A new technique to work out a corpse's time of death
economist.com
economist.com
> geoFOR: A collaborative forensic taphonomy database for estimating the postmortem interval
If it the output is wrong, innocent people go to jail, the guilty go free, and families are confused with misinformation about their loved one's death.
There is no discussion of the system's accuracy or an attempt to determine it empirically (possibly an omission by the article and not the developers). It does say,
> At present the model offers a range of days within which it is 80% confident the true pmi lies ...
Taken literally, that is the AI stating the AI's accuracy instead of an empirical test against known real-world outcomes. And look at the training data:
> About 1,800 of these are real-world cases involving the discovery of a body. The rest are drawn from forensic experiments at so-called “body farms” in Texas and Tennessee, in which corpses are left to decay for weeks and months under varied circumstances.
Yet elsewhere they say that real-world cases are highly unreliable:
> Working out when a person has died is the most basic (but frustratingly imprecise) part of a forensic investigator’s work. For decades these specialists have had to rely on intuition, combined with observations of the state of the deceased and clues such as temperature, both of the dead body and the environment. Different bodies decay at different rates, however, and individual circumstances can throw off the most careful pmi calculations. A body found in a ditch in northern England in 2004, for example, was given an erroneously late time of death because the ditch was shielded from sunlight and the colder-than-expected conditions had helped preserve the corpse. / Forensic-science journals are full of such cases ...
It would seem that you are training the computer to make the same errors with the same biases that the foresnic investigators already make - unless they have some objective source for the actual time of death in those real-world cases.
Hopefully, it's just a limitation of the article, not of the actual software.
https://undark.org/2019/11/11/how-microbes-could-aid-forensi...
Glad to see were going to start letting the justice system become dependent on black box stochastic parrots.
After all the government + technology is always right and incorruptible it's not like horrible things like hundreds of innocent government employees being imprisoned and prosecuted over errors in software ever happens, right? right?