Can AI steward sim racing incidents? What it gets right, and what it really doesn't

We build a tool that uses AI to help review racing incidents, so you might expect us to tell you it can replace human stewards. It can't. Here's a more honest answer, because the details matter to anyone thinking of using AI for something that affects other people's race results.

What AI is good at

Describing what's on screen. Modern models are good at watching a clip and saying what happened: which cars, which corner, who was ahead, where the contact was. This is useful because describing is where stewards spend a lot of time.

Applying a written rule consistently. Given a ruleset, a model won't get tired, won't favour friends, and won't apply the rule differently in round six than in round two. Consistency is a real strength, and a big part of why leagues use it. See how to run consistent decisions.

Structuring the output. The same headings every time (what happened, which rule, how confident) means two stewards start from the same summary.

Speed. A first draft in a minute instead of twenty is a real change for a team of volunteers.

What it struggles with

This is the part that rarely gets said out loud.

Depth and distance. A flat video doesn't tell you how far apart two cars really were. Whether a car was alongside by "half a car length" or "a quarter" is often a pixel-level judgement from an angle that wasn't designed to show it. Models can be confidently wrong here.

Camera angle. A cockpit view hides the car behind. A chase cam hides what's beside you. A broadcast view compresses distance. A single clip is often one view of a three-dimensional event.

Intent. Was that a late lunge or a misjudged braking point? Was a car's move a deliberate block or a reaction? Footage shows what happened, not what the driver meant. Humans misjudge this too, but at least they've felt the car.

What happened before the clip. Who had the position at the start of the braking zone? Was one driver already on a defensive line? If the clip is trimmed too tight, the answer isn't in the footage at all.

Netcode. As we covered in netcode and incidents, the clip shows one driver's version of events. The model has no way of knowing it was delayed or predicted.

Overconfidence. The worst failure isn't a wrong answer, it's a wrong answer that sounds sure. Models produce fluent, plausible reasoning whether or not they've really got the information. That's why we show a confidence signal, and why low confidence is a reason to look again, not a reason to trust a penalty.

Context outside the footage. League history, previous warnings, whether the driver is new, a known lag issue. A model only knows what it's given.

What goes wrong in practice

The risky pattern isn't the AI making a bad call. It's people trusting the output because it's well written. A tidy paragraph with a rule number reads as authoritative, even when the evidence is thin.

A few ways leagues get into trouble:

  • Applying penalties straight from an AI verdict without anyone watching the clip
  • Not telling drivers that AI was involved
  • Treating high confidence as certainty
  • Using AI to settle a conflict where the real problem is that the clips don't match

How to use it safely

  1. The AI drafts, a human decides. Always. This is also how we describe our product.
  2. Watch the clip yourself for any penalty. Read the report as a guide to what to look for, not as a replacement for looking.
  3. Give it more than one angle if you have them. More views reduce the biggest weakness.
  4. Include context. The ruleset, the corner, what happened beforehand.
  5. Treat low confidence as a signal. It often means the clip really is ambiguous.
  6. Tell your drivers. People are more relaxed about AI assistance when you're open about it and when there's always a human making the final decision and an appeal.
  7. Check it against your own record. If the AI keeps leaning one way on a type of incident, find out why.

So, can AI steward?

Not on its own, not yet, and probably not in the way people fear. What it can do is make human stewards faster, more consistent and less burnt out, and that matters. Most of the harm in leagues comes from slow, inconsistent and tired decisions, and that's the part it improves.

If you want to see how that works in practice, you can try Sim Racing Steward free. Look at the confidence and the evidence it cites, and decide for yourself how far to trust it.