Skip to main content

Knowing is not doing

Why frontier AI models can recite the rules and still get the ethics wrong, and what it means for compliance and ethics leaders

AI systems can state policies, explain principles, and produce convincing rationales. But knowing the rule does not guarantee that a system will follow it when authority, pressure, incentives, or competing goals enter the equation.

Knowing is not doing examines the growing “say-do gap” in AI: the distance between articulating an ethical principle and behaving consistently when that principle is tested. As AI moves from answering questions to taking action inside business workflows, that reasoning gap can quickly become a control gap—with legal, regulatory, and operational consequences for the organizations deploying it.

What this white paper reveals

  1. Knowing the rule is not the same as following it. AI systems may correctly explain a policy or ethical principle and still reason around it when completing a task.
  2. More reasoning does not necessarily mean better ethical judgment. A fluent, structured explanation can conceal fragile or inconsistent decision-making.
  3. AI failures tend to follow four patterns. Systems may agree instead of reasoning, reason backward from a preferred outcome, match words instead of principles, or lose sight of the person or value at stake.
  4. The risk grows as AI gains agency. When systems can book, buy, screen, change records, or make decisions, failures no longer remain in drafts waiting for human review—they can be executed at scale.
  5. Accountability cannot be delegated to the tool. Organizations need clear ownership, documented controls, ongoing monitoring, and a reconstructable record of how AI-enabled decisions were made.

A careful AI policy is not enough

The organizations best prepared for responsible AI will not simply have the strongest policy on paper. They will test how AI behaves under pressure, govern it like any other actor with access to company systems, and treat ethical performance as a capability to measure, not a quality to assume.