A man spent a year in jail on murder charge that hinged on disputed AI evidence | Hacker News Reader

A man spent a year in jail on murder charge that hinged on disputed AI evidence

theregister.com

372 pointsBender204 comments

The article says that employees of the AI company (ShotSpotter) manually reviewed and classified the sounds as gunshots:

> records showed that ShotSpotter actually initially picked up what sounded like a firework a mile away, and this was later reclassified by ShotSpotter staff to be a gunshot at the intersection where and when Williams was seen on camera.

So the AI didn’t even make the call. The staff did, manually. I assume that means the actual audio is available and entered into evidence?

If humans are making the call then blaming AI seems like a stretch. That’s almost like blaming the motion detection algorithms for triggering video recordings that were later reviewed by humans. It’s still humans reviewing the recordings and making decisions.

It may not be right to blame AI, but it's certainly the case that they were presenting it as "the system says there was a shot fired here." The bigger point is that the "system" can't be trusted. And the fact that they need to manually classify things is proof of that fact. (It's way more complicated than just an audio recording. It's recordings from numerous locations, attempting to triangulate the original location, without being confused by echoes and other things.)

> It may not be right to blame AI, but it's certainly the case that they were presenting it as "the system says there was a shot fired here."

I think a crucial distinction is to clarify who “they” is: Who was making the claim? From the article, it seems that the prosecutors may have been the ones exaggerating the evidence:

> Prosecutors said ShotSpotter picked up a gunshot sound where Williams was seen on surveillance camera footage in his car, putting it all forward as proof that Williams shot Herring right there and then.

Naturally the defense wanted to see the actual evidence, so that’s what they got down to.

I know everyone wants to vilify AI, but I think the actual bad actor in this article might be the prosecutor trying to misrepresent the evidence.

>I think a crucial distinction is to clarify who “they” is: Who was making the claim? From the article, it seems that the prosecutors may have been the ones exaggerating the evidence:

The prosecutors are the "they". The further removed you get from people who know the 'AI' the more the validity of the evidence provided by the AI relies merely on what others have said.

The people working on the system know that it can be faulty. Other software developers have an idea that it could be faulty. Prosecutors are told that it's alright. They turn around and argue it in a court of law. And the further you get the less people have any idea that the system is faulty. Down the line people will go "the machine says so, therefore it must be so".

>Down the line people will go "the machine says so, therefore it must be so".

Thats the real feature of AI, devision laundering: a flawed decision made by someone for their own gain now appears impartial and trustworthy, and come wothout accountability

Sure. But the point is that they misrepresent the evidence by pointing to the system as trustworthy. Even without tampering with the evidence the system isn’t trustworthy beyond a reasonable shadow of a doubt.

That's the point? These systems invariably turn out to not be any sort of AI at all, but sales people ready to make a call any way their customer (police) wants it made.

Later it's presented as AI to a jury. It's basically the same scam as lots of forensics, junk science to get a desired result.

I generally agree, but wouldn't necessarily say a lot of forensics is junk science. The science is generally sound (procedures and theory). Usually it's the prosecutors that present the forensics in a way that misrepresents what the results actually mean. Or lab techs make mistakes or perform misconduct (How to Fix a Drug Scandal on Netflix is interesting).

Fingerprinting is junk science. Polygraphs are junk science. Tons of things that have been or are admissible as evidence are junk science.

I wouldn't call polygraphs forensics. Most states do not allow them as evidence.

Do you have a link for the fingerprinting claim?

(Not who you responded to)

https://www.crammlawfirm.com/fingerprint-match-jails-innocen...

This isn't the case I originally recalled, and doesn't say how many points were matched, but I remember a 5 point match being used to convict a person who was later proven innocent.

That shows its flawed, not junk.

Oh, yeah. I've heard of them using too few points. I would have lumped that into misconduct or mistakes by the lab tech or prosecution.

Yeah, knowing or not, comparing with too few points is misconduct - that should be the fault of someone besides the defense, but that is somewhat tangential - does anyone with more knowledge know where the line is drawn? I imagine there are some outlandish coincedences happening.

Fingerprints aren't "junk science", but they do have error margins and plenty of room for mistakes. The problem is that many juries, judges, prosecutors, and even defenders forget about this.

This is the case for many types of evidence; witness statements for example can also be notoriously unreliable. It's not bad to use this kind of evidence, but you do have a problem if you consider any witness statement to be absolute truth and never consider any "reasonable doubt" there may be, and this is how you end up convicting people based purely on "I saw the suspect commit the crime from 50 metres away in the dark and I'm sure it was him!" or some such.

If you have six independent witness statements like that, or a witness statement and fingerprints then the reasonable doubt starts becoming a lot less reasonable.

This is also the problem with the case in this article: a piece of evidence over which it seems like reasonable doubt should exists is the singular piece of evidence tying this person to the crime. The problem here isn't the evidence as such IMHO, but that juries and/or defenders (depends on the case) seem exceedingly bad at sceptically examining physical evidence.

> Fingerprints aren't "junk science"

Fingerprints are directly-recorded facts. What may be junk science is what expert witnesses present about the interpretation of fingerprints.

> The problem is that many juries, judges, prosecutors, and even defenders forget about this.

Juries as triers of fact, and judges insofar as they are ruling on the facts of the case, aren't, in any case, supposed to allow themselves fo be guided by extrinsic knowledge, they are supposed to act based on facts (both direct and the expert knowledge relating to the import of direct facts) provided as evidence, both testimony and exhibits, in the case.

> Fingerprints are directly-recorded facts.

That's not really how it works; false positives exist, especially if you consider that a lot of times you have incomplete or smudged fingerprints and there is actually some amount of subjectivity involved in determining a match. See e.g. [1]: "only two properly designed studies of latent fingerprint analysis had been conducted. These both found the rate of false matches (known as “false positives”) to be very high: 1 in 18 and 1 in 30."

Again: useful evidence for sure, but considering them undisputable facts is not wise.

[1]: https://theconversation.com/fingerprinting-to-solve-crimes-n...

> That's not really how it works

Yes, it is.

> false positives exist

Matches (positive or negative) are not the same thing as fingerprints. Fingerprints are directly recorded facts. Matches are matters of interpretation (the fact of a match reported by a particular system is a distinct direct fact from the fingerprint itself, but again the significance is...what expert testimony exists to establish or challenge.)

The reality is that when "fingerprint evidence" is brought to court that it's not foolproof and has a sizeable margin of error that should be accounted for when considering the verdict, something that is currently often ignored. You can argue semantics all day, but that's the way things work right now. This seems like an exceedingly pedantic point to make.

> The reality is that when "fingerprint evidence" is brought to court that it's not foolproof

No evidence is foolproof. My entire point is that the issue isn't with “fingerprints” or the knowledge that juries and judges bring extrinsically (which was the explicit claim made upthread), but with the expert testimony that contextualizes fingerprints.

(This was even more clearly the problem with fiber evidence when the FBI crime lab was presenting pure bunk expert testimony in virtually every case.)

The "fact" of a fingerprint is useless without matching. Forensics is the process of interpresting and finding matches for facts.

You pedantism is wholey off base here.

Not fingerprinting, but related issue for microscopic hair comparisons:

The Justice Department and FBI have formally acknowledged that nearly every examiner in an elite FBI forensic unit gave flawed testimony in almost all trials in which they offered evidence against criminal defendants over more than a two-decade period before 2000.

Of 28 examiners with the FBI Laboratory’s microscopic hair comparison unit, 26 overstated forensic matches in ways that favored prosecutors in more than 95 percent of the 268 trials reviewed so far

https://www.washingtonpost.com/local/crime/fbi-overstated-fo...

> Fingerprinting is junk science.

How then do you explain the success of Touch ID on Apple devices and the equivalent systems on other phones and tablets?

Random people can't just walk up and unlock a Touch ID device with their fingerprint, which suggests that fingerprints are in fact a very good way to tell people apart.

Yes, it is possible to make a fake fingerprint that Touch ID will not be able to distinguish from my fingerprint. But that fake fingerprint will only unlock my device, not your device, so it is still correctly distinguishing our fingerprints.

Because other surfaces are not a purpose built sensor designed to detect the features of a finger then repeatedly trained with a specific set of finger scans?

Like are you joking?

Saying TouchID validates fingerprinting is like saying MRI machines validate psychics.

Are you joking?

It seems you two have a different idea of what is being discussed. The underlying fundamentals of finger prints are solid. The part that can make them ineffective is when they are using insufficiently tested tools or matching on too few points. And of course misrepresentating what a match actually means and how it pertains to the case.

> The underlying fundamentals of finger prints are solid.

What exactly are these underlying 'fundamentals'? Nobody even knows if everybody has unique fingerprints, it's just an assumption.

Beyond that, it's not all that relevant since as you mentioned how the matching is done is largely subjective, and it depends a lot on the number of "points" used and who does the matching, and that should really bring the accuracy even more into question. What exactly is the statistical probability of an incorrect match based on the number of points used? Good luck answering that question.

IMO the situation is made a lot worse by the fact that the public generally assumes fingerprinting to be extremely accurate (or 100% accurate).

Sorry, but the security required for your Apple devices is less exacting than the demands of justice.

late edit: how often do people try to unlock your unattended apple device?

Ok, but supposing what you say is true, they are junk science when it comes to forensics. There is research on capturing the full 3D structure of fingerprints to actually capture the uniqueness of a fingerprint[0]. But even if the 2D scan is enough not every scanner is capable of collecting that structure.

Given how much R&D goes into making better fingerprint scanners, it seems odd to then claim that a fingerprint taken off of a victim has anything near that level of detail. I guess its sound to say that it may be used as a piece of evidence to further narrow down the target, but its completely disingenuous to claim that it's sound evidence.

[0]: https://ieeexplore.ieee.org/document/9160923

> How then do you explain the success of Touch ID on Apple devices and the equivalent systems on other phones and tablets?

Because on those devices the fingerprint is only used as a second factor (the first factor being physical possession of a particular device).

If Apple used fingerprints like law enforcement do, that is running a fingerprint against the big DB of all prints, then they would certainly get some arbitrary matches.

How many people have you let try the unlock your device with their fingerprint? Tens or hundreds of Millions? That'd be a closer comparison for how a search of fingerprint database work.

Searching a big DB is one way to use fingerprints. It is a crappy way, though.

Another way fingerprints are used is to compare fingerprints found at a crime scene to people who have been identified using other methods that have nothing to do with fingerprints.

E.g., if you have someone murdered in his office when working late at night, and security cameras show that five other people were in the building at the time of the murder, and you are able to get a good set of fingerprints off the murder weapon you don't need a database. You take the fingerprints from all five of those other people, and if one is a very good match and four do not match, you concentrate most of your effort on the one that matched.

It's actually quite similar to how DNA evidence has been used and misused. I don't know how they compare DNA nowadays, but when DNA evidence was first making waves getting people convicted they only compared two DNA samples at a few points. For a given sample there would likely be several people in the country that matched.

That's fine if used right, like fingerprints in the earlier murder hypothetical. Narrow it down to only 5 people who could have committed the crime, get a DNA sample from the crime scene that must be from the criminal, and if that matches exactly one of those 5 suspects it is strong evidence they are the criminal.

Have no suspects yet, run that same sample through a database, get exactly one match, and conclude that must be the criminal. Totally bogus. A lot of people were convicted in the early days of DNA matching that way.

The database method can be made sound, but only if the database includes everybody. Match against a database that includes everyone, only get one hit, and you've probably got your criminal. But if the database includes everyone you are likely to get several matches.

As I said, that was how it used to be. I know DNA sequencing has gotten faster and cheaper over the years, but I have no idea if routine forensic DNA matching now matches enough to make matches unique except in the case of identical siblings.

Suggest you try doing some rock climbing and then seeing how well touch ID works on your phone.

I'm curious if you think that an accused person could be used to setup a Touch ID, then have the prints collected at a crime scene used to unlock the device.

Do you think the forensic prints would meet the standard required by Touch ID?

No, the procedures and theory behind, e.g., bite mark analysis or fire analysis, are total junk.

Do you have some links for this? My understanding is the interpretation part can be BS, but stuff like finding and identifying potential ignition or accelerant locations and types should be more settled.

https://theintercept.com/2019/05/05/forensic-evidence-aafs-j...

https://www.rollingstone.com/politics/politics-features/jeff...

That doesn't exactly support your claims for most of forensics. If you read/skim the NAS report these articles are based on, you will see that much of the issue is not with the underlying science, but rather in the "interpretation" part. Yes, the odontology seems to be unsupported. But even the fire analysis that you list as an example is said to have a scientific basis like chemistry, but that the interpretation part has issues.

While my stance on forensics is not as negative as the parent comment, I don't see how you are drawing such a sharp distinction between the "science" and the "interpretation". They are not separate things.

When a scientist proposes a new test, they include measurement techniques as well as a method of interpretation. If a medical test is used in a way that is frequently prone to misinterpretation, it would be fair to call that entire test useless.

I disagree. The canonical example here are p-values. Completely valid and accepted interpretation which will forever be abused. In the medical realm, we see the same things with false-positive and false-negative rates causing massive confusion. An even more recent example is COVID vaccine efficacy. I would take the bet that 90% of the general population could not give the accurate definition of what the definition of efficacy is.

Just curious about the downvotes.

1. Was it he said statistics lie sometimes?

2. Or, questioned 90% efficacy of some Covid vaccines?

I think you are getting downvoted because (hopefully) people recognize I didn't question their efficacy. Just that 90% of people know what that means.

"If a medical test is used in a way that is frequently prone to misinterpretation, it would be fair to call that entire test useless."

And yet there are many examples of this with malpractice, unnecessary deaths, etc. Lyme tests are notorious for false negatives depending on the lab you use. It's also a widely held belief that you want tests done in the middle of the week instead of on a Friday or weekend because some techs just want to get out of there and are more prone to mistakes.

The reason I'm drawing a hard line is because of fact vs opinion. This same line gets drawn in court. They will ask experts if it's a fact or an opinion that they testified about (usually to attack a witness). Autopsies are full of facts, and then there are opinions on what those facts mean. If it's an opinion, it's possible to counter it with another professional opinion.

The only sort of opinion they should be sharing is included, excluded, or undetermined. Hair is not a "match", but it could include or exclude someone and maybe mathematical probabilities.

It may very well be the case that a version of fire analysis that is scientifically rigorous can be done, but this would have nothing to do with how it is typically practiced.

Which is what @giantg2 said:

> The science is generally sound (procedures and theory). Usually it's the prosecutors that present the forensics in a way that misrepresents what the results actually mean.

https://news.ycombinator.com/item?id=28265884

That’s not the same claim. The practitioners who come up with the results are not operating in a scientifically sound or rigorous way. One documentary I watched about a wrongful murder conviction (a man was found to have burned down his house to kill his entire family) had a discussion of how the fire investigators would talk about “feeling the beast” and other nonsense. It is very flattering to describe this as in any way “evidence-based.”

If you are using forensics in a trial, the validity of the "interpretation" part is key to it's use. The reliability of these interpretations has been historically hugely overstated.

Given this history it is absolutely reasonable to reject most forensic evidence unless there actual good science supporting the reliability of expert interpretations.

The usual example is DNA evidence. The two most obvious fallacies are actually called The Prosecutor Fallacy and The Defendant Fallacy because it's so easy for either side to mis-sell what the evidence really means.

It's basically impossible for someone who isn't pretty handy with Bayes to work out the significance for themselves.

The average jury member has no chance.

Not all forensics disciplins are created equal and some of them, such as hair forensics, are definitely junk science.

How is hair analysis junk science? Maybe you just mean that people misuse the facts to come to incorrect conclusions?

It's only supposed to be used to include or exclude someone, and based on the numbers/probabilities it's generally only useful for excluding someone. Of course that's not how most prosecutors use it...

Hair analysis, as in looking at two hair samples under a microscope and determining if they are a match, is definitely junk science. The FBI has specifically acknowledged this, dozens of cases have been overturned because subsequent genetic testing has shown hair samples did not, in fact, match. In at least one case someone was convicted based almost entirely on hair sample analysis, and later DNS testing showed that not only was the hair not a match, the hair found at the crime scene belonged to a dog. https://www.nemannlawoffices.com/blog/fbi-forensics-convicte...

Citation strongly needed on forensics bring generally sound.

There have been plenty of references in this post to the NAS study on the state of forensics.

Basically, the majority of forensics are based on sound theories and practices. The issues are when professionals and prosecutors try to "interpret" the facts.

I would be interested in any citations showing that forensics are generally unsound.

As soon as the analysts making judgments about the evidence are in communication with the prosecutors or police, it's junk science. You wouldn't do a psychology or medical experiment without at least single blinding, and certainly not with the experimenter telling the subjects what result they want!

Look at this damming quote - "ShotSpotter strongly insisted it had not improperly altered any data to favor the police's case". That shows they don't have a system in place to prevent data being altered or to prevent themselves knowing what the police want, so all their findings are suspect.

Most forensic science is junk.

That said, human in the loop ML can be extremely useful. Surfacing possible positive examples for human review and/or cutting down 'obvious' negatives can multiply the productivity of human reviewers in many contexts.

> Most forensic science is junk.

I'm ignorant on the topic. Got links?

Forensics is full of gee-whiz-cool csi stuff which eventually is shown to be worthless. Polygraphs are a great example; they're still very wisely used in the US, despite piles of evidence showing they're prone to both false positives and false negatives. Sibling comments saying they aren't forensics because they've been discredited are missing the point: these are typical crime tech, and have been lately discredited, but are still in use in the US regardless.

https://en.m.wikipedia.org/wiki/Polygraph

For a more recent example, here's an article on denim analysis...

https://www.nytimes.com/2020/04/08/science/denim-pattern-mat...

The underlying problems seem to be bad incentives (the prosecution is paid+evaluated on its ability to put people in jail) and the court's willingness to accept insufficiently validated tech.

https://innocenceproject.org/what-is-bite-mark-evidence-fore...

To add to that list: fire investigation. There have historically been a lot of myths in the field regarding forms of "proof" that a fire was caused by arson, and many of them were historically used to obtain (likely false) convictions.

https://www.abajournal.com/magazine/article/long_held_belief...

> human in the loop ML can be extremely useful

Human-in-the-loop ML can be extremely useful in avoiding obvious mistakes.

Human-in-the-loop ML can be extremely useful in fabricating desired results.

Not quite sure what point you're trying to make here...

Misusing science is a problem regardless of whether ML is the science in question. It's /already/ a big problem within criminal justice (polygraphs, psychopath tests, etc).

Avoiding 'obvious' mistakes is helpful in reducing the amount of work human raters/evaluators need to do. Automatically classifying 90% of data and leaving 10% for humans can reduce human workloads by 90%, or enable new use-cases where doing 100% of the work simply wasn't feasible. Using ML in this way doesn't make it 'fake AI.'

It really boggles my mind how desperate people are to try and include and blame the police for everything.

>sales people ready to make a call any way their customer (police) wants it made.

You really believe people signed up for a dangerous amd thankless job with high liability and low pay so they can feverishly conspire to have this guy arrested?

Edit: This place is becoming reddit. Someone posting links which dont substantiate their statements gets upvoted, meanwhile my replies questioning this get hit by a downvote brigade.

Just fyi - Police are generally quite well paid for the area they live in. For example, in Alameda (the county I live in) it's common for police to make over $450,000 a year[1].

Also, while police to have a more-dangerous-than-average job[2] (and face unique risks - they are probably the most likely profession to die of a gunshot on the job), there are much more dangerous jobs[3].

So if you have been excusing police because you think they are accepting poverty for danger, I'd ask you to consider that Police are more like extremely well paid garbage collectors[4].

[1] https://transparentcalifornia.com/salaries/2019/alameda-coun...

[2] https://www.bls.gov/iif/oshwc/cfoi/police-2018.htm

[3] https://www.cdc.gov/niosh/topics/logging/default.html

[4] https://qz.com/410585/garbage-collectors-are-more-likely-to-...

With $450,000 as yearly pay i should give up being a software developer. Alameda even has a violent crime rate of half the US average.

https://www.city-data.com/crime/crime-Alameda-California.htm...

I do wonder if the pay negatively affects the confrontations that the individual officer will risk, e.g. let a suspect escape due to not having good enough backup.

That's because he's probably lying/exaggerating. I highly doubt your average officer makes anywhere near that. I don't make over 200k as an experienced developer, either.

People are terrible with nuance. OP did a good job at contextualizing danger of policing, without claiming it's "safe", as you'd see on reddit, but I'm guessing something is getting lost here.

I'm very sceptical someone is in the multiple sixes of figures without serious responsibility and credentials. In my area officers make similar (lower) pay than tech jobs, and every area I've seen is similar. So people on reddit will be complaining about officers making 80k-100k out west, when in reality that's just marginally better than bartenders.

So yeah, they're making like 60k-100k, or more out west. Here in the midwest it's more like 40k-80k. This compares to developer jobs I've seen, but it's not apples to apples. Cops end up years behind in pay, but end up with better benefits. So they'll be at 70k 3-10 years in, instead of 0-5 in.

In either case there's other tiers and roles, was assuming grunts. There might be other cases where "super grunt" jobs exist, and get paid very well. Just like with engineers it seems like the way to advance is increased specialization and moving to management.

You can spin your wheels at low level engineering jobs for decades if you want, too. Usually, people also have to move towards leadership roles to advance in engineering...

So I digress, but I guess it's fair to say police officer is of similar value as engineer in terms of job perks... but engineer is considered a good job.

Where in source [1] does it show that it's common for police to make over $450k a year?

The 'Avg Total pay & benefits' sorted by title has Chief Information Officer/Registrar of Voters as the entry into that level of remuneration. Police specific above that you only have the Sherrif, which is an elected official.

Police pay for the vast majority of roles is significantly below $450k.

That's a good critique! Looking at historical data it seems like "commonly above $300,000" would be more accurate. It's worth saying that listed police salaries are far below these numbers and the officers get there through overtime and other extra payments. Thus the use of 'common' instead of average because you have to look back at what people were actually paid.

Edit: I can no longer edit my original post so this will have to do.

https://transparentcalifornia.com/salaries/search/?a=alameda...

How are supervisors and elected officials making 300k (after all benefits considered) in an area with a high cost of living equal to "police commonly make above 300k"?

Second, it would be absurd to account for extra income to say they make what you're saying. That's like saying a cashier at walmart works three jobs to make an average of 70k a year, so cashiers at walmart commonly make 70k a year.

I can't remember which department, but I remember the local SF/Bay news many years ago talking about some supervisor at that department somehow making $1.5 million the year before (or something insane like that.) They explained how it was all possible, which to me seemed like he had just been with them for so long, he knew exactly the optimal way to play the system and exploit all the loopholes in it.

As for your second part, I guess it would seem unlikely for some positions, but I know a doctor who made like 3x his base salary each year through all sorts of absolutely legal, ethical, and hospital-sanctioned means. Taking every single weekend he could on call (even offering to "relieve" other doctors who had been assigned that weekend) for example allowed him to make much more than you would ever think possible.

Many of these tricks are not secret. They are commonly known and coordinated among emergency service personnel. They cooperate to take vacation, etc. in ways that will maximize the overtime of coworkers. That overtime can then be recycled into time-and-a-half leave that will reciprocate into the maximum overtime for the first party.

An administration that squashes all overtime faces extreme wrath. They can be cutting employees' take home in half. I and plenty of others have seen people lose their houses or suffer other financial crumbles when the expectation of built-in pay far beyond their published salaries were cut off. You can argue that it is foolish to depend on such extra pay and I would agree. But when that significant extra has been their reality for all 10 years of their employment with the department, you can see how they would just plan it into their future finances.

> Second, it would be absurd to account for extra income to say they make what you're saying.

Except it's extra income that they make by being a police officer. A business that hires an off-duty police officer for security isn't hiring them because they're a big burly-looking dude; they're being hired because they are a police officer. An off-duty officer will typically wear their police uniform while working "off duty", and can take actions which an ordinary civilian could not (like conducting arrests).

> it would be absurd to account for extra income to say they make what you're saying.

It's not what I say they make, it's what the public records say they make. You can go look yourself.

For example, from my first link, Sergeant Stuart Eugene Barnes made $125,380.76 in salary and $172,330.17 in overtime. I think it fair to say that's compensation from their job as a police officer.

Oakland officers seem to be able to significantly rack up the overtime these days.

https://oaklandnorth.net/2021/04/28/oakland-police-overtime-...

Kind of makes sense though, most officers start there then move to a safer place once they get enough experience. Remember seeing billboards advertising $80k starting salary 10+ years ago while driving through town.

I looked at your link and I dont see any police making over 450K a year. In fact, I dont see any making over 100k a year except supervisors and elected officials.

It sounds like what's common is that your police do a lot of overtime: https://transparentcalifornia.com/salaries/2019/alameda-coun...

From your source (for the rank of Sergeant):

    Regular pay: $139,197.60

    Overtime pay: $160,477.73

    Other pay: $13,292.65

    Total pay: $312,967.98

    Benefits: $127,240.58

    Total pay & benefits: $440,208.56

Based on Alameda county, here's some "living wage" stats: https://livingwage.mit.edu/counties/06001

The way I interpret that is that they're paid with the idea that they'll have two kids or a Sergeant is a "Manager" (Management $124,884, from the living wage source). I'm curious what the reason for overtime is, but I suspect that it's staffing requirements.

Debating pay to danger ratios is a pretty fruitless effort. People don't get paid for danger.

See: https://militarypay.defense.gov/Pay/Special-and-Incentive-Pa....

TLDR: "Service members will receive $7.50 for each day they are on duty in an IDP area up to the maximum monthly rate of $225."

Moreover, I was a 4 year Corporal living in the barracks with the equivalent of two deployments under my belt raking in about $20k a year.

To my knowledge, this is a fairly standard rate. Side note, we used to crack all kinds of jokes around, "I guess I earned my 7.50 today."

I think what people do get paid on frequently is the availability of people who can and will do the job. A limited and capable talent pool will always make more, especially when demand is high.

I think you're conflating issues with the institution of policing within the United States with the behavior of individual police officers. Without getting into the politics or philosophy of the comportment or purpose of a police force, there are some things in situations like this that are fact:

1. Police officers (and more importantly, prosecutors) are incentivized to make arrests and have those arrests stick.

2. It's a well understood psychological phenomenon that individuals are much more lenient of their own actions. They call this attribution theory in psychology [1].

3. While policing is a risky job, it comes with privileged treatment, powers, and trust that a common citizen does not have. As such, and as public servants, we hold those with these powers to a higher degree of responsibility and behavior than those without such powers. To hold police accountable, we need a minimum level of oversight and control. Historically, this has either not happened or been circumvented by way of police unions establishing ill-advised contracts with the local governments they work with. [2]

I'll agree on points that individual police officers aren't evil, but I'd need a lot of convincing to believe that many policing organizations within the United States are not being held accountable to an acceptable level; nor are they being trained in ways that encourage appropriate interactions with the public they serve.

[1]: https://en.wikipedia.org/wiki/Attribution_(psychology)

[2]: https://chicagounbound.uchicago.edu/cgi/viewcontent.cgi?arti...

1. What are those incentives you mention?

2. That applies to everyone, not just police.

3. This is a fallacy due to media portrayal. If you are late to your job what happens? Probably nothing or you get asked not to do it again. The same mistakes you make, police officer will get investigated by internal affairs and punished over.

This is equally a fallacy - "if you are late to your job" "will get investigated by internal affairs". No, they won't. But they will be protected by one of the strongest unions in the country, if not the strongest.

Policing is also a lot less risky a job than is commonly perceived.

What freaks me out about forensics is no one at all is doing any validation testing. And they aren't doing blind testing either.

Police are telling ShotSpotter to alter evidence from gunshot-detecting AI

https://www.vice.com/en/article/qj8xbq/police-are-telling-sh...

https://news.ycombinator.com/item?id=27959755

It's scary that prosecutors thought that the ShotSpotter evidence is suitable for a court. It's amazing technology that can help police get to a crime scene quickly but is far too easy to spoof or get the wrong answer due to reflections and other sounds etc.

Hopefully cases like this will eventually stop the system being used in court.

I mean, drug dogs are still a thing even though we know they tend to respond to owners desires regardless of if there’s drugs present.

Wow i have never heard of that. Curious to see a link/article on the issue, if you can provide.

This is a decent summary of court cases related to the dogs: https://www.nlrg.com/criminal-law-legal-research/bid/84670/C...

I think this part is pretty key, and has always been a big open question in my mind:

"The court further recognized a potential problem in assessing reliability related to handler cuing, as even well-trained dogs could respond to subconscious cues from their handlers. For this reason, the court determined that a critical factor in determining reliability is the record of false positive alerts made by the dog."

The subconscious cues, or even deliberate cues that only the handler and dog know, could allow an officer with ill intentions to make some kind of signal only known to him and the dog, which causes the dog to make an "alert", giving the cop probable cause.

Some departments deploy dogs that alert on all vehicles when they want an excuse to search. In at least one department, this was referred to as “probable cause on four legs”.

If judges continue to allow this quality of evidence, the police will continue to abuse the legal environment that creates.

Let's face it - most judges are, at least subconsciously, on the same side as the police and prosecutors even though that's not how the system should work.

>though that's now how the system should work.

I'm assuming that's an unfortunate Freudian slip of a typo.

not how the system should work, even though it is now how it seems to work.

You know what you call a judge, a prosecutor, and a cop eating at the same table?

Your average lunch near the courthouse.

I don’t know about that. Judges have front row seats to the 10,000 ways that cops, prosecutors and defense attorneys screw up every day.

If anything, I think the expense of defense and the poor quality of public defenders probably accounts for bias.

Magistrates aren't even required to pass the bar. I had one think that I was calling him prejudiced when asking to dismiss with prejudice (and many other examples with other magistrates, but this one is somewhat funny). I had a judge say that you can't use statements made at a prior trial to discredit the same witness at the current trial if their testimony conflicts. He also denied said that a trial de novo is a "complete do-over" yet denied a petition to dismiss because there's no record of the issues being raised before the end of the trial. So is it a complete do-over? The whole reason they do de novo trials for appeals from the magistrate is because the magistrate doesn't have a court reporter. Not to mention one of the complaints was that the magistrate wouldn't even hear the petition, responding with "That's not gonna happen". Makes sense since he was a retired police cheif. The judge also called me a sovereign citizen when I asked to speak, even though I met the definition of a litigant under the state constitution. There were some minor things too. The main point is, he knew rights had been violated in an irrevocable way as part of that case but decided that it didn't matter.

My impression is that if they do recognize it, they don't care. I also think many dont recognize it. After all, are they going to side with a layman citizen or the police and prosecutor? Even with a defense attorney, it often comes down to the police officer's word vs the accused. Who are they going to believe?

That’s bullshit. Lay magistrates deal with minor issues like traffic infractions, violations and minor misdemeanors.

Misdemeanors can keep people from getting a job. There can be pretrial restrictions, including custody. They also tend to be the ones having preliminary hearings and can issue warrants, even for felonies.

These are not just traffic violations and minor issues. And even if they were, that doesn't mean that people's rights stop mattering, nor is that a valid excuse for gross incompetence.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3078300/

Backs up the GP claim but I can’t find any evidence of this being replicated in another independent experiment. It’s at least a little bit strange that no one tried to replicate these rather concerning findings in the past 10 years since publication.

They probably refused new experiments because of this. It's better to have one test with this result that you can refute as a one off result that hasn't been replicated than dealing with multiple tests with the same result.

An internet search for drug dog false alert returns numerous results.

If you like that you'll love the Wikipedia entry for polygraphs :)

https://en.wikipedia.org/wiki/Polygraph#Effectiveness

For those interested, I recommend reading about "Clever Hans", the horse that could supposedly do math: https://en.wikipedia.org/wiki/Clever_Hans

"Drug-sniffing dogs in traffic stops often wrong"

https://www.chicagotribune.com/news/ct-xpm-2011-01-06-ct-met...

I started reading "Amazon Unbound" by Brad Stone, and in the part where he discusses the development of Alexa, my laymen self had never considered how difficult it would be to train that thing. I think the term was "far-field" speech recognition, and all the problems that come up when adding in real world variables. People talking from across the room while others are talking, or the TV is going, or as you said, reflections and other materials in the room.

I agree that it is disturbing they tried to use this as evidence.

The only people saying ShotSpotter is amazing technology are ShotSpotter's PR team.

By all other accounts, it doesn't work at all and is losing contracts.

> "...a study of Chicago police data found that over a nearly 22-month period ending in mid-April, almost 90% of ShotSpotter alerts didn’t result in officers reporting evidence of shots fired or of any gun crime."[1]

> "In four years, police have made two arrests while responding to a ShotSpotter activation. Lt. Shawn Takeuchi, an SDPD spokesman, could only confirm that one of those arrests was directly related to the activation, but declined to give more information on both."[2]

1. https://apnews.com/article/chicago-police-crime-shootings-be...

2. https://www.voiceofsandiego.org/topics/public-safety/shotspo...

>Hopefully cases like this will eventually stop the system being used in court.

I don't see any incentive here. The only consequence suffered by the prosecution was having to drop that evidence. That's it. Maybe there's a civil case later, but that will be separated by a large time gap, etc.

"Hopefully cases like this will eventually stop the system being used in court."

Maybe it will stop this system from being used in court, but new ones will take its place. It would be great if we could change the structure of the justice system to prevent these types of issues from popping up again. I have no idea what a solution would look like.

If you did that then everyone on here would be ranting about old judges and a luddite system instead

Those same complaints are still applicable today.

You don't have to ban new technology, just impose standards on it.

My personal theory has long been that "technology" is the word people use for things that don't reliably work. It's never clean drinking water from the tap or the seed drill, it's stuff like this or clunky meeting software.

> It's amazing technology

Serious question, what part of this technology is amazing? Its just sounds trained against audio labeled as gun shots. How is this any different then detecting faces in images or whatever?

> Crucially, Williams' lawyers – public defenders Lisa Boughton and Brendan Max – said records showed that ShotSpotter actually initially picked up what sounded like a firework a mile away, and this was later reclassified by ShotSpotter staff to be a gunshot at the intersection where and when Williams was seen on camera. ShotSpotter strongly insisted it had not improperly altered any data to favor the police's case, and said that regardless of the initial real-time alert, its evidence of the gunshot was the result of follow-up forensic analysis, which was submitted to the courts.

"Strongly insisted" links to: https://www.theregister.com/2021/08/02/nvidia_cuda_openai/

> One of the pieces of evidence against Williams claims ShotSpotter's sensors in Chicago identified gunfire where surveillance cameras had seen Williams stop his car by a south-side Chicago block, right when and where the cops said Herring was shot.

> However, Williams’ lawyer submitted paperwork [PDF] claiming ShotSpotter actually detected a firework a mile away from that location, and that ShotSpotter later reclassified the bang as a gunshot and the location as being where Williams was seen on camera, Vice first reported.

> Williams' lawyer demanded the court hold an inquiry into the ShotSpotter evidence, and the prosecutors simply withdrew it.

> ShotSpotter responded by denying at length it improperly altered any data or evidence, and hit back at any suggestion it had done so to help the police make a case. It said its software generates real-time alerts automatically, and staff later analyze the microphone readings to submit forensic reports for the courts, and these final reports can therefore differ from the initial alerts.

> "The idea that ShotSpotter 'alters' or 'fabricates' evidence in any way is an outrageous lie and would be a criminal offense," it said in a statement. "We follow the facts and data for our forensic analysis. Period."

> Update: The case against Williams was dismissed by the judge at the request of the prosecution, which admitted it now had insufficient evidence. Williams had spent the best part of a year in jail awaiting trial.

Wow, we need a civil rights lawyer to represent Williams in a wrongful imprisonment suit against the city, if not also a civil suit against ShotSpotter (or maybe even a criminal complaint for knowingly providing false testimony?). People shouldn’t be able to just dust their hands of this.

"People shouldn’t be able to just dust their hands of this."

Yet the system will let them. I've had a prosecutor and trooper hold a charge that they know was incorrect and subject us to pre trial restrictions found only under that charge. The prosecutor also used their position to block our ability to have a witness remotely testify by telling court scheduling not to even speak to us (inappropriately using their position to influence the court). The state police said they see nothing wrong with knowingly holding an incorrect charge and subjecting people to pretrial under it. The Bar said that even though the conduct we described would constitute prosecutorial misconduct, they won't pursue complaints against prosecutors unless the court makes a determination supporting that. That would cost thousands of dollars to hire a lawyer, make an appeal, etc. The ACLU said they have bigger issues to deal with. I submitted a complaint to DOJ, but I'm not holding my breath.

Literally everyone we dealt with as a part of the system made a bunch of mistakes and didn't give damn about correcting them or protecting people's rights. It's us vs them and all about the money.

>So the AI didn’t even make the call. The staff did, manually.

(a) The staff of the company who makes the AI, which would have an incentive to make the AI's original judgement appear accurate?

(b) The staff of a company paid by the state/police, which aims to please its customers?

(c) Was the staff trained/experts in recognizing such sounds, or were they random AI company employees, which just made a judgement call? If so, was this call presented as AI-supported to the trial, and was it sold as anything beyond "some devs, no gun experts, mind you, heard this recording and said this must be a gunshot".

It could come at the behest of police as in "it would be really helpful to our case if there was ShotSpotter evidence to support our cased". It could come from the fact that ShotSpotter is losing customers, and willing to do anything to make the product "feel" more valuable to the police to ensure cash cows keep producing.

The modern criminal investigation has so much negative baggage attached, from historical shenianigans, that it's easy to find ways to question the prosecution's case.

If manual reclassification is routine then what does that say about the AI?

That it doesn't give the answers the customers (police and prosecutors) want.

Which is mostly unrelated to whether or not it's accurate.

I think it’s related. If it is good enough to put someone in jail there shouldn’t be any kind of need for manual overrides at all.

"Good" in what sense? Police and prosecutors don't necessarily care about their evidence being good in the sense of accurate. It just needs to convince a jury.

There is a very, very, very long history of police and prosecutors using bad (in the sense of accuracy) evidence to put people in jail.

I’m making a normative claim. Obviously it does not describe reality.

There was a submission here a few weeks back about ShitSpotter and how they seem willing to put in overrides at request of the police without necessarily having a basis for it.

If that's what happened in this circumstances it would be, police check ShoySpotter for more evidence, ShotSpotter hasn't automatically marked anything, police request manual review for gunshot they are sure must be there, ShotSpotter manual review process comes up with something and marks it as a shot with a human override.

There's a lot of places bias, or outright corruption, could cause problems in that process, if it's what happened here.

It's the AI companies to blame, always. Obviously the algorithm is without blame, it's not a blamelable entity.

What almost invariably happens in cases like this or the cell phone surveillance boxes (Stingrays) is that if the underlying technology or behavior is challenged or questioned strongly enough, the evidence introduced by them, if not the case itself, will be withdrawn, in order to try to avoid any negative precedent against the device or technology being set.

This is missing a key crucial insight: the overall number of evaluated cases increases when AI is helping meaning that borderline cases which may have been ignored prior are now being evaluated. Said another way, with AI handling the obvious cases, the number of borderline cases being evaluated by humans increases. With humans having large amounts of inaccuracy, borderline cases before which may have been ignored due to not having manpower to evaluate are now being evaluated. This can and does lead to many more erroneous accusations despite the fact that all the borderline evaluations are being handled by humans.

Another factor here is that simply evaluating more data is also likely to lead to more erroneous evaluations since random chance has a much larger impact.

But this is so convenient! "Your honor, this is the AI algorithm that said it was a gunshot, AI can't lie and is based on pure Science!" Yeah, later it turns out people made that decision because police asked them to do it, but the defense may not know it.

It sounds like they reclassified it after the police said "Gee, are you sure you don't have a gunshot? in a slightly different location?"

I wonder how much of "AI" introduced in court (or for a warrant) involves some element of individual human attention though... They just call it "AI" without qualification until they get caught (if even then).

Would it be better though if it had been purely the output of a machine learning algorithm? (Which after all is not infallible). In that case, the guy probably would have been convicted on that alone...

Coming next year: "A man spent a year in jail on child pornography charge that hinged on disputed Apple CSAM evidence".

I know you're being coy to slam on Apple, but this already happens in the uS __without__ Apple's (imo misguided) strategy and it happens frequently.

Prosecution over thumbnails, file hash matches with tenuous connections to individuals, Crime-Tech has given a new position for law enforcement to craft fanciful narratives to support prosecutions of just about anyone. I'm not specifically citing or inferring there have been *targeted* instances of this, but it happens very often towards US citizens already.

Stingrays, Cellebrite, Forensic "science", ShotSpotter, it's absolute chaos and anyone can be a victim here.

My goal here is not anything about Apple and its CSAM strategy, but to make it very clear that the imagined scenario happens now and happens frequently.

I'd be interested in reading about some of these real cases if you have any links handy.

For that to happen: they'd have to be matched 30 times, the Apple employees would have to mistakenly manually confirm the match 30 times, and the defence would fail to request to see the underlying images that matched the hashes during discovery.

It seems... far fetched.

For now.

When your slippery slope includes the abolition of pre-trial discovery, it's a very slippery slope indeed.

When does slope become cliff?

When it's your phone.

Did you just use a slippery slope argument to justify slippery slope arguments? Impressive. How deep can we go?

You don't know it's 30 and you don't know it's going to stay so even if it is.

Right, because imagine the spin.

"Our CSAM matching found 29 matches that are either confirmed or strongly suspected in your iCloud drive. But because Apple values your privacy, we are not turning this information over."

No. Expect that number to go to '1', or very close, and very soon. Because as much as this is a PR nightmare for Apple right now, the above scenario is even worse.

You might be surprised to know that many other major tech companies have already been using the same CSAM technology for the last two years and how many of these situations have you heard of? I can't think of any.

This year in jail was awaiting the “speedy trial” that is supposed to be guaranteed by the US Constitution. I think that in itself is as much of an issue as the quality of the evidence. How can the government justify taking year away from the life of a man who has not been convicted of anything?

A year? Curtis Flowers spent over 20 years on Death Row without being convicted of anything. A couple years ago he was set free! Why? Because of a true crime podcast called In The Dark. Seriously. Listen to season 2 and your mind will be blown. The “justice” system in this country is horrific.

https://en.wikipedia.org/wiki/Curtis_Flowers?wprov=sfti1

I agree with your distress at the state of the justice system, but it doesn’t make sense to say that someone was on death row without having been convicted. The article you link to explains that his convictions were overturned. Although I would never rely on information in a Wikipedia article, you did link to it, so presumably you find it credible.

Every time he was “convicted”, the prosecution was found to be in blatant violation of laws on appeal. They intimidated witnesses, kept black people off the jury, and more. So no, he was never rightfully convicted! The prosecution broke so many laws, way more than Curtis Flowers ever did! And they just did it over and over again for 20 years. None of them have faced any consequences. They effectively imprisoned an innocent man for 20 years because they didn’t like him. And big surprise, Curtis is black.

What you are describing is different from convicting someone withoht trial, it's way worse.

The prosecution breaking the law doesn't mean the trial didn't happen. "rightful conviction" and "conviction" are very different things.

It's typical for a defendant to waive their right to a speedy trial. If you don't waive it typically you get the book thrown at you and in the little time you have to prepare, something sticks. I agree this should be changed but there is not much political will behind it, as prosecutors view it as a necessity to try everyone fairly. The real solution is to not bring so many pointless cases to trial.

Knowing that, I think I would demand a speedy trial, but most people have horrible lawyers.

Huh, thanks for pointing that out. I had no idea that waiving the right to a speedy trial was even a thing. If you do not so waive, what is the standard for a trial date that is considered constitutionally speedy at the moment?

It depends on state. In Kansas, it is 180 days if out on bail, or 90 if held. Note this does not include time before you are arraigned. In some states it does (personally, speedy trial should include the time before you are arraigned, so that the state is forced to hurry up and deal with cases or dismiss them).

I'm unsure about federal guidelines. I think many states have laws that could be challenged in federal court, and their de facto behavior most certainly can, if you have money.

A lot of times, people have to waive their right to a speedy trial. Apparently, in some districts they don't like it when the accused doesn't waive their right, so much that the defense attorney will get assaulted by the judge.

https://youtu.be/jAv1QUkHFSY?t=397

This. "Michael Williams, 65, who denied any wrongdoing, sat in jail for 11 months awaiting trial for allegedly killing Safarian Herring, 25."

Do not waive a speedy trial ever. The judge will want you to, your attorney will want you to ($), the DA will want you to. Tell them to go fuck themselves.

'Florida Judge'

It's interesting to me that there's no working process to introduce/accept new types of science in court. Polygraphs aren't allowed (though the police still find ways to use it to bully people). So, there's some process. But you still see stories about subsets of fiber, hair, arson, fingerprint, dna, etc, science, that shouldn't have been brought to court.

The legal system will, over time, also have to accept and adapt to the idea that free will is an illusion. (See “neuro-law” for more information on this.)

Unless you are aware of any breakthroughs in fundamental physics or neurology which are being kept from the rest of us, “free will is an illusion” is not a very useful theory when it comes to public policy.

The idea that the people who run the legal system have the free will to change it is also an illusion.

Link: https://en.wikipedia.org/wiki/Neurolaw

> The rapid growth of functional magnetic resonance imaging (fMRI) research has led to new insights on neuroanatomical structure and function, which has led to a greater understanding of human behavior and cognition. As a response, there has been an emergence of questions regarding how these findings can be applied to criminology and legal processes.

I hope they have a more solid foundation of evidence than fMRI.

Eg from https://www.fastcompany.com/90520750/duke-university-researc... :

> The researchers reexamined 56 peer-reviewed, published papers that conducted 90 fMRI experiments, some by leaders in the field, and also looked at the results of so-called “test/retest” fMRIs, where 65 subjects were asked to do the same tasks months apart. They found that of seven measures of brain function, none had consistent readings.

Also https://www.pnas.org/content/113/28/7900 :

> Functional MRI (fMRI) is 25 years old, yet surprisingly its most common statistical methods have not been validated using real data. Here, we used resting-state fMRI data from 499 healthy controls to conduct 3 million task group analyses. Using this null data with different experimental designs, we estimate the incidence of significant results. In theory, we should find 5% false positives (for a significance threshold of 5%), but instead we found that the most common software packages for fMRI analysis (SPM, FSL, AFNI) can result in false-positive rates of up to 70%. These results question the validity of a number of fMRI studies and may have a large impact on the interpretation of weakly significant neuroimaging results.

Why would it? Aren't judges and legislators equally lacking in free will too?

> “I kept trying to figure out, how can they get away with using the technology like that against me,”

I have a bad feeling this will be a theme over the next decade as government and law enforcement around the world starts to integrate "AI" (or rather call it what it is: statistical models + mass surveilence), without enough scepticism built into their process. Some of us know better, but it looks like society and government will still have to learn the hard way that these tools are not a highlight pen, they are a very crude spray can, and even then they will miss spots.

It seems to me jail time without verdict is the real problem in this scenario — it’s not clear to me how that’s not “guilty until proven innocent”

Precisely. This happens to people all the time regardless of the technology involved. Being suspected of a crime should never ruin your life. At most it should pause your life.

House arrest with deferral of your debts and obligations and mandatory reinstatement of your job if you are exhonorated or the charges are removed seems like what you would need at a minimum to live up to "innocent until found guilty".

That's exactly what judges decide to do if certain criteria are met, though.

The person in this case, however, had a criminal history including attempted homicide so the judge decided to opt for detention while awaiting trial instead.

Sounds like perfectly normal procedure to me and happens other countries as well.

In Germany and Austria you get compensated for your time incarcerated if found innocent. So as to basically "fine" the government for not having speedy trials.

that sounds fair except not great that it's just penalizing the taxpayer, not anyone in government

We need to quit thinking about "penalizing the taxpayer" type of mentality. It's the governments income. You don't get to act like a taxpayer for all the gas you buy or interest you pay in, but suddenly when money funnels to the government your money and how it's spent matters. The government is just another corporate entity. Not like you get the money they tax back anyway.

Wow - any software system that is used to submit evidence to the court should have a preferably cryptographically secure chain of custody of all artifacts - including the raw recorded audio and the code used to process it. Any manual 'post processing' should be tracked down to a person, a date, and a verifiable reason for making changes to the model/decision. For 'black box' models an independent organization should be able to test them to confirm stated accuracy/precision.

I don't think the answer is to do away with these systems but as we yield more and more authority to machines regulation will need to catch up so they aren't just used a fig leaf for corroboratory-evidence-as-a-service.

The US has the most troublesome track record of law enforcement abuse of "evidence".

https://innocenceproject.org/what-is-bite-mark-evidence-fore...

https://bostonreview.net/books-ideas/nathan-robinson-forensi...

https://archive.nytimes.com/www.nytimes.com/books/first/k/ke...

Wow there are microphones used to detect gunshots in Chicago? I remember seeing such technology in a relatively obscure cyberpunk comic book when I was a kid, but I never thought this would be legal.

Do the microphones record conversations too? Can the government spy on its own citizens? I guess some secret court ( https://en.m.wikipedia.org/wiki/United_States_Foreign_Intell... ) allowed it?

Oh boy, wait until you learn it’s not just Chicago where they deploy this crap [1]

[1] https://www.shotspotter.com/cities/

There are LED street lights in San Diego California that do just that. They have cameras and microphones. The microphones are used for Machine Learning of what people say. [1] There are a handful of other cities deploying these. It is a clever way to deploy CCTV much like what the UK has but without drawing much attention to it.

[1] - https://www.smartcitiesdive.com/news/san-diego-faces-lawsuit...

This is about ShotSpotter.

https://hn.algolia.com/?q=ShotSpotter

>“The problem? When we made it possible for anyone to generate art with artificial intelligence, barely anyone used it to make actual art. Instead, our AI model was forced to make videos for random inputs, trolling queries, and NSFW intents.”

I am struggling with the use of the word "forced"; the researchers put their software out there to be used by anyone with an internet connection. The researchers created that dynamic.

And how are they not seeing that AI created no-humans-used-to-make-this-porn could be a net good.

I get that their system may sometimes need manual classification of signals but how is it possible to change the triangulated location? That is the core part of the technology and doesn't require ML to implement. Seems like it's all just Theranos style vapor.

Locating on an echo at one more receivers rather than a perhaps quieter more direct path. Putting the wrong impulses together when computing a location. Computing the wrong start time for one or more impulses.

If you want to read up on all the ways this kind of thing can go wrong, the technical term here is "source localization."

The military has no problem detecting gunfire direction and estimating range from a single microphone array on a mast. Shotspotter is a distributed network of multiple sensors across a wide area. If they can't make it work after 15+ years then they are hopeless.

Boomerang is colocated with the target. Multipath is less likely to be an issue because there is likely to be line of sight to the source. Selecting the right impulses is easier because the spacing of the microphones provides some tight timing constraints. You’re not likely to see inconsistencies in timing the impulses lead to location error because any effects of the environment are likely to be consistent across the pulses. Finally, the location accuracy requirements are far different. ShotSpotter needs to put a dot on a map accurate enough to find shell casings or a victim in a wide geographic area. Boomerang just needs to give a good bearing from the receivers to return fire. The range doesn’t have to be all that precise, and the mathematics suggests it isn’t.

These are two very different approaches to solve different problems.

Fake AI?

Who would have thought?

Shocking, shocking I tell you.

https://venturebeat.com/2021/04/05/government-audit-of-ai-wi...

Well that's still a failure on both the software programmed by humans causing a false positive and then doubled down by a wrong human review, based on a false alert. All involved stakeholders have failed.

How is this a failure by software? The software categorized the sound as a fire work.

> How is this a failure by software?

The outcome - a guy being charged with a crime using bogus evidence is the story. The software may have been right, but the people who operate the software didn't trust it. Regardless, the story is trying to point out we might not be ready to put our faith in AI for criminal prosecutions at this time.

If shotspotter had not existed, I think this man would have still been charged with the murder. It seems beyond dispute that the victim was shot in the car of the accused. The case against him was seemingly lost for two reasons: shotspotter's fuckery, and the police not bothering to collect other evidence in a timely matter. It should not have been hard for the police to determine whether there was a gunshot fired in the car, or whether the passenger was struck by a bullet fired from outside the car. There was really no reason for shotspotter to play a role in this case, it could have been cut and dry but they went and muddled the whole thing up with their janky tech.

If you use a hammer to drive a nail, and the nail goes in exactly as intended, but then the carpenter rips out the nail, how is that a failure of the hammer?

Yeah, looks like it's the actual software that's undermined the case against him. A loud bang which coincides - perhaps pretty loosely - with some footage, some supposed audio expert contractor's human employee - possibly knowing other details of the case - thinks it's consistent with a gunshot but the case collapses when records reveal the AI program they ran originally classified the loud bang as a distant firework.

'The system' is software + people who run it. Their failures are combined. As an external person i dont care which part of the company fucked up, they shoupd carry responsibility

The software didn't raise a false positive though, by my reading.

It identified it as a firework a mile away! The humans at the software company, REclassified it as a gunshot at Williams' location.

Interesting that there's no mention of the actual gun shot, which appears to have happened some time later?

Not only is this "tool" bullshit, it seems the employees of said org are morally bankrupt enough to falsify evidence in order to put an innocent man in jail.

The whole thing would be utterly unbelievable if it wasn't happening in America.

EDIT: typos

I’m not a Luddite by any means but it is time to be more honest about the limitations of AI.

The only time I hear about people who talk about AI are average joes, ignorant journalists, generic business analysts, and upper management that don't know anything about programming. I wish I could point to these people and say "look at video game AI, we can't even make those without it cheating or following a routine to simulate human playstyles. What makes you think we can do that in real life without doing the same?"

Video games are a bad example because it’s pretty feasible to make an AI that would win every time.

They are a perfect example because of that. They are in controlled environments. Life is the testing ground of everything. To think we can trust a computer to have reasoning skills in our environment ethically is absurd and a total affront to freedom.

Only in very simple games. There's a reason that there's not more AI multiplayer grinding bots.

Not really relevant to this article, but relevant to the title: if AI keeps banning accounts for reasons that seems nonsense to humans, how will we protect people from getting flagged and locked up 'just in case' by the machine?

AI is just not advanced enough to jail people, by a very large margin. And yet we will see it happen.

The scariest part is how quickly policing agencies are to embrace this tech as well but the government as a whole is slow to change anything else. Why is it that the state can be proactive about jailing me but not as much when it comes to quickly resolving cases and protecting my rights?

> Soon after another vehicle pulled up alongside, and someone in a passenger seat took out a gun and shot Herring in the head, Williams told police.

I dont understand what Shotspotter has to do with this case at all. The question the case seems to hinge on is “who fired the gun”, not “was a gun fired at this location”.

What am I missing?

>What am I missing?

The fact that some of the evidence the police used to say who did what where was manufactured by the company.

The headline is:

> A man spent a year in jail on a murder charge that hinged on disputed AI evidence.

But after the reading the article, the evidence seemed irrelevant?

Maybe the headline should have been:

> AI evidence was manually reclassified after human review in murder case

Incorrect

The police used the altered location of the shot to imply that the defendant's story about what happened wasn't true. Which was that another person in another var at a different location did the shooting.

They used the altered location data alongside video evidence to imply the defendant and the victim were alone at the shot location and therefore the defendant was the only possible suspect.

> The police used the altered location of the shot to imply that the defendant's story about what happened wasn't true. Which was that another person in another var at a different location did the shooting.

Ok that would make more sense, but I definitely did not take that away from the article. I’m sure there are better sources on this case though.

>I’m sure there are better sources on this case though.

Yet another example of why you cannot get your information from one source. Even if they are not trying to mislead, the one author might not have all of the information to provide in thier writings.

We have this system in our city. Can’t speak to the accuracy and whether or not it should be admitted evidence. It has in some cases revealed the approximate locations of victims who have been shot where they potentially could have died without the AI heads up as to where they were.

Prosecution based on evidence provided by closed source and closed hardware based devices should never have been admitted in courts of laws in the first place.

Anyone else surprised that such a system is running across cities?

“It's said that in May last year, Williams was driving through Chicago one night hoping to buy some cigarettes.”

Ah yes, I remember hearing that a lot growing up.

Editors, do your job!

What do you mean?

I think OP meant that "driving through Chicago" was pretty large/vague area for someone just trying to buy cigarettes.

It's like saying I was driving through the United States to go to my local corner store.

I wouldn't have thought about the poor writing if OP didn't mention it, but it is still a bit pedantic in my opinion.

More likely "it's said" - a claim that some unspecified people said something, expressed using a phrase people often interpret to mean "it is generally believed that..." or "there is a saying that goes...". GP sarcastically said sure "it's said" that x occurred, I heard it all the time as I was growing up.

It's not good journalism. If someone said something, say who said it. If it's an anonymous source say so. "It's said..." in an article like this means "I vaguely think this has been said somewhere but I can't be bothered to check or quote my sources".

Thank you everyone (including OP) who clarified that it was "it's said".

“It’s said” — by whom?

Weasel words: “it’s said”

The sentence I quoted makes no sense in a news article.

- The use of the passive “It’s said” has no business in a news item. Who said it?

- Idiomatically “It’s said” is typically used to identify a bit of folk wisdom, not a particular incident.

It’s terrible writing and it should appall a decent editor. This isn’t a blog post or a comment, it’s an item in an English language professional online publication. Do better.

This seems like the definition of circumstantial evidence doesn't it? AI, human, accourate or not, he shouldn't have been in jail without something more.

Reply on news.ycombinator.com