Ghost is an AI-powered privacy agent. We use machine-learning models to find your exposed personal data, draft removal requests on your behalf, and prioritise the risks that matter. Because AI sits close to legal action and to highly sensitive personal data, we hold ourselves to a stricter standard than the industry norm. This page sets out the rules we live by, in language a non-technical reader can check us against.
Our principles
- Privacy as default. We minimise the data sent to AI providers, redact what we can before sending, and never use customer data to train any model. Every third-party model provider we use is bound by a no-training, no-retention contractual clause.
- Transparency. When you read an AI-generated output inside Ghost, it is labelled. When an AI took an action on your behalf, the audit log records which model, which version, and what inputs it saw. You can ask us to surface that record at any time from your dashboard or via privacy@useghost.me.
- Human-in-the-loop on consequential actions. AI suggestions never auto-submit legal removal requests, never send messages on your behalf, never close a case, and never make a billing decision. A human (you, or a Ghost analyst named in the audit trail) approves every consequential action.
- Bias and fairness review. We audit AI outputs quarterly for differential performance across demographic groups — on name-matching, risk scoring, and removal-letter quality — and publish the methodology in our annual transparency report. Where performance is uneven we down-rank or retrain the affected components rather than ship them.
- Accountability. Every AI system at Ghost has a named human owner who is responsible for its safety, accuracy, and retirement schedule. There is no anonymous, autonomous "the AI".
- Right to redress. If an AI output affected you and you believe it was wrong, you can request a human review at no cost. We respond within five business days and document the outcome.
How we evaluate models before shipping
- Capability evaluation on a frozen test set of historical removal cases, with held-out brokers, jurisdictions, and identity patterns to detect overfitting.
- Safety evaluation covering prompt-injection, personal-data exfiltration, refusal-to-defame, and refusal to act against any person who has not consented.
- Bias evaluation across a representative sample of name structures, address formats, and pronouns drawn from public census-style data — never from customer records.
- Red-team review by an internal adversary team attempting to misuse the model for surveillance, harassment, or re-identification. A failing review blocks release.
- Shadow deployment against live cases for at least two weeks before any model becomes the primary actor on a consequential surface.
Acceptable AI use by customers
You may not use Ghost AI to investigate, stalk, dox, profile, or produce synthetic media of individuals other than yourself, except under a Ghost for Business contract with documented consent from every enrolled person. The full set of restrictions is in our Acceptable Use Policy; the AI-specific rules are a subset of it. Violations are reviewed by a human safety team and may result in suspension, termination, and (for severe cases) reporting to law enforcement.
Data flow to and from AI providers
When a request is routed to a third-party model provider, Ghost sends the minimum prompt necessary to complete the task and never sends raw identifiers where a pseudonym will do. Each contracted provider is on our sub-processor list, processes data only on our instructions, retains no copies beyond the request lifetime (subject to short-term abuse-prevention windows mandated by the provider), and is forbidden from using the data for any training or evaluation purpose. The current provider list and the safeguards in place are in our Privacy Policy.
Reporting AI safety issues
If Ghost AI produces output that you believe is unsafe, inaccurate, biased, or that violates this policy, write to ai-safety@useghost.me. We acknowledge every report within 48 hours, complete a first investigation within five business days, and publish a quarterly summary of issues received and remediation actions taken.