
Show your work, AI. Then let us check it.
An AI agent's explanation is not the same thing as a record of what it did. Microsoft's new AI Transparency Initiative takes that distinction seriously, proposing more consistent ways to inspect models and agents across organizations. More than 40 partners are involved. The promising idea is evidence that can travel across tools, rather than a different reassurance from every vendor. But an initiative is not an audit. The November 17 announcement does not establish a completed standard, a certified product or a measured reduction in incidents. Its work includes standards participation and planned integration. Customers should welcome that direction while asking when usable specifications, test results and independent checks will arrive. The distinction becomes sharper when agents act. A system might explain why it wanted to retrieve a document; an external record can show which identity was used, whether access was allowed and what actually happened. Neither record necessarily reveals the model's complete internal decision process. Combining them can still be more informative than treating the model's own account as the sole witness. The morning's model-safety disclosures reinforce that need for care: monitorability depends on what evidence a monitor receives and the conditions of the test. Our assessment is cautiously positive. Better records could make troubleshooting, accountability and switching vendors easier. The test is whether those records are usable when something goes wrong—not how many organizations endorse the announcement. Transparency earns its name when a customer can challenge a claim and follow the answer back to evidence.

































































































































































































