Automatic marking of spoken English increasingly feeds audio or text straight into a large model rather than extracting hand-designed features first. That removes a bottleneck and removes a safeguard at the same time, because those features encoded what assessors believed should count.

The risk is shortcuts. A powerful model can learn highly non-linear mappings, and if some incidental property of the recording correlates with the grade, it will happily use that instead of the thing being assessed. The grader looks accurate on held-out data while measuring the wrong quantity.

In language assessment that matters more than in most applications. A shortcut latching onto accent, recording equipment or first-language background produces a system that is unfair in a specific and hard-to-detect way, and the person receiving the mark has no way to see it. Controlling that reliance is a fairness requirement, not only an accuracy one.