai, gemini.google.com), non-punitive anonymous surveys, and deliverable analysis (LLM-typical phrasing, uncorrected hallucinations). In France, 73% of private-sector employees use AI at work (Ipsos bva, May 2026).
Why do employees use AI in shadow mode?
Four documented reasons: free consumer tools, immediate productivity gains, slow official rollout (often several months), and lack of an explicit policy, so silence gets read as tacit permission. 89% of office professionals who use AI for work first adopted it in their personal life.
How do you frame shadow AI without killing innovation?
Three non-negotiables: publish a one-page AI policy (allowed tools, banned data, outputs requiring review), quickly open an official channel of equivalent performance, and prioritise training over sanction. Blocking without an alternative creates permanent workarounds.
What does a shadow AI incident actually cost?
Illustrative case: an 80-person HR consultancy finds that 45 consultants use ChatGPT on client data. Legal audit, client communication, training and three resignations bring the bill to about €95,000. Indirect costs (trust erosion, GDPR non-compliance) are typically far higher.
See also
Further reading
The State of Generative AI Adoption in 2025, Federal Reserve Bank of St. Louis, 2025
Sources
- Generative Artificial Intelligence, Adoption Rate for Work: Employed Adults (series RPSGENAIUSAGESHAREWORK, Real-Time Population Survey, workers aged 18 to 64), FRED, Federal Reserve Bank of St. Louis, Q2 2026 value released 4 August 2026. https://fred.stlouisfed.org/series/RPSGENAIUSAGESHAREWORK
- The State of Generative AI Adoption in 2025, Bick, Blandin & Deming, Federal Reserve Bank of St. Louis, 13 November 2025. https://www.stlouisfed.org/on-the-economy/2025/nov/state-generative-ai-adoption-2025
- CNIL, "Artificial intelligence" topic page (in French). https://www.cnil.fr/fr/intelligence-artificielle
- PagerDuty Shadow AI Survey, conducted by Wakefield Research, fielded 9-20 April 2026 (1,250 office professionals in non-IT roles, companies with $500M+ revenue, US/UK/JP/AU, margin of error ±2.8 pts). https://www.pagerduty.com/newsroom/shadow-ai-workplace-survey-2026/
- AI at work: behind mass adoption, shadow uses, Ipsos bva × Greenworking, 23 September 2026 (fielded 12-25 May 2026, in French). https://www.ipsos.com/fr-fr/ia-au-travail-derriere-ladoption-massive-des-usages-fantomes-et-des-managers-en-premiere-ligne