9 ms·

This is a reasonable argument, but I'm not sure you're aware of allegations of this company altering evidence to better fit a police narrative: https://www.vice

• pseudo0 4y ago
  I agree that there is definitely moral hazard when it comes to after the fact "corrections", particularly if the analyst knows the outcome desired by police. The company's track record looks pretty poor in this respect. In general though, I think there are legitimate reasons for manual analysis. Any machine learning approach to a problem like this is going to have to balance false positives and false negatives, and there is necessarily going to be a somewhat arbitrary cutoff. Detecting bang-like noises in a large city is likely going to have a pretty conservative cutoff to avoid DDOSing the police with calls. For example say the company sets their cutoff for automatic reporting at 90% confidence, but when the police ask them to review a specific time period, it turns out there was a shot detected with 89% confidence, and manual analysis indicates a false negative. That is probably still useful information, and it would definitely be useful to include this data point with the correct classification in future iterations of the model.