AI facial recognition could sink this murder probe A murder case in Cleveland, Ohio, could collapse because the city's police relied on AI-based facial recognition software to obtain a search warrant. This despite a specific warning from the provider of that artificial intelligence – our old friends Clearview AI – that results from its facial recognition search software should …
I have had a case which involved bite mark evidence to put a suspect at the scene although I didn't make the comparison - a dental expert did that. I wouldn't call it fringe except insofar as there are few cases where it's applicable. I'd also limit it to situations where the bite marks are of a quality comparable with those from tool marks or ballistics. That would exclude bite injuries on a victim.
John Oliver did a piece on this, but basically: no. Bite mark analysis is extremely unreliable and very often suffers from both false positives and false negatives. It's allowed as evidence in court because it's only up to the judge if evidence is admissible or not, not actual forensic scientists.
Or in other words: bite mark analysis is as reliable as "psychic witnesses" (which some judges *also* allow).
Well that says it all, doesn't it ?
On top of that, who are these Keystone cops who take facial recognition from one video to accuse someone of a crime they had a record of on another video ?
They found a gun. Well duh, it's the USA. You can find a gun almost anywhere there. Would be a lot easier if guns were forbidden, eh ?
So, did the gun match the bullets ? Did it ? Because I believe we have actual reliable technology to determine that. And if it did, then screw this facial recog stuff, that guy is guilty.
But did the gun match the bullets ?
You believe it because that's what TV shows keep saying.
In the real world, no, that's not reliable either. And yet, because lots of people watch those TV shows as documentaries, it's indeed used as if it were.
https://www.scientificamerican.com/article/the-field-of-firearms-forensics-is-flawed/
Agree that the report is very interesting but it does not exclude a 'clear match' scenario where there is a sufficiently high confidence, it very rightly points out that such a match is a high bar to meet.
The lack of any such evidence in this case would suggest that they could not even reach the 'looks similar if you squint' level of match (presuming the seized weapon is is actually the same calibre as the fatal bullet).
Firearms identification wasn't my area* however I note that the tests were based on new firearms and the whole basis of marks examination is that an object which makes a mark picks up unique characteristics over the course of time so that an old object will leave a mark with more identifying characteristics than a new one.
It also concerns me that a lot of effort in these reports seems to have gone into the inconclusive responses. To my way of thinking there are only three possible responses to a comparison: they could match, they couldn't match and I can't draw a conclusion. The last shouldn't really affect the court's decision. Also, "could" and "couldn't" carry very different levels of confidence. If a particular blood sample couldn't have come from some particular person the court could reasonably decide it didn't. If it could that doesn't mean must have** and the court should take into consideration that it could also have come from many other people and should be provided with frequency figures to decide whether, in the particular circumstances of the case, that was likely.
But returning to this case - "thinking" the gun was responsible shouldn't be enough. Thinking isn't evidence and if that's all they were offering it suggests they didn't have evidence or even that they had evidence agianst them.
* Fortunately. One of the very few times I was asked to look at a firearm for some other purpose it turned out to have been booby-trapped. That was spotted by the colleague who examined it. A similarly booby-trapped forearm exploded and killed two police officers who'd opened it.
** I have experienced prosecution trying to drag "could have" on a hair sample into "must have". The FBI had a reputation for claiming great success in matching hair. I couldn't understand it because I never found it very conclusive. I later read that a proper investigation trashed that reputation. The only situation where I found hair examination could become conclusive was in the identification of three very badly burned bodies where a sample recovered from one was compared with samples from the hair brushes of the known victims and fortunately could only have come from one.
Nowadays you can get DNA from a hair. But DNA can be misleading too. DNA from the daughter of the Long Island killer was found on one of his victims, because the victim rode in the killers family car. Police confirmed wife and daughter were outside the country on vacation at the time of that murder.
May I point out from the linked article, A man was "shot in a bathtub" and "officers found the man dead from what was initially believed to be a suspected overdose"
If these Keystone Cops struggle with the difference between a shooting and an overdose, I certainly wouldn't trust them to nab the right man
As you say, they have just seen another video, with the first having being too dark/grainy to identify a face, and said "that must be him!" Before passing it to Clearview to charge a random bloke who happened to be wearing the same colour T-shirt.
And then even if the keystone cops were right in their hunch, the chance of Clearview finding the right personentry in their database is significantly less than 1, as Clearview clearly state in their documents
I'm amazed that the case made it to court, never mind that the police are appealing to keep their shoddy evidence
"So, did the gun match the bullets ? Did it ? Because I believe we have actual reliable technology to determine that. And if it did, then screw this facial recog stuff, that guy is guilty."
IANAL but my understanding of US rules of evidence is that if the search warrant was invalid, nothing they found under it is admissible - I believe they call it "fruit of the poison tree."
So no evidence derived from the gun recovered in an improperly issued search warrant can be used. I believe that there are nuances around it but as a rule of thumb "If they wouldn't have it without the improper search, it doesn't exist for the purposes of investigating or prosecuting the case"
That's why the screwup by the cops is so egregious, because even if the guy IS guilty, now no evidence recovered in the search, or anything derived from it, is usable at all and if they want to prove their case they have to do it without any of that. They'd have been better off not making the invalid search in the hope that they could gather a VALID lead to that evidence later, one that wouldn't poison the results of recovering it.
So in future the cops will use AI to find a suspect, then arrest and manufacture evidence instead and not mention the AI aspect saying to themselves that it's fine because they know he's guilty. Like they always have in some parts of the world.
The problem as always is not the tools, it's the fools using them.
<......"The problem as always is not the tools, it's the fools using them".......>
In this case it IS the tool that is the problem. Even the supplier as good as admits that it is of no use, so why do they make it available? (OK, I do know why - money,as usual!)
Not! I wouldn't imagine it is a bundle joy elsewhere either.
The accused has previously been a naughty boy if the online records of the Ohio Dept of Corrections and Rehabilitation were to be believed but on the evidence presented tainted or otherwise I wouldn't hang a dog.
What was it in the old Superman tv series ... Truth, Justice and the American Way? Of course he would now be an undocumented/illegal immigrant to be deported on the next flight to his natal Krypton.* :(
* although more likely Rwanda as I recall the planet was in a rather fragmentary state.
> Identifying the person in the supplied footage albeit not the footage of the actual murder.
Boggle!
It sounds like the cops had made up their mind, based on the suspect wearing clothes and being able to walk, then set about creating producing evidence to support their suspicion.
Still, it's not as bad as the town where the newly elected fire chief was revealed as being a convicted arsonist.
You do have to wonder just what is going on there
> newly elected fire chief was revealed as being a convicted arsonist. You do have to wonder just what is going on there
What is going on there, is good old-fashioned nepotism, apparently:
"Despite Simmons’ record, he’s got emergency services in his blood. His dad, Herb Simmons, was the long-time director of the St. Clair County Emergency Management Agency."
wait until you hear about the country that elected a felon and rapist, and insurrectionist traitor, unbelievable the level of corruption that these third world countries allow and have such a weak constitution and law system thats incapable of prosecuting rich people.
it will blow your mind
It sounds like the cops had made up their mind
Because police and prosecutors are incentivized to close cases, not to arrest and convict the right person. If they can find someone they think a jury will convict, many of them will happy put an innocent man behind bars. And there is no arguing with this fact, because there is an ever growing list of people being proven innocent years later and it is later revealed that the police and/or prosecutors knew they were innocent and withheld evidence of that from the judge/jury because they just wanted to stamp a "closed" on the case file. It is even more true in high profile cases where the police feel pressure to make a quick arrest.
The Central Park Five being a great example of this, even though a certain orange someone still refuses to accept the fact of their innocence.
I think the facial recognition did its job. They put the data in, it said I think it is this guy and then the cops started the investigation. This is just the same as a possible eye witness going through a book of pictures and saying the same. At least the AI doesn't think "I know this guy didn't do it but I never liked him so I'll get him investigated".
It is then up to the police what to do. If a witness picking a portrait out of a gallery is sufficient for a warrant, it should be the same for an AI. Neither should be sufficient evidence for a conviction and I believe the term used is beyond reasonable doubt. I'd hope at least some checking was done, would a person looking at the CCTV and the person think they are the same?
The main issue is the lack of resources to investigate every crime committed. If the options are to put up a wanted poster in the hopes a member of the public recognises them or use an AI to look through millions of images to see if they can identify them, I know what a resource strained force looking to make a quick conviction would do.
Perhaps it should show 5 best matches as suggestions and tell the police force to get up and do some real work? But I suspect this would only lead to everyone being treated as guilty and searched/monitored without sufficient evidence.
well, resources are an issue...as is the cooperation of the victim and neighbors of the victim (and perps).
But another big factor is also score keeping. Police like to talk about "arrest rates" for crimes -- an arrest (and later found "not guilty", or even "insufficient evidence for prosecution") is better than no arrest. Prosecutors like to talk about convictions -- a wrongful conviction is better than no conviction. A hung jury beats not going to trial. A plea bargain beats going to trial ("Admission of guilt! Sure, lesser charge, but we all know what he did"). It's easier to prosecute a crime victim that fights back than it is the criminal -- as the criminal generally had a plan going in, the victim made it up as it went along. The victim who shoots or stabs their attacker is an easy win on the score card for all parties (arrest, prosecution, conviction!).
I do believe there were even scandals in a few big US cities where crime rates were deliberately INFLATED to get more money to "fight crime".
And unfortunately, many of the people who claim "the system is stacked us!" don't even bother to show up to vote in local elections, where the prosecutors and people who manage the police departments are chosen.
I don't know what the answer is. But it is all related to people...
@AC "I think the facial recognition did its job. They put the data in, it said I think it is this guy and then the cops started the investigation. This is just the same as a possible eye witness going through a book of pictures and saying the same."
It did not do its job it made 7 different matches, the AI report turned up eight photos, 2 were Tolbert 6 weren't. Possible eye witness to shopping in a convenience store 6 day later, and saying the shopper matches 7 different people.
> I think the facial recognition did its job.
Based on what do you think that, when the company who makes the software says: 'that results from its facial recognition search software should not be "used as admissible evidence in a court of law or any court filing."'
> They put the data in, it said I think it is this guy and then the cops started the investigation.
The "data" being some other video showing a someone with similar cloths whos walk the cops felt was somewhat similar.
> At least the AI doesn't think "I know this guy didn't do it but I never liked him so I'll get him investigated".
I am sure that's a great consolation to the victims of false accusation /sarcasm
> But I suspect this would only lead to everyone being treated as guilty and searched/monitored without sufficient evidence.
So you do see the problem?
"it should be the same for an AI"
People have the ability to say "I don't recall" or "I don't know" or "jeez, they all look like scumbags". And there are probably penalties if you knowingly pick the wrong person.
AI, on the other hand, will say it's absolutely positively this dude (while pointing at a picture of Daffy Duck).
Worth making the point that the legal wrangling at hand here is not about the reliability of the AI/facial recognition software. It's about the prosecution's (apparent) failure to turn over relevant material (including the facial recognition support).
Whether any challenge to the quality of the facial recognition documentation happens pre trial or during trial, I have no idea.
https://www.criminalattorneycolumbus.com/what-is-discovery-in-criminal-law-in-ohio/#:~:text=In%20Ohio%2C%20discovery%20is%20governed%20by%20Criminal%20Rule%2016%20of,evidence%20and%20information%20before%20trial.
This kind of harks back to the overreliance on finger-prints that led to some pretty serious legal bothers a while back.
Fingerprinting was never as accurate as claimed in the early days, and its apparent accuracy was a result of high rates of recidivism: if your fingerprint looked like a match in the database, it probably was a match, because that person was likely to commit a repeat offence.
But then we started keeping fingerprints of people who were innocent on file, including witnesses who only gave their fingerprints to rule them out if their fingerprint came up in the current crime investigation.
Then when you look in the national database for a match, there was a much greater chance that the fingerprint that looked similar was not a career criminal, but someone who was once an innocent bystander, or worse: the victim of a crime. Imagine that: you've been beaten up by a drug dealer for simply being in the wrong place at the wrong time, and just as you're getting over that, the police start questioning you about a burglary because there was a print found that looked sort of similar to yours....
"But then we started keeping fingerprints of people who were innocent on file, including witnesses who only gave their fingerprints to rule them out if their fingerprint came up in the current crime investigation."
This could be a good thing. Instead of the CSI process of the computer stopping on the first match, the search could produce a "heat map" of close matches. Here's the suspect but then he's right in the middle of a dense cluster of other close matches.
Police and other people in the fingerprint biz have resisted this sort of examination in the past. Because it calls into question the validity of the entire process. But times are changing. All we need now are juries who understand statistics, error magnitudes, etc.
"Then when you look in the national database for a match, there was a much greater chance that the fingerprint that looked similar was not a career criminal, but someone who was once an innocent bystander, or worse: the victim of a crime."
You'd think a halfway competent defense attorney could handle this. "Yes. My client was at the scene of the crime. He was the one being robbed."
Anecdote: Many years ago, I applied for a concealed weapons permit at the local sheriffs department. The process included submitting a set of fingerprints for the requisite background check. In the days when this involved paper cards, nasty black ink and probably someone at FBI headquarters examining them with a jewelers loupe. I asked how long the permit process would take and was informed it would be about eight weeks. Two weeks later, the sheriffs clerk called me up to tell me that the check had been completed and the approved permit could be picked up at their office. Upon arriving, the clerk told me that she had never seen an application go through so fast. "I think the FBI keeps my file near the top of their in-basket." After a somewhat horrified look on her part, I added, "Seriously, I've been through a number of checks for security clearances and whatnot. They already know who I am."
Sometimes it's advantagious to have fingerprints and facial recognition flow through the system effortlessly.
Apple says that about one in a million people look similar enough to me that they can unlock my iPhone with faceID. That means if the police had a database of everyone’s faceID about 300 or so Americans would match the faceID on my phone. There are probably more people matching the photo they took (ignoring the little detail that there was no evidence that the killer was in the photo).
The technology is just fine. What’s wrong is the absolute lack of knowledge of how statistics works. If the is zero other evidence then the face match means nothing. It is a little bit of circumstantial evidence but not enough for a search warrant. It’s enough to check if the person lived nearby, and ask them where they have been, but a search warrant requires more.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
