Hamelberg AI

Practical reading guide

Start with the claim. Then ask what the source can actually show.

An AI headline, a research finding and a public reaction can sound equally certain. They are different kinds of evidence. Use these checks when moving between Hamelberg AI’s three projects or reading the original sources they link to.

1. Identify who is speaking

A product launch can establish what a company announced. It does not, by itself, establish how reliably the product works in ordinary use. A study may test performance on a defined task; its sample, setting and comparison determine how far the result travels. A news account may accurately report either kind of claim without independently proving it.

On The Brief, read the summary and then open its source. Look for words such as “says”, “proposes”, “finds” and “may”. They signal different relationships between an assertion and the evidence. Research summaries indicate when a preprint or an abstract is the available basis. A short account is a starting point for reading, not a substitute for the underlying paper or reporting.

2. Check the unit being counted

One story may be covered by several publishers. The AI Empowerment Observatory groups source pages about the same development so additional coverage is not automatically counted as another event. Its people and AI readings are separate: a reported improvement in system capability does not automatically mean a benefit for workers, patients, students or other people.

The Observatory’s discovery markets are a bounded source sample. A market label says where reporting was found, not necessarily where the event happened. Incomplete source text is excluded from the analytical totals. The published numbers describe the evidence retained under those rules, rather than all AI news or all effects of AI.

3. Keep observations separate from explanations

The European Security Monitor records how sampled public videos and official government communications frame security topics. A country tag describes the subject of a source. Comment categories and platform engagement describe recorded responses; they do not identify a national audience or reveal why a person reacted.

When comparing countries or time periods, check the collection window and missing sources. A change in the chart can reflect new publications, archive access, translation quality or classification status. It cannot, on its own, establish hostile intent, causal influence or a change in public opinion.

Keep a short evidence record

Copy these prompts into your own notes. Keep the record next to the claim when sharing it; a source link without the limit can still leave readers with the wrong impression.

Claim, in its original scope
What exactly is said to have changed, for whom, and during which period?
Source and access
Record the original URL, author or institution, publication date and the part you could actually inspect. An abstract and a complete study are different levels of access.
Measure and comparison
Write the unit beside the number. Identify the baseline: another group, an earlier period, a benchmark or no comparison at all.
Missing evidence
List important exclusions, unread source text and outcomes that were not measured. Leave unknowns as unknowns.
What would change the conclusion?
Name a plausible check: a quality audit, a different source sample, a longer follow-up or evidence about who carried the correction work.

A worked reading, not a new finding

“The organisation says its new assistant halves drafting time” establishes an announcement. Without a method, it does not yet establish a measured effect. Your record might read: “Company claim; task and comparison unspecified; quality and checking time not reported.” That statement neither endorses nor disproves the claim. It accurately describes the evidence available to you.

If a later evaluation measures time on a defined task, update the record rather than treating the announcement and evaluation as equivalent sources. If checking work moves to another team, account for that work before translating the result into an organisation-wide time saving.

For a published workplace study with a clearly defined productivity measure, read our note on productivity and human control. For an invented dataset that makes denominator choices visible, read the source-counting example.

A four-question check

  • Source: Who made the original claim, and can I open that source?
  • Scope: Which people, sources, countries and dates were included?
  • Method: Was this a reported claim, a measured result or an AI-assisted classification?
  • Limit: What might be missing, uncertain or different outside this sample?

These checks are useful even when a finding is persuasive. They help preserve the distinction between a source’s statement and what a reader can reasonably conclude from it. See who publishes these projects and how to contact us →