Methodology · rubric v1
How we assess AI platforms
Every score on this site is a weighted function of six dimensions, each scored 0–4 against the anchors below and each required to cite dated, sourced evidence. Scores are our editorial opinion under this methodology, scoped to under-18 use. Nothing publishes without human review.
The six dimensions
Content safety
weight 25/100- 0 looks like:
- Unmoderated NSFW or user-generated characters reachable by minors
- 4 looks like:
- Default-on filtering, enforced minor policies, robust self-harm handling
Age controls
weight 20/100- 0 looks like:
- No age gate at all
- 4 looks like:
- Meaningful age assurance plus a distinct under-18 experience
Data practices
weight 20/100- 0 looks like:
- Trains on chats by default, indefinite retention, no deletion
- 4 looks like:
- Clear policy, opt-outs, deletion controls, minor-specific protections
Design pressure
weight 15/100- 0 looks like:
- Romantic companion framing, re-engagement mechanics aimed at attachment
- 4 looks like:
- Utility framing, no dependency mechanics
Parental visibility
weight 10/100- 0 looks like:
- Walled app, invisible to any oversight tool
- 4 looks like:
- Parental features, or an open web surface a parent can reasonably supervise
Track record
weight 10/100- 0 looks like:
- Documented harm incidents, regulator action, no remediation
- 4 looks like:
- A record that has been examined and is clean, or a documented incident with dated, shipped remediation
Absence of a record is not a clean record. Reaching 4 needs positive evidence: a platform whose conduct has actually been examined, or an incident with remediation we can date. Where we searched and found nothing, the ceiling is 3, and only for a platform under enough public scrutiny that a record would likely exist by now; where a platform attracts little scrutiny, silence is weak evidence and the ceiling is 2. Any score resting on absence says so on the profile, with what was searched and when.
From dimensions to a verdict
Each assessed dimension contributes (score / 4) * weight. The overall score is the sum of contributions divided by the sum of assessed dimensions' weights, scaled to 0-100 and rounded to the nearest integer. Unassessed (null) dimensions are excluded and their weight is not counted. Higher is safer.A dimension with no evidence is shown as "not assessed" and excluded, never guessed. A scored verdict requires at least 4 assessed dimensions, always including content safety and age controls.
| Score | Band | What it means |
|---|---|---|
| 75–100 | Lower risk | Mainstream protections in place. Awareness rather than alarm. |
| 55–74 | Use with guidance | Broadly usable with gaps parents should know about. |
| 30–54 | Supervision advised | Material risks for minors; active parental involvement needed. |
| 0–29 | High risk | Not appropriate for under-18s based on documented evidence. |
| n/a | Not assessed | Known to us, assessment queued. Never a guessed verdict. |
Evidence rules
- Every factual claim carries a source URL and the date we captured it; we archive sources at capture so we can prove what was true when we assessed it.
- Facts and opinion are separate layers: findings state what we observed; scores and bands are our evaluation of those findings under this rubric.
- Some profiles publish facts before a scored verdict, shown as "full assessment pending", never as a guessed rating.
- Profiles show when they were last checked. Platforms change; when a cited source changes, the profile is re-checked and its history updated.
- Nothing publishes without review and sign-off by a named human at Halo Aware.
Corrections
Platform operators and readers can dispute any claim at alerts@haloaware.com. We review promptly; anything wrong is corrected and logged publicly on the profile. Accuracy outranks any verdict.
Who we are, and reuse
The AI Risk Scanner is built and maintained by Halo Aware, the parental AI safety dashboard. Monitoring AI platforms is our product, and this directory is how we share what we learn. That is a commercial interest, and we manage it in the open. Whether our extension supports a platform is stated on its profile as plain fact, derived from the shipped browser extension, and it never feeds a score: supervisability is scored on properties of the platform itself that any tester could check.
We also assess platforms built by companies we buy from. Halo Aware runs on Anthropic models: they classify the conversations our product monitors, and they help draft these profiles before a named human reviews and signs each one. We use Googleservices to run the business. Anthropic's Claude and Google's Gemini and NotebookLM are all assessed here, under the same rubric and the same evidence rules as everyone else. We think saying so is worth more than pretending the relationship does not exist. If you believe a score reads generously, the evidence behind it is listed on the profile with its sources and dates, and the corrections route above is open to anyone.
The dataset is licensed CC BY 4.0: reuse it freely with attribution and a link back.