Questions
Frequently asked questions
What is AnomalyShield?
A risk-scoring system that ranks accounts by the estimated risk of age-inappropriate
adult-to-minor contact, so that trained reviewers look at the highest-risk cases first. It combines a
content-based score and a social-contact (network) score into one 0–100 composite. It is a
prioritisation tool, not an adjudication tool.
What does a score actually mean?
It means "look here first" — nothing more. A score only changes the order of the human review
queue. It never suspends, reports, names or otherwise acts on anyone by itself. A high score is
not a finding of guilt.
How accurate are the top-scoring flags?
At the very top of the ranking, most flags are genuine (illustratively about 9 of the top 10);
accuracy falls as you flag more widely, because genuine cases are rare. The full calibration, precision@k
and uncertainty analysis is in the
Methodology &
Confidence report. All system-specific numbers are illustrative until replaced by a real evaluation.
How many accounts would be flagged in Sweden?
It depends on the setting. Flagging the top ~0.05% of adult Facebook accounts (~2,900) would
surface roughly
570 genuine cases and about
2,330 false alarms for human review — catching
about one in five of the estimated real cases. Tighter settings are more accurate but narrower. See the
Sweden estimate.
How many people in Sweden have a sexual interest in children?
Depending on definition, on the order of
40,000–210,000 men; the commonly cited ~1%
clinical base rate gives about
42,000. Far fewer offend, and only a few thousand are actively,
detectably offending online at any one time. Full breakdown and sources in the prevalence brief
(
EN ·
SV ·
FI ·
DE).
Does immigration change that number?
No. Pedophilia is a sexual age-orientation found across all populations at broadly similar rates;
there is no credible evidence it is several-fold more common in any ethnic or religious group. Child
marriage is a different phenomenon (post-pubertal adolescents, driven by law and custom), and the "legal
from 9" premise is inaccurate. The model therefore uses one population base rate and flags behaviour,
not origin — weighting by ethnicity would manufacture discriminatory false positives and breach GDPR
and the EU AI Act.
What happens to a flagged account?
A trained human reviews it. If the reviewer finds apparent exploitation, the case is escalated
through lawful channels — the NCMEC CyberTipline in the US, the IWF in the UK, EU trusted-flagger
routes — initiated by the human finding and fully logged. The score is never the basis for action on its own.
Are people ever named or exposed publicly?
No. Public naming, "watch-lists" and any kind of exposure are prohibited uses. A false
accusation in this domain can destroy an innocent person's life, which is exactly why the scored list stays
inside the safety team and every flag is reviewed by a human.
Are these numbers real measurements?
Population and usage figures are real and sourced (Statistics Sweden, DataReportal). The
detection rates (precision, recall, base rate) are illustrative — drawn from the reference model —
because no evaluation of a live system was available. They are placeholders to be replaced with measured
values.
What about privacy and the law?
The design is built around the GDPR (meaningful human review under Article 22, heightened
protection for children's data, data minimisation and full auditability) and treats the system as likely
high-risk under the EU AI Act. Fairness is audited across languages and groups.
Where do the figures come from?
Public statistics (Statistics Sweden), platform data (DataReportal/Meta), and peer-reviewed
research (Dombert 2016; Seto; the Dunkelfeld/Troubled Desire project; the 2025 Lancet meta-analysis) plus
official child-safety bodies (NCMEC, IWF, WeProtect). Each report carries its own source list.