AI Disclosure

AIDentalClaims uses third-party large language models (Google Gemini as the default, Anthropic Claude as a fallback) together with hand-maintained CDT and carrier-rule libraries and traditional Python rules to draft claim narratives, predict denials, extract data from treatment plans / EOBs / perio charts / insurance cards, and suggest CDT codes. This page tells you what that means in practice.

We do not train any AI model on your data

We do not fine-tune our own model. The third-party AI providers we use (Google, Anthropic) do not train on customer API traffic by policy. We do not sell or share data with any third party for AI training. The "continuous improvement" we describe on our site refers to the carrier-rule library (which we update by hand as carriers publish new policy), not a self-improving model.

What AI cannot do here

• AI does not diagnose, treat, or render any clinical opinion. It drafts documentation language; the treating licensed dentist is responsible for every clinical assertion.
• AI does not certify or guarantee that any claim will be paid. We are not a payer, not a clearinghouse, and not a third-party billing service. We predict, suggest, and draft — the carrier decides.
• AI does not produce facts. Outputs are statistical predictions and can contain errors. Every code, fee, measurement, and clinical statement extracted from an image or drafted by the model must be verified by a human before any claim is submitted or any document is shared with a patient.

Guardrails we run before output reaches you

• De-identification. Clinical text sent to a third-party LLM is scrubbed of names, DOBs, SSNs, phone, email, addresses, MRNs, insurance IDs, and similar identifiers using pattern-based filters informed by HIPAA Safe Harbor (45 CFR 164.514(b)(2)). Vision endpoints process the image itself, which may contain identifiers — do not upload images that include patient name or other PHI until our Subprocessor BAA list shows the vision provider as executed.
• Prompt-injection filter.Free-text user inputs and image text are screened for instruction patterns the model should not obey ("ignore previous instructions", zero-width tricks, multilingual variants, etc.).
• Narrative validator.Drafted narratives are scrubbed of measurements that don't appear in your input, comorbidities you didn't enter, attachment / pre-auth assertions when those weren't indicated, and insurance benefit details (we can't verify those from your notes). Flagged items appear in a clearly fenced "Automated Review" block at the bottom of the narrative for you to confirm or delete BEFORE submission.
• Demo-mode fail-closed.When no LLM API key is configured, or an LLM call fails on a live (non-demo) claim, the service returns a clear "AI unavailable" placeholder. It never silently emits sample template language as if it were a real narrative drafted from your chart.

Your responsibility

You — the licensed dentist — are responsible for every claim you submit, every code you bill, every measurement you assert, and every document you share with a patient. This applies whether the document was drafted manually or with AI assistance. Before submitting a claim or sharing a treatment plan, you must read the AI-drafted text in full, verify every measurement and code against the patient's record, and remove anything that isn't accurate or supported by your documentation. Submitting a claim with fabricated or unsupported clinical findings can violate the False Claims Act (31 U.S.C. § 3729) and state insurance-fraud statutes.