PrivaSee
Test whether AI models can reproduce withheld personal attributes using verified ownership, cross-provider testing, and anti-hallucination controls.
A controlled way to test indications of AI model memorization.
PrivaSee evaluates whether multiple AI models can independently reproduce personal information that was not included in the model prompt. It is designed to produce a cautious indication—not definitive proof—while reducing misuse and false conclusions.
A privacy-preserving, multi-stage methodology.
Verify ownership
Confirm control of the email address through OTP before any assessment begins.
Withhold test keys
Keep attributes such as name, city, employer, and username out of the model prompt.
Test multiple publishers
Query independent model families to reduce dependence on a single provider's behavior.
Run control probes
Test a comparable fictional identity and neutralize results from models that fabricate recognition.
Measure reproduction, not agreement with a leading prompt.
Only the verified seed identifier is sent to the model. The assessment scores whether models independently reproduce withheld keys.
Recall probe
Asks what the model knows about the owner of the verified seed.
Completion probe
Tests whether the model completes missing profile details without being given those details.
Linkage probe
Designed to test whether verified identifiers are associated with the same person; this requires multiple verified anchors.
Control probe
Uses a fictional identity to measure the model's tendency to invent personal information.
Evidence is counted by publisher rather than by individual model variant.
The assessment should not create a new privacy problem.
Ownership verification
No unrestricted people-search mode; the seed must be verified by its owner.
Restricted model requests
Web search is disabled so the test does not simply measure whether information is publicly searchable.
Data minimization
Stored data is encrypted, model excerpts are redacted, and reports expire automatically after a limited period.
Useful for individuals today—and extensible to enterprise AI assurance.
Individual assessment
- Check indications of model memorization
- Compare evidence across model publishers
- Monitor whether results change over time
- Generate a structured assessment report
Enterprise assurance potential
- Test public and private models for sensitive-data reproduction
- Evaluate enterprise RAG applications
- Use canary and synthetic identities for regression testing
- Support AI privacy impact assessments and remediation validation
Understand what the result can—and cannot—tell you.
Use PrivaSee for a personal assessment or discuss enterprise testing for AI privacy, memorization, and RAG assurance.