Hacker News Reader

Comment by Isamu | Hacker News Reader

Here’s the crux:

>At a Jan. 9 hearing, McMonagle agreed with defense attorneys, saying the AI identification of Tolbert was akin to an anonymous informant, which is not enough to establish probable cause.

Face matching can be used but it is not admissible evidence in court, and so you can’t use it for a search warrant, you have to find other probable cause.

View on HN

Parallel construction is the term for using a covert, inadmissible method to get your suspect, then using that illicit information to build a legitimate case that you can present in court.

Perhaps the fact that they were unable to build that legitimate case means he is innocent. Or perhaps they just got too lazy with the wrong judge this time.

Not exactly. It is used to get to certain stage of an investigation, and before proceeding, other means are found to get to that same stage. In this case, parallel construction would have been done after AI identification but before getting the warrant. The parallel construction would/should have been used to justify the warrant and make the search legal. They didn’t do that, instead choosing to write the warrant in a way that was misleading at best, lying at worst, about justification for the search. Now the results of the search are being thrown out, which included finding the murder weapon in his house. So it doesn’t sound like he is likely innocent.

A gun. How do we know it was the murder weapon.

I think there's a difference from "parallel construction" and what could have happened in this case.

If they felt this guy was their suspect, there's nothing against using AI to determine that they should focus on that guy and look closer. They could instigate further, even (for example) run across someone who actually told them "yeah he said he shot a dude that day" and establish other facts and legally get a warrant.

That's parallel construction.

No it's not. Parallel construction is when you hear Joe has coke in his car, and choose to follow Joe, pull him over when he speeds, and develop PC for a search from that. The grounds for the search of the car are independent.

Here the AI has primed them to falsely recognize the guy. There's no independence the way there is with parallel construction.

Your example is not parallel construction either. If you hear that Joe has coke in his car you already have probable cause to search Joe's car. You don't need to wait until he speeds, since speeding wouldn't give you a reason to search his vehicle.

Parallel construction would be: you pulled Joe over for [whatever reason] and searched his car without probable cause (maybe he seemed shifty or he was being a dick) and found coke, so you subsequently went to an informant who told you that he witnessed Joe selling coke out of his car earlier on the day of the search (which would have given you probable cause to search his car if you had talked to the informant before pulling Joe over).

The basis of parallel construction is that a real investigation would have uncovered the evidence anyway. What many people don't get is that law enforcement actually has to do the legwork of a real investigation (but knowing what to look for, and where, etc.) to show that the excluded evidence should be made admissible. It's a remedy for law enforcement to cure procedural violations of due process under the Fourth Amendment.

I understood it to come in when the police don't want to admit intelligence from foreign sources in drug cases also.

Yes that's one of the common situations where parallel construction is required. But it still requires the police to do the parallel investigation so that they can establish that they could have come across that evidence in an admissible manner.

I don’t think this priming theory is workable. If some witness on the street gives the police a handful of names, and they discover she didn’t actually witness the crime, are they now prohibited from investigating those people because they’re “primed”?

Facial identification is particularly difficult to do right. If you lean on people at all they will pick the closest match rather than admit they don't know. Not from dishonesty, it's just that we're not that good at recognizing strangers.

It's only parallel construction if the original means of determining the suspect was unconstitutional.

Vs just "investigation"?

AI = anonymous informant. I see what the judge did there :)

He’s retired and I thought it might be accidental.

Which is kind of fair enough, isn't it? AI models are right more often than they're wrong, but they're not nearly infallible enough to be used in a court. So the output of them should be treated like an intelligence tip, i.e. a fact that probably points you in the right direction for gathering more solid evidence... but has a chance of pointing you in the wrong direction entirely. Which is exactly how most of us use LLMs right now.

> AI models are right more often than they're wrong

lol, what models are you using?

Mostly just ChatGPT 4o and o1. I feel like I get an 80% hit rate on software questions.

You would think that a real police officer would have reviewed all of the video themselves and checked the work of the facial recognition system. The software is just there to help narrow down what the officer needs to manually review. Like doing a web search and manually checking the top results to see if they are what you want.

Reply on news.ycombinator.com