How a case is verified

A case is an institution (a court system, a journal, a bounty program, a platform, sometimes a city, a country or a profession), the assumption that broke (that writing costs effort, that volume means interest, that a polished text means a competent author, that a reply comes from a person), the first-order effect with the numbers the source gives, the adaptation the institution has made so far, and who pays the externality. Dated, sourced, sober. No forecasts.

The ledger has two halves. A side effect is the default. A resilience case is one where the response is the story: the shock, what the institution did, and the measured outcome. Both share one record shape, so one page grammar serves both.

1. Discover. Once a week, four families of web queries: institution-agnostic phrases in English and German, twenty niches a model proposes given what the ledger already holds, the practitioner channels where "we are drowning in AI-written X" is said first, and the positive phrasing that finds institutions that coped.
2. Extract. One model reads each page and decides whether it describes a case. If it does, it fills the fields from the page only, and lifts every number, quote and date verbatim into a list of claims. The tenth article about the same institution updates the existing case instead of adding one.
3. Verify. A second model from a different provider reads the fetched pages, never the web or its memory, and gives every claim a verdict: supported, partly, not supported, or unverifiable. A case becomes verified only when every claim is supported. Anything less stays a candidate and never shows here.
4. Review. One person reads the week's candidates and what changed, keeps, kills, or marks a case to write about. A killed case never counts.

Each case page shows the evidence pages, the quotes, and the second reader's verdict on every claim. That table is the whole promise of the site: "every number checked twice" is literal.

How far an institution has reacted is the trend signal, not the day's headline. Categories tend to move up the ladder, and a rung moving is what the Side Effects page exists to show.

none

nobody has reacted yet

detect

the institution can tell machine from human, and stops there

verify

every submission is checked before it counts

price

a cost on the sender; the filter is money again

tie

submissions are tied to a person

reform

the rule itself was rewritten

Every case names who carries the externality in a sentence. The extractor also files it under one of ten groups, so the Who Pays page can add them up:

Volunteers and reviewers

unpaid gatekeepers: peer reviewers, maintainers, moderators, volunteer editors · volunteers-reviewers

Clerks and public staff

paid staff of an institution: judges, clerks, case workers, librarians, teachers, nurses · clerks-public-staff

Students and learners

students, trainees, juniors who lose the practice that trained them · students-learners

Job candidates

applicants and candidates drowned in a flooded pipeline · job-candidates

Platform workers

gig workers, freelancers, independent researchers paid per task or bounty · platform-workers

Patients and clients

the people the institution serves: patients, litigants, customers, applicants, readers · patients-clients

Creators and small publishers

authors, musicians, indie developers, small outlets displaced by volume · creators-small-publishers

Public budgets

taxpayers and the budgets of courts, agencies, universities · public-budgets

The general public

everyone: the record, the commons, trust in a procedure · general-public

Other

a group the taxonomy has no row for; the case names it · other

The ledger is published under CC BY 4.0: use it, cite it. JSON and CSV carry every verified case with its evidence URLs; the Atom feed carries the newest fifty. New cases land on Monday mornings, Berlin time. Cite a case by its page, which is stable: the id in the URL decides.

Suggested citation: targeting.ai, The ledger of unintended AI side effects, https://society101.targeting.ai, retrieved on a date.

The schema, the ladder and the first six cases come from the essay Friction Was the Filter (Towards AI, September 2026), which read James S. Coleman against Germany's social courts, curl's bug bounty and 15,899 AI-written peer reviews. The collector is open about its limits: sources are the web as the queries find it, English and German first, and a verified case is only as good as the page it rests on. Corrections reach the author through targeting.ai.