An alignment researcher and Congressional staffer have very different backgrounds, context levels, and priorities when reading the same AI model safety evaluation — and so will come away with separate conclusions, assumptions, and action items. You just learned about the importance of considering upstream context and downstream consequences: now, you will apply this knowledge to distilling audience-specific information from a real report.
The Five Steps
In this exercise, you’ll pick a real report and work through it in five steps:
- Identify upstream. Whose data, claims, and access was this report built from? Pick the upstream actors from the list.
- Identify downstream. Who reads this report next, and what does each reader need from it? Select the audiences.
- Clip. Pull the passages that actually matter from the full report. Most of it won’t make the cut.
- Distill. Rewrite each clip for a specific audience. A researcher, a policy staffer, and a journalist need very different things from the same paragraph.
- Thread. Connect each distilled piece to the audience it serves. If a piece doesn’t serve anyone, it shouldn’t be in your report.
Worked Example
Here is one report taken through the first two steps. The Future of Life Institute's AI Safety Index grades AI companies on safety, and tracing where its grades come from and who reads them explains why the report is shaped the way it is.
The Future of Life Institute’s AI Safety Index, a scorecard that grades AI companies on safety.
- Written by: the Future of Life Institute, an advocacy nonprofit.
- Upstream: the labs’ own public documents (system cards, safety policies) and a panel of expert reviewers. FLI has no inside access; everything it grades comes from what companies chose to publish.
- Downstream: journalists (who need the headline grades), the graded labs (who need to know where they lost points), and the public (who need a simple comparison). Notice the grades compress detail by design; that’s what its audiences need.
Your Options (Pick One)
Four real reports are on the table, each written by a different kind of author: an AI company reporting on its own model, a government institute, an international inspectorate, and an export-control enforcer. Pick the one whose upstream sources and downstream readers you most want to work through.
- Claude 4 system card (Anthropic, 2025) — an AI company’s own report on its newest model’s capabilities, risks, and safety testing.
- Frontier AI Trends Report (UK AI Security Institute, December 2025) — a government institute’s public assessment of where frontier AI capabilities are heading, based on its testing of leading models.
- IAEA safeguards report on Iran (GOV/2026/8) — international nuclear inspectors’ quarterly findings, the closest existing analogue to what an AI treaty inspectorate would publish.
- BIS settlement order against Seagate (US Commerce Department, 2023) — the enforcement document behind a $300 million fine for shipping hard drives to Huawei in violation of export controls.
For each: where did it come from, who reads it next?
The exercise takes the five steps in the order that makes them workable: you clip and distil first, then trace the actors, then thread. The two identification steps above are steps three and four on the rail.