Legal & transparency

How Fixiom uses AI

Fixiom is AI-powered, and we'd rather you know exactly what that means — what generates the answers, what keeps them honest, and where the limits are.

Effective date: July 21, 2026

What powers it

Fixiom's reasoning runs on a frontier large language model accessed through Anthropic's Claude API, combined with live web search, government vehicle-safety data (NHTSA), retrieved real-world documents, and — when you connect an OBD2 adapter — measurements from your own vehicle. The AI never runs on your phone; requests are processed on our servers and the provider's API, then discarded from the reply path (see Privacy).

The guardrails

A general-purpose chatbot will answer any car question — confidently, and sometimes wrongly. Fixiom's diagnostic brain runs inside a set of engineered guardrails designed for the failure modes that cost you money in this trade:

Specs are cross-checked, not recalled. Before stating a factory number (capacities, torque, fluids, pressures), Fixiom checks its own knowledge against live sources and attributes the figure to your exact engine and generation. Agreement → stated. Disagreement → you get the range and the reason. Unverifiable → it says so and points you to the OEM source.
Diagrams are real documents only. Fuse legends and wiring/component diagrams are retrieved from real sources and shown only when they can be verified for your vehicle's generation — with the source visible. Fixiom never generates a diagram image with AI, because a plausible-looking fake is worse than nothing.
Bulletins and recalls come from the source. TSB and recall checks run server-side against NHTSA's own data for your exact year, make, and model. The AI cannot "remember" a bulletin into existence, and a bulletin you cite gets verified before Fixiom agrees or disagrees.
A code that doesn't fit gets called out. A turbo code on a non-turbo engine (or similar configuration mismatches) is flagged rather than explained away.
Honest confidence, enforced. Diagnoses carry an explicit confidence level, and the system is capped from claiming a diagnosis is confirmed until a confirming test or your own verification supports it. When evidence is thin, it says so and proposes the test that would settle it.
No guessed part numbers or component identities. If the exact part for your vehicle can't be verified, Fixiom declines to guess and routes you to the right catalog or a visual check.
Hardware honesty. Generic OBD2 has real limits (no factory-locked ABS/airbag codes, no bidirectional tests). Fixiom states them instead of pretending, and won't claim to have read something it didn't.
Evidence over vibes. With an adapter connected, the AI prefers to measure — setting live watch conditions and updating its assessment on captured data rather than assumption.

What gets sent to the AI

Your messages, the vehicle's identity (year/make/model/engine, or VIN if provided), trouble codes and summarized OBD2 readings, and any photos you choose to attach. This content is processed under commercial API terms and is not used to train AI models. We don't send your name or contact details — we don't have them.

The limits — read these

AI can still be wrong. The guardrails reduce wrong answers; they cannot eliminate them. Treat output as a well-reasoned starting point, verify specifications against manufacturer service information, and apply your own judgment before acting.
It is not professional advice. Fixiom is an informational tool, not a certified mechanic, and no output creates a professional-client relationship.
Safety-critical work needs professionals. Brakes, steering, airbags/SRS, fuel systems, and high-voltage components can injure or kill. Fixiom flags these situations, and you should take the flag seriously.
Coverage varies. Vehicle data quality differs by make, model, year, and market; the app is upfront when your vehicle's data is thin.

Report a bad answer

A confidently wrong answer is the exact thing we engineer against — if you catch one, we genuinely want it: hello@fixiom.ai. Include the vehicle and roughly what was asked.