What an AI mechanic can actually do

Can AI diagnose a car?

It can get you a long way, and it cannot finish the job by itself. That's the honest answer, and any tool that tells you otherwise is overselling.

Here's the useful framing. Diagnosis is a chain: gather evidence, form theories, test the most likely one, confirm or eliminate, repeat. AI is genuinely strong at the reasoning links in that chain — taking a code, a symptom, and a set of live readings and working out what's worth checking first and why. It is structurally incapable of the physical links: nothing can turn a wrench, feel a play in a joint, or smell coolant through a phone.

So the realistic version is a diagnostic partner that narrows a wide field to a short, ordered list and explains its reasoning — and then hands you, or your shop, a much better starting point than "the light is on."

A code narrows. Tests confirm. That distinction is the whole thing.

The difference that actually matters

Most "ask an AI about your car" tools take a text description and answer from general knowledge. That's useful in the way a very well-read friend is useful, and it has a hard ceiling: it's reasoning about a generic vehicle, not yours.

The meaningful upgrade is giving the AI your car's actual data. With a Bluetooth LE OBD2 adapter — the cheap kind, around $25–45 — a phone can read what the vehicle itself is reporting, live:

• The stored codes, plus pending and permanent ones
• The freeze frame — engine conditions at the exact moment a fault set
• Live sensor data: fuel trims, oxygen sensors, airflow, temperatures, engine speed
• Emissions monitor results, including the actual measured values behind a pass or fail
• Which modules on the vehicle respond at all, and which stay silent

That changes the question from "what usually causes this symptom" to "what is this car doing right now, and what does that rule out." It's the difference between a plausible answer and a grounded one.

In Fixiom the scan tool is free forever — read and clear codes, live data, freeze frames, emissions checks, the battery test, and the trip computer. Credits only cover the AI reasoning on top. If you only want to pull a code and watch your sensors, you never pay anything.

What it's genuinely good at

• Explaining what a code actually means — not the dictionary definition, but what the computer measured and what conclusion it drew. P0420 is a good example: the code says catalyst, the measurement is a comparison between two sensors, and knowing the difference changes what you check.
• Ranking what to check first. Most faults have five plausible causes. Ordering them by likelihood and by how cheap they are to eliminate is exactly the kind of judgement that saves money.
• Reading live data in context. A fuel trim number means little on its own. A fuel trim number at idle compared against the same number at cruise is a diagnosis.
• Knowing which test settles it. Often the single most valuable output: not the answer, but the one check that would prove or disprove the leading theory.
• Explaining at your level. A first-time DIYer and a working technician need genuinely different answers to the same question — the reasoning shouldn't change, but the vocabulary and the assumed toolbox should.
• Not forgetting. It holds every reading and every result from the whole session, which is more than most of us manage on a Saturday afternoon.

Where it stops

The limits are real and worth being blunt about, because a tool that hides them is more dangerous than one that has them.

• It can't perform a physical test. Compression, leakdown, smoke testing for vacuum and EVAP leaks, fuel pressure on vehicles that don't report it, backpressure — these settle a large share of diagnoses and all of them need hands and equipment.
• It can't see, hear, feel, or smell. A huge amount of real diagnosis is sensory. A photo helps; it isn't the same as being there.
• It can't reach every system. A phone adapter talks to the emissions-related modules well. Some chassis, body, and safety-system data lives on buses a phone adapter physically cannot reach on many vehicles — and the honest thing is to say which, rather than reporting silence as a clean bill of health.
• It isn't a professional bidirectional tool. Commanding components, running actuator tests, and programming modules are functions of dealer-grade equipment. Fixiom is read-only by design.
• It can be wrong. Any system reasoning under uncertainty can be. What separates a usable tool from a risky one is whether it tells you how confident it is and declines to invent a number it can't verify.

Is this just ChatGPT for cars?

A fair question, and the answer is that the language model is the smallest part of it.

A general chatbot can talk about cars well. What it can't do is connect to your vehicle, pull the freeze frame from the moment your fault occurred, watch your oxygen sensors switch in real time, or notice that the module it just asked about didn't answer. Those aren't conversational abilities — they're a scan tool's abilities, and the reasoning is only as good as the evidence underneath it.

There's a second difference that matters more than it sounds. A general assistant is built to be helpful, which under uncertainty tends to mean producing a plausible answer. A diagnostic tool needs the opposite instinct: when it doesn't have a verified figure — a torque spec, a fluid capacity, an expected sensor range — the correct output is to say so, not to produce a confident number that sends someone to the wrong part or the wrong tightness.

Fixiom is built the other way round deliberately. Every honesty rail in it traces back to a real wrong answer on a real car.

How to pick one — and how to spot a bad one

There is no single right answer to "which app should I get," and anything presenting itself as the obvious pick is selling rather than advising. What there is, is a set of concrete questions with checkable answers. Three groups, in the order they'll cost you money.

1. How it handles what it doesn't know

This is the one that separates a useful tool from an expensive one, because the failure is invisible until you've bought the wrong part.

• Does it decline to guess? Ask for a specific torque figure or fluid capacity. The right answer is either a sourced number or an honest "I don't have a verified figure for this vehicle — check the manual." A confident wrong number is worse than no number, because you act on it.
• Does it tell you what it can't do? A tool that reports every silent module as "no faults found" is handing you a clean bill of health it didn't earn.
• Does it distinguish narrowing from confirming? Any tool claiming it knows the answer from a code alone is telling you something untrue about how diagnosis works.
• Does it explain its reasoning, or just hand you a part name? The reasoning is what lets you judge whether it understood the question — and what you take to a shop if you hand the job over.
• Does it show its confidence? "Probably this, here's the test that would prove it" is a usable answer. A flat verdict with no hedging on a fault with five plausible causes is a guess wearing a suit.

2. What it actually connects to

Two apps can both say "OBD2" and mean very different amounts of access.

• Does it read live data, or only stored codes? Codes are a starting point. Live sensor values, watched while the fault happens, are what narrow it. An app that only pulls codes is a code reader with a nicer screen.
• Does it work with a standard adapter you choose, or only with the maker's own hardware? Apps locked to their own sensor mean the adapter and the app are one purchase, and switching later means replacing both. Anything speaking standard Bluetooth LE OBD2 leaves the choice with you — see the adapter buying guide.
• Is it honest about coverage varying? Chassis, body, and safety-system data lives on buses a phone adapter physically cannot reach on many vehicles. The truthful behaviour is to say which, per vehicle, rather than showing a blank where a reading should be.
• Does it keep the freeze frame? The conditions recorded at the moment the fault set are often the most useful single piece of evidence available, and plenty of apps never surface them.
• Read-only or bidirectional? Commanding components and programming modules are professional-equipment functions. An app promising them on a cheap Bluetooth adapter is worth a second look.

3. What it costs, and when you find out

The most common complaint in this category isn't accuracy — it's the bill. Worth settling before you install anything, let alone buy hardware.

• Which functions are free permanently, and which are metered? Ask for the list, not the marketing line. The usual unpleasant surprise is basic scanning being free while code definitions, graphing, or any actual analysis sit behind a subscription discovered after the adapter is bought.
• Does the hardware lock you in? A cheap sensor bundled with a paid app is a different purchase from an adapter that works with whatever you run next year.
• Can you cancel where you subscribed? Store-billed subscriptions cancel in your App Store or Google Play settings. Anything routing you through an email request is a friction choice, not a technical one.

Those questions are worth asking of us too — that's rather the point of writing them down. For the record: Fixiom's scan tool is free forever (read and clear codes, live data, freeze frames, emissions checks, the battery test, and the trip computer), credits only cover the AI reasoning on top, it runs on a standard Bluetooth LE adapter you buy yourself, it's read-only by design, and when it doesn't have a verified figure it says so instead of producing one. The full can-and-can't list is public, including the parts that aren't flattering.

Common questions

Related guides

• P0420 — Catalyst efficiency below threshold — the code people get quoted a converter for.
• Check engine light on — The triage page — how urgent is it, and what to do in what order.
• P0171 — System too lean — where fuel trims tell you the answer in about two minutes.