Departments run unsanctioned ChatGPT plugins on case data — classic shadow AI. What organizational risk does this describe?

Shadow AI is the organizational risk of unsanctioned generative tools processing municipal data outside approved governance, logging, and security boundaries. Departments running unsanctioned ChatGPT plugins on case data is the classic pattern: case content leaves the controlled environment, vendor terms and retention are unknown to the AI CoE, and inventory, DLP, and audit trails cannot see the processing. The risk is governance bypass and uncontrolled exposure—not innovation itself.

Model drift: unsanctioned generative tools losing accuracy on municipal case data over time describes performance degradation over time, not the governance-bypass risk of using unsanctioned tools itself. Prompt injection: unsanctioned assistants mishandling crafted inputs in municipal case workflows is an adversarial input-manipulation risk, not the unsanctioned-use governance risk described in the stem. Training data bias: unsanctioned triage tools skewing municipal case referrals across districts is a fairness problem of skewed priorities, not the shadow-AI risk of operating outside governance.

Exam caveat: shadow AI names the unsanctioned-use risk; options that describe model drift, prompt injection, or training-data bias name distinct accuracy, adversarial-input, or fairness failure modes, not the shadow-AI risk in the stem. Match the stem’s description of unsanctioned plugins on case data to the risk label, not to the fix.

Operational check: discover unsanctioned plugins via network and SaaS logs, force case-data workloads onto sanctioned private assistants, require CoE intake for new generative tools, and educate departments that convenience plugins on case files equal shadow AI.