Tracker Methodology
How we analyze, score, and verify the privacy policies of AI platforms.
Grounded in primary evidence
PrivacyGPT is built on three core pillars: falsifiability, visible citations, and strict neutrality. Unlike static review blogs, we do not guess or rely on hearsay. Every data point in our tracker is mapped to a specific clause in the company’s publicly available legal terms and cited with a clickable direct URL.
Core assessment fields
The questions every entry answers, and how each answer is graded.
Model Training by Default
Does the platform use user conversations to train or fine-tune its models by default?
- Opt-in Default: Data is private until user consents (highest rating).
- Opt-out Default: Platform trains by default, but user can disable it.
- No Opt-out: Platform trains by default with no settings to disable.
Opt-Out Ease
How difficult is it for an average consumer to stop the company from training on their data?
- Excellent: A single toggle in settings without losing chat history (e.g. Claude).
- Moderate: A settings toggle but history is disabled (e.g. Google), or a dedicated form.
- Poor: Requires emailing support, submitting a manual ticket, or is completely unavailable.
Human Review of Chats
Are conversations read by human annotators or trust/safety moderators?
- Human review represents a distinct threat vector for data exposure (chats are cached, viewed by third-party contractors).
- Opt-out configuration: Can users request exemptions from human review, or is it mandatory for safety monitoring?
Confidence & verification ratings
To remain honest, we label the verification source and confidence levels for every entity in the tracker.
Verified
Explicitly stated in the current public privacy policy, terms of service, or official help articles with direct link.
Inferred
Logically deduced from parent company policies, technical docs, or developer platforms, but not explicitly stated in consumer policies.
Needs Review
Contradictory information exists, policies have changed recently without official clarification, or data requires active human verification.
Transparency rubric & point system
We compute a normalized score (0 to 100) based on weighted category scores.
Default weights
| Category | Default Weight |
|---|---|
| Default Model Training Disabled | 30% |
| Ease of Opt-Out | 20% |
| Data Retention Period | 15% |
| Deletion Rights (chats & account) | 15% |
| No Human Review of normal chats | 10% |
| Third Party Sharing restrictions | 10% |
Category scoring rules
1. Model Training by Default
- 100 ptsPrivate by default (no training on user chats).
- 0 ptsTrains by default (requires opt-out).
2. Opt-Out Ease
- 100 ptsSettings toggle or already private.
- 40 ptsRequires manual form submission or is restricted to EU/UK.
- 0 ptsNo opt-out available.
3. Data Retention Period
- 100 ptsZero retention (chats deleted immediately).
- 80 ptsShort retention (under 30 days, e.g., when history is off).
- 50 ptsMedium retention (30 days to 18 months).
- 0 ptsLong, indefinite, or 5-year research retention.
4. Deletion Rights
- 100 ptsFull account & thread deletion processed within 30 days.
- 50 ptsPartial deletion (e.g., chats are deleted but model-ingested training posts cannot be purged).
- 0 ptsNo deletion requests honored.
5. Third-Party Sharing
- 100 ptsStored in isolated customer VPC, zero third-party sharing.
- 85 ptsShared with infrastructure/model providers under strict non-training, security contracts.
- 20 ptsShared for targeted advertising or across general consumer platforms.
6. Human Review of Chats
- 100 ptsNo human review.
- 80 ptsHuman review is restricted to abuse/security flags or safety reports.
- 30 ptsChats are sampled and reviewed for model annotation/improvement.
Total Score = (Σ (Category Score * Category Weight)) / (Σ Weights)