3 Levels of AI Use Disclosure
Moving beyond simple “AI or not” checkboxes is necessary for effective governance.
The content explores how to categorize AI use into three distinct levels to improve oversight. Level 1 covers human-led work with minor AI edits, such as fact-checking or copy editing. Level 2 involves human-led outlines filled in by AI, which requires deeper scrutiny to ensure the output matches the original intent. Level 3 covers one-shot generation where the AI does most of the mental work. The text also explains that it is mathematically impossible to prevent all “jailbreaks” because the variety of possible prompts is infinite. A recent example shows how an attacker used Morse code to bypass safety filters and move funds by exploiting how models decode different formats.
For public sector organizations, these points shift the focus from total prevention to active risk management. Rather than trying to eliminate every possible vulnerability, the goal is to make finding exploits more difficult and expensive than the reward for the attacker. Implementing a tiered disclosure system helps teams decide how much scrutiny a document or piece of code needs based on its AI involvement. It also shows that safety training often fails when inputs are encoded in non-standard formats. This highlights the need for multi-layered security and human-in-the-loop oversight when deploying AI in public services to meet safety standards.
How is your organization currently measuring the level of AI involvement in its workflows?
#AIGovernance #PublicSector #AISecurity #RiskManagement #AIPolicy
https://insight.trustible.ai/p/3-levels-of-ai-use-disclosure