Hacker News Reader

Comment by rayiner | Hacker News Reader

The National Academy of Sciences report mentioned in the article is pretty damning about the whole field of "forensic science." https://www.ncjrs.gov/pdffiles1/nij/grants/228091.pdf.

Pretty much only DNA analysis rises above shamanism,[1] and juries and judges have incorrect perceptions about the accuracy of those tests. While the often-quoted false random match probability is astronomically low, that error rate is dominated by lab error, which can rise to 1-2%.

As an aside, the Innocence Project has a better description of the details of what happened: http://www.innocenceproject.org/cases-false-imprisonment/kir...

View on HN

The biggest scandals are fingerprint analysis and eyewitness testimony. The New Yorker had a big article on fingerprint analysis years ago and little has improved.

In essence, FBI and other "experts" have frequently attested to partial matches being reliable when they're pretty much subjective. Similarly, the FBI would claim matches based on far fewer points than other countries deem acceptable.

http://www.newyorker.com/magazine/2002/05/27/do-fingerprints...

Meanwhile, the unreliability of eyewitness testimony is legion, and it falsely convicts more people than anything else.

http://www.scientificamerican.com/article/do-the-eyes-have-i...

Forensic Science is basically a fiction, one which unfortunately has great power to ruin the lives of the people it touches.

The glorification and exaggeration of the CSI television franchise doesn't help the matter either.

Anything that Sir Arthur Conan Doyle glorified (Sherlock) is guaranteed to be soft-headed thinking—the man was infamous for it. Houdini had a long-standing "feud" with his good friend Doyle, because Doyle would popularize all of these crank theories & practices.

That report's really long. Is there any suggestion that, besides using validated techniques, the forensic investigators themselves be cross-validated -- where they don't know if they're getting cases from the "training set" or the "test set"? (And neither do the people they're directly in contact with.)

If it's justified for machine learning for ad systems, I fancy that sort of check is justified for justice, too.

Reply on news.ycombinator.com