More than one in four government professionals now use AI tools that their IT departments never approved, a practice known as shadow AI. The figure, 27% according to Thomson Reuters' 2026 Future of Professionals Report, is slightly lower than the 34% average across all industries, but it still creates substantial governance, security, and accuracy risks for agencies that operate under public accountability and strict regulatory obligations.
Shadow AI refers to any use of chatbots, automated coding assistants, or other AI tools by employees without IT's knowledge or consent. It is the modern successor to shadow IT, yet the difference is critical. These tools do not simply store or transmit data - they generate content, draw inferences, and shape decisions, often leaving no record of how they arrived at an output.
The behavior is not purely a matter of employee intent. Government procurement cycles are lengthy, while free or consumer-grade AI tools are immediately accessible. A survey by the National Cybersecurity Alliance found that 43% of AI users have already shared sensitive information with AI tools without their employer's knowledge, evidence that the gap between approved and actual use runs wide across sectors, including the public sector.
Three risks agencies cannot ignore
Unsanctioned AI use introduces several concrete dangers that directly affect public-sector operations.
- Data security - When employees input information into public AI tools, that data can leave an agency's control entirely and sit on third-party servers with no visibility into retention or use. For agencies handling citizen records, law enforcement data, or classified material, this is a real exposure point. IBM research notes that one in five companies in the United Kingdom has already experienced a data leak tied to generative AI use. In government, the consequences go further: FOIA obligations, statutory retention requirements, and public accountability standards were not designed with unsanctioned third-party AI platforms in mind.
- Accuracy of information - Generative AI tools produce confident, plausible, and frequently incorrect output, a well-documented limitation known as hallucination. Without a review process, these errors can move directly into memos, public statements, or policy documents. The risk grows as reliance on AI increases. Separate research from cyber-risk monitoring firm UpGuard found that roughly one-quarter of employees now consider AI tools among their most trusted information sources, rivaling the trust they place in their own managers.
- Process consistency - When employees depend on different tools, prompts, and informal standards, the resulting work product varies in format, quality, and traceability. That inconsistency complicates audits, obscures accountability, and undermines the standardized processes public institutions rely on. Zylo, a technology management platform, notes that shadow AI use is difficult to detect through conventional network monitoring because it often occurs within otherwise approved platforms or embedded workflows, making it harder to determine who did what, and on what basis, after the fact.
Guardrails, not bans
Outright prohibition tends to push the behavior further out of view rather than eliminate it. Government agencies that have managed the issue effectively tend to converge on three practices.
First, a clear AI policy that defines which tools are approved, what categories of data may or may not be entered into them, and who is accountable for reviewing AI-assisted output before it is used. IBM recommends pairing policy with operational guardrails - sandbox environments, access controls, and defined escalation paths - rather than relying on static, one-time guidance. Resources like AI Learning Path for Policy Makers can help leadership draft governance frameworks that match the sensitivity of government work.
Second, cross-functional training. Most employees turning to unsanctioned tools are not trying to circumvent policy; they are trying to meet workload demands with the tools available to them. Onspring's analysis shows that training programs that involve IT, legal, and security functions together are far more effective than policies issued by IT alone, especially when paired with sanctioned alternatives that meet the same operational need. Ongoing education through initiatives like AI for Government can reinforce which tools are approved and how to use them responsibly.
Third, technical controls. Network-level controls, endpoint monitoring, and application allow-lists narrow the gap between approved and actual use. These measures work only if employees are simultaneously given sanctioned tools that meet their needs - otherwise, the guardrail may simply drive behavior further underground.
Why this matters for government
The question facing every agency is not whether employees will use AI, but whether that use will occur within an approved framework or outside of one. Building that framework requires two constants: fiduciary-grade tools vetted for the sensitivity of government work, and sustained human oversight over any AI-assisted output before it becomes public record. Agencies that establish clear policy, invest in training, and pair those measures with technical enforcement are better positioned to capture AI's benefits while preserving the data security, accuracy, and procedural consistency that public accountability demands.
Your membership also unlocks:
