Hamelberg AI

Analysis · Reading public data

More articles do not necessarily mean more developments.

A dashboard can count correctly and still be read incorrectly. The first question is what each row represents; the next is which rows never reached the chart.

Hamelberg AI evidence note · Published 14 November 2026 · All numerical examples are invented for explanation, not observations from the live projects.

Start with an explicit unit

A source page is a publication. A development is the underlying event or announcement that several publications may describe. A classification is an interpretation of material that was available to read. These units answer different questions.

If five outlets cover one AI launch, five source records preserve useful evidence about coverage. Treating those records as five launches would inflate the count of developments. Collapsing them into one source would instead hide differences in reporting. A useful archive can retain all five sources while grouping their shared event once.

A worked example: 12 pages, four developments

Suppose one weekly collection contains 12 source pages. Six discuss the same product launch, three discuss a workplace study, two cover a court decision and one reports a school pilot. There are 12 publications but four distinct developments under that grouping rule.

Hypothetical coverage and development counts
SubjectSource pagesGrouped developments
Product launch61
Workplace study31
Court decision21
School pilot11
Total124

The launch accounts for half the publications and a quarter of the developments. Neither fraction says that half or a quarter of AI’s real-world effects came from that launch. The denominator describes this collection. If a second event was missed completely, grouping the retained pages perfectly would not reveal it.

Grouping also requires judgement. A follow-up about a genuinely new deployment may deserve a separate development even when it mentions the original launch. A correction to a previously reported number may belong with the existing development. Keep source links and the grouping rationale available so readers can question that decision.

Missing text changes the denominator again

Now imagine a video archive containing 40 collected videos. Usable transcripts are available for 25; 15 remain unclassified. Suppose 10 of the 25 classified videos contain a particular security theme. The defensible classification statement is “10 of 25 classified videos contained the theme.”

Reporting 25% of all collected videos would quietly assign absence to the 15 unread videos. Reporting 40% of all public videos would quietly extend the result beyond the collection. Even a range of 25% to 62.5%, obtained by assuming none or all of those 15 contain the theme, would describe only this 40-video archive. It would not correct for videos that were never collected.

A useful caption

“Theme present in 10 of 25 classified videos; 40 videos collected in total, including 15 awaiting usable transcripts. Counts describe the retained sample.”

The caption keeps the missing material visible instead of allowing the chart to imply completeness. The numbers here are hypothetical.

Why a country comparison is harder

A country tag can describe a video’s subject without describing the location or nationality of its audience. If one country has more sampled sources, more accessible transcripts or more active channels in the collection, raw counts can rise without any change in the underlying prevalence of a narrative.

Dividing by the number of classified sources helps answer a narrower question about the classified sample, but it cannot repair an unrepresentative collection. Transcripts might be missing more often for certain languages or formats. Repeated releases from a single institution might dominate one country’s total. Record those differences before interpreting a ranking.

Audience comments add another selection layer. Commenters chose to participate, accounts may not correspond to unique people, and the country mentioned in content is not an audience-location measure. Agreement with a source’s message cannot establish that the source caused an opinion, or that a national population shares it.

Apply the distinction to the projects

The AI Empowerment Observatory’s methodology explains how reporting is grouped and how retained evidence enters its readings. The European Security Monitor’s research guide explains its source collection and provisional classifications. Use the live pages for actual totals; this note supplies a way to interrogate those totals, not an additional dataset.

Before sharing a chart, write down its unit, time window, inclusion rule and missing evidence. Then write one sentence about what it cannot establish. If that sentence changes the intended headline, keep the limitation with the finding rather than relegating it to a separate page.

Use the reading guide’s evidence record →