When AI-generated summaries are wrong, junior staff often end up carrying responsibility for outputs they did not actually verify.

Junior researchers and field teams are increasingly using AI to synthesise messy field notes into clean summaries. This module teaches teams how to recognise when they have stopped being analysts and started being "liability sponges" — signing their names to aggregated narratives created by an algorithm.

The pattern is structural. The person positioned to absorb blame when something goes wrong is almost never positioned to halt the workflow when they see it going wrong. That is what "human in the loop" usually means in practice, and the fix is to redesign the loop — not to add another review stage on top of it.

The seam is visible in rhythm long before it is visible in evidence — but only if the institution knows what to do with suspicion.

Before anyone can be accused of anything, there is suspicion. Most institutions either skip from unease directly to indictment, or talk themselves out of noticing at all. This module trains managers to read the third option: calibration.

Production can now be proxied by machines. Navigation cannot.

The true test of professional ownership is not the quality of the artifact — it is the author's ability to navigate the logic of what they have produced. This module operationalises that test as a routine practice of live co-design rather than a prosecutorial interrogation.

Because commercial AI is trained to be polite and agreeable, it inherently dilutes the severity of community grievances. We map exactly who pays the price for that sanitised data.

Community grievances and social realities are rarely friction-free. This module introduces the Victim Register mechanism: training teams to stress-test their own reports by shifting the vantage point to the person downstream who absorbs the cost of the smoothed, sanitised abstraction.

The standard review cycle asks is this defensible to the client? This module trains teams to ask, first, is this defensible to the people whose lives are described in it? Reports that fail the first review get published and quietly resented. Reports that fail the second get contested — sometimes quietly, sometimes with lawyers.

If you cannot legally look under the hood of your vendor's algorithm — or specify AI-use expectations before a subcontractor starts work — you are not governing. You are hoping.

This module embeds the prior five into the documented structure of the practice: vendor contracts, subcontractor agreements, and the recruitment process itself. The goal is that when an incident surfaces, the organisation has a defensible position before the media, the regulator, or the client asks.

• • Vendor procurement — the Calvin Convention. Non-negotiable clauses: the right to interrogate the model, demand evidence access, and halt processing. Vendor Interrogation Scripts for the procurement conversation.
• • Associate and subcontractor contracts. Contract language that sets AI-use expectations before work begins, not after something goes wrong.
• • Recruitment policy supplement. Extending the live-edit logic into the interview process. Hiring the liability sponge in is as expensive as creating one inside.
• • Documentation discipline. What has to be logged, by whom, at what granularity, to make an AI-use claim verifiable six months later.
• • Incident handling without scapegoating. The difference between "an employee was dismissed" and "the organisation lost a contract" is often whether a written guideline existed beforehand.