How it works · version 4.0
How the Observatory reads the news.
AIEO does not ask you to keep up with every AI headline. It gathers the coverage, groups reports about the same thing, checks what came before, and asks what the news shows for people.
The evidence path
Four steps from news to meaning
Every public card links back to the sources behind it.
Reading the weekly picture
One reading per development, on each side.
The same developments are read twice: once for people and once for AI and its operators. The analytical denominator contains only developments with a completed reading and at least one complete source whose saved content hash matches the classification input. Each side has four categories: gains only, losses or limitations only, both, and no direction stated. Each set adds up to the same complete-content total.
A source does not need to discuss both sides. A benefit for people can be recorded when no AI direction is stated, and vice versa.
What counts as a gain or a loss?
For people, the reading concerns ability, access, opportunity, control and welfare. For AI, it concerns capabilities, use, reach, resources and operating limits. Reporting may include attributed claims, forecasts, risks and recommendations. The source framing remains explicit; these counts do not establish independently verified effects.
Mixed means the source describes gains and losses together. Those findings are retained, including different effects on different groups. They are not cancelled into a neutral or unclear result.
No direction stated and incomplete evidence
A complete written source may not state a direction for one side. That is different from a source that could not be read completely. Missing bodies, unfinished readings, and audio or video summaries without complete transcripts are excluded from analytical totals and from the Brief export. Their source links and exclusion reasons remain in the collection audit. A missing companion article does not exclude a development if the reading used another qualifying complete source; only qualifying sources are passed to the Brief.
Content availability is checked automatically and does not certify that a classification is correct. Missing publishers can bias this sample. The complete-content results must not be presented as percentages of all AI news. Explore the source records and exclusions. Operationally unfinished classification runs still block publication.
How to read the patterns
People and AI are assessed separately.
The Observatory uses everyday wording and keeps the source evidence behind each card visible.
AI gains or limitations concern the system and its operating context. They do not make a claim about consciousness. Corporate growth can support an AI-side reading without proving a benefit for people.
Independent directions do not automatically establish a linked relationship. A statement connecting both sides is needed for that stronger interpretation. This week's main view presents the independent readings, without an additional paired-pattern subtotal.
How source pages become developments
Each AI-related source page published during the weekly period counts once. Pages about the same concrete occurrence are grouped into one development.
Extra pages = source pages minus distinct developments. They show added attention, not an added event.
How first recorded and seen before are decided
A development is first recorded when AIEO does not link it to an accepted development in the disclosed history. A later report about an earlier event adds attention but does not create another event.
Possible links are kept separate while they are checked. Earlier weekly releases can be corrected, and prior versions remain archived.
Why relationship patterns can overlap
The people side and the AI side are considered separately. A development may show people gaining, people losing ground, AI moving forward, AI being limited, or more than one of these at once.
A separate check asks whether benefits differ across groups. A development can therefore appear in more than one relationship pattern. This overlap applies to the detailed pattern counts, not the five-category weekly outcome breakdown.
What AIEO does not claim
The discovery markets are a research sample, not a complete census of global AI news. The market where a source was found is not necessarily the place where the story happened.
The Observatory records what the available news supports. A card does not prove a universal outcome, private experience, intention or objective system performance.
What is public—and what remains protected
AIEO publishes the concepts, sources, weekly scope, counts, revisions, limitations and the meaning of every public card. This is enough to understand and challenge the output.
Exact prompts, calibration thresholds, private orchestration code and abuse-resistant operating details are not published. This protects the integrity of the system without hiding how readers should interpret its findings.
Comparing editions
The independent-direction policy begins with the 24–30 August 2026 edition. Earlier single-category relationship counts are not directly comparable. Weekly source-page and development counts still describe reporting volume, not the amount of AI activity or its effect on society.
News discovery covers five search markets. Market counts can overlap when a source appears in multiple searches. English, French and Chinese written sources can all contribute.
See it in use