Latest Posts

December 2 closes the transitional window for marking synthetic content from systems already on the market. The Code of Practice is not a shield.

A gap appears, a team builds an internal tool, the tool becomes the dependency. Three gates stop that sequence, and a written record of the search makes them compound.

OSFI's July 2026 bulletin tells institutions to set limits on agent autonomy and strengthen senior management literacy. Both rest on a behavior nobody's measuring.

Leaders say they'd pass an AI audit today. Far fewer call their governance mature. The distance between those two answers has a name, a level, and a next rung.

The two-body picture of People+AI work shows accountability, not a headcount. Real decisions run on more sources, and collapsing them hides the disagreement worth keeping.

Colorado's replacement AI law reads as a notice bill. The draft rules decide coverage and human review by what the reviewer actually did, and that can't be attested.

A TrustCloud essay reduced the CISO's AI job to four board questions. Three map to systems that log evidence. The fourth is about people, and nothing logs it.

On October 1, Connecticut stops accepting the AI tool as a defense to a discrimination complaint. Here's what's live, what waits, and what the statute assumes about us.

Two states now write the clinician's review of AI-generated notes into law. Both assume the review works. The error studies say that's where the risk lives.

An AI system can follow our instruction badly, or it can serve an intent nobody in the room recorded. The second failure produces good work, which is why it survives review.

Gartner's first Hype Cycle for AI Governance plots 32 innovations. Its own survey says the organizations with the best policies only partly follow them.

The honest position on People+AI measurement is that the baseline does not exist. Here is the protocol for building one, starting with the step everybody skips.

A rollout can raise capability and leave judgment exactly where it was. The metrics we collect cannot tell the difference, because they are all measurements of the tool.

Annex III human oversight duties apply from December 2, 2027. The competence and override records they ask for take about that long to build. Here's the calendar.

Illinois priced each false citation at $1,500. Connecticut wrote verification into its rules. The USPTO disciplined a lawyer no court had sanctioned.

Discovery tools answer which AI systems are being used. They cannot answer whether anyone verified the output, and that is the exposure that reaches customers.

Approval volume rises, reviewer context falls, and the gate keeps producing signatures. AWS now names rubber-stamping as an architectural anti-pattern.

Six questions separate meaningful review from a nominal approval step. Five of them can be answered from records. The sixth requires watching someone work.

Seats provisioned and courses completed are inputs. The standard governance catalog measures systems and processes. Almost none of it measures the operator.

The Digital Omnibus rewrote Article 4 six days before the AI Act's biggest date. The literacy duty now states that it requires no specific level in any person.