> human in the loop ML can be extremely useful

Human-in-the-loop ML can be extremely useful in avoiding obvious mistakes.

Human-in-the-loop ML can be extremely useful in fabricating desired results.

Not quite sure what point you're trying to make here...

Misusing science is a problem regardless of whether ML is the science in question. It's /already/ a big problem within criminal justice (polygraphs, psychopath tests, etc).

Avoiding 'obvious' mistakes is helpful in reducing the amount of work human raters/evaluators need to do. Automatically classifying 90% of data and leaving 10% for humans can reduce human workloads by 90%, or enable new use-cases where doing 100% of the work simply wasn't feasible. Using ML in this way doesn't make it 'fake AI.'

> Most forensic science is junk.

I'm ignorant on the topic. Got links?

Forensics is full of gee-whiz-cool csi stuff which eventually is shown to be worthless. Polygraphs are a great example; they're still very wisely used in the US, despite piles of evidence showing they're prone to both false positives and false negatives. Sibling comments saying they aren't forensics because they've been discredited are missing the point: these are typical crime tech, and have been lately discredited, but are still in use in the US regardless.

https://en.m.wikipedia.org/wiki/Polygraph

For a more recent example, here's an article on denim analysis...

https://www.nytimes.com/2020/04/08/science/denim-pattern-mat...

The underlying problems seem to be bad incentives (the prosecution is paid+evaluated on its ability to put people in jail) and the court's willingness to accept insufficiently validated tech.

To add to that list: fire investigation. There have historically been a lot of myths in the field regarding forms of "proof" that a fire was caused by arson, and many of them were historically used to obtain (likely false) convictions.

https://www.abajournal.com/magazine/article/long_held_belief...