Walmart employees are out to show its anti-shoplifting AI doesn’t work
arstechnica.com
arstechnica.com
An AI that fails to recognize some thefts but does recognize others can still be considered to work as long as enough thefts are caught. Imperfect, but given the low cost of a failure relative to not using an AI, imperfection seems quite tolerable. Even the false positives seem to be of little concern as they involve a shopping associate coming over and having to assist an individual. As long as the experience for the shopper remains more pleasant than waiting in line for one of the cashier operated checkouts, it is still an overall win.
So yes, the AI is imperfect, but where is the information that justifies the conclusion it doesn't work and not the conclusion that it doesn't work perfectly?
So it's really a that the parameters of the tech are overlapping poorly with the parameters of customer convenience, perception, treatment, and other aspects of the next consumer experience and op. costs.
This is probably a great example of understanding the limitations and special needs of AI in the real world. It's not magic, it's just 'some magic' that we have to develop processes to work with.
You can trick any system, but that's not telling on it's overall effectiveness.
So I stopped shopping there. Wendy's FTW.
https://www.google.com/amp/s/www.cio.com/article/2437873/45-...
Fast forward several years, there’s more staff stationed with self-checkout due to theft.
I think retailers are going to calculate the cost of headcount and theft and make a decision based on an equilibrium.
Of course, the best outcome for Walmart is to have an automated LP system to reduce staffing cost for the long term.
Now they have 2-3 employees monitoring the self checkout.
There is a certain point where employers become more focused on reducing shrinkage than headcount.
(I’ve been a cashier at a grocery store)
That store must have lost thousands every day from that, but they're an insanely busy store and probably more than make up for it in sales.
Also, I remember when I worked at a grocery store, the policy was if you seen shoplifters you weren't supposed to approach them or do anything other than to just tell the manager. There were a few thefts from that place and this was before the days of self checkouts. I remember I seen a guy once stuff some things in his backpack, I told the assistant manager, he followed him around through the store, but in the end the guy just left and nothing really happened. He kept whatever he took.
But, they scale. There's always a threshold for the accuracy/efficiency point. It's not meaningful to identify errors, even if they seem glaring, if overall you're reducing 80% of loss at 10% of the expense.
On top of that, you have to also account for the concept that the employees themselves might steal from the store.
I've just been generally surprised at how often I see an ML solution implemented the complaint is "it isn't perfect", when perfection was not the goal metric. I think people just expect robots to do everything exactly right.
Also, scanning 1 item while holding two of it in your hand.
I've also walked out of Lowes with a couple dollars in random hardware in my pocket before. Went in for some screws so didn't grab a cart, decided to pick up some other stuff since I was there, and stuffed them in my pocket temporarily. Forgot about them during checkout. I actually went back and paid for those though because I felt like such an idiot about it.
Technically it's your kids stealing, but since you are responsible for their actions, it's on you.
When this happens to me I usually let them know the next time I'm at that store (or if I discover it in the parking lot I go back in). Sometimes they charge me for it, usually they just say thanks for being honest.
Often the staff look like they are on autopilot and aren't reading anything and just going through their shift on muscle memory so I imagine this happens a lot. Is it stealing if you tried to pay but the staff did something on the machine and it didn't charge you? I don't know.
Any implementation should use this as a supplement only, not replacement for humans. For example a human watching 10 checkouts and quickly manually reviewing scenes flagged as suspicious. Any implementation should also account for the fact that loss will still occur and account for that in the RoI.
The fact that the technology is very error prone is not surprising (to anyone that actually builds AI tech). The issue here sounds like poor implementation and unrealistic expectations from management on what AI truly can and can’t do.
You probably know the weight of one jug of milk, but even if you don't, it's easy to "learn" the weights of new SKUs statistically.
My guess is that they have more tolerance for items not of the exact weight and/or more accurate scales.
They tried this. Was a disaster.
Self checkout without that works well in many supermarkets, IKEA etc. There are losses but with random checks they are lower than the staff costs - so It's a win.
Who are you checking, the cashier or the buyer? My guess is a deterrent for some kind of collusion.
Same with Fry's. Not the same with Costco; there, you're subject to the membership agreements.
At Best Buy I stopped shopping there for years. I don’t remember why I eventually went in one again, but in the years since I’ve never been asked once. I do wish they’d get rid of the guy at the door entirely. It’s weird. And they always greet you from the side or back.
This seems a really odd place to draw the line on surveillance.
What’s the beef with my choice?
It just seems odd that they've captured your face walking in and all over the store with their other cameras, but then you choose to cover the camera at the checkout. Why? What do you hope to gain/prevent by doing that? Why are you risking getting in trouble?
Why is that one camera where you draw the line?
If the cost of having self-checkout machines that actually _work_ and let you complete your purchase is that some items slip by, that might well be worth it.
In contrast, Safeway chain self-checkouts are such a miserable experience that for more than two items, even waiting in a long cashier line is worth it. Their system requires you to place every purchased item on a scale after the scan, and a good fraction of the time the scale fails to read the expected added weight, forcing you to wait for staff to unlock the machine again. For a cart full of purchases, this will happen half a dozen times.
And I don't understand what this has to do with COVID-19 at all in the article.
Finally, what do employees care about theft? It's not coming out of their paycheck. False positives from an anti-theft system seems like a small job nuisance at best -- not something to whistleblow with a "slickly produced video".
Nothing about this article makes sense to me, unless it's motivated by an Everseen competitor -- whether another company or an internal division of Wal-Mart that thinks they have a better solution. In either case, why would the general public care at all?
The article seems to be mainly about theft during self-bagging.
We didn’t have self checkout back then, though.
I think their is a massive divide from Walmart's Corporate/Tech branch and that of the typical store employee. I overheard one employee saying back in December as we were days away from NYE that she had to remain at work despite having been approved for leave weeks earlier as most of her coworkers simply just didn't show up to work and demanded a raise from her manager. I now always wear earbuds when I go inside of one.
Whereas for the tech employee or risk management office drones they don't face those concerns and instead have to justify a growing budget for their department, perhaps even their jobs in this environment, with these kind of antics.
I inadvertently noticed how easy it is to use their buggy and slow check-out kiosks against them, but since I only ever buy motor oil at Walmart anymore I don't see this being a useful risk. I actually went back and paid for the item when I noticed it made the noise that it scanned but didn't include the item to my checkout price, which was a total pain in the ass that meant standing in line at the nearly always understaffed customer service/return area.
Bobby Lee, the ex-owner of massive Bitcoin exchange in China and the brother of Charlie Lee creator of Litecoin, was ex-Walmart tech. In a recent podcast outlining his previous history in the Bitcoin ecosystem he made it clear they have no shortage of talent there. To what end, is another matter entirely.
what is the positive to false positive rate? projects can have metrics that those in the field are not privy too and it could down to they are preventing a sufficient dollar amount to not consider the false positives are an issue. Again, what are the percentages?
People take fruit and vegetables, and put the money in the box--or if the farmer is tech-savvy enough, through their phones. And even if a few people don't pay, they still get more money from their inventory than shipping it to a wholesaler or a grocery store. (As far as I know, direct to a local gourmet restaurant pays best.)
For an extra employee on the duty roster, you're probably paying minimum $7 per hour, every hour that store is open to the public--probably 16 hours a day, every day, for a big box. That's $41k per year, plus the per-employee costs for 2 or 3 people, which are probably another $60k, unless you're an employer that doesn't offer benefits (i.e. Wal-Mart).
The employee would likely have to stop $200/day of petty theft. That doesn't seem impossible, but the problem is that if you pay your employee minimum wage, they don't actually care that people are stealing from their employer, because they have no actual incentive to do anything beyond what is the bare minimum requirement to avoid getting fired, which may be just catching one shoplifter per day.
And then, after all, if one caught every shoplifter, every day, people would stop trying, and one might miss quota on later days. The person whose job depends on shoplifters has a perverse incentive to encourage shoplifting, just as cobra-catchers have an incentive to breed and release snakes.
So the problem is more that it's too expensive to hire someone at a rate high enough to make them care about doing their job well, instead of just keeping it indefinitely while gaming the right metrics.
It does.
> The employee would likely have to stop $200/day of petty theft. That doesn't seem impossible, but the problem is that if you pay your employee minimum wage, they don't actually care that people are stealing from their employer, because they have no actual incentive to do anything beyond what is the bare minimum requirement to avoid getting fired, which may be just catching one shoplifter per day.
> And then, after all, if one caught every shoplifter, every day, people would stop trying, and one might miss quota on later days. The person whose job depends on shoplifters has a perverse incentive to encourage shoplifting, just as cobra-catchers have an incentive to breed and release snakes.
You just explained the the broken window type fallacy of LEO and Police in general. When the crime no longer exists, you create it... by force. Oddly enough, it also applies to loss prevention.
> So the problem is more that it's too expensive to hire someone at a rate high enough to make them care about doing their job well, instead of just keeping it indefinitely while gaming the right metrics.
Agreed. And again, it applies to more than just shoplifting prevention.
https://en.wikipedia.org/wiki/Cobra_effect
Long story shortened:
1. British are ruling India
2. Concern about cobras
3. Establish dead cobra bounty
4. Cobra population crashes
5. Entrepreneurs breed cobras
6. Bounty program halted
7. Breeders release stocks
8. More cobras than before
The source story is just an anecdote, but there are plenty of historically verifiable examples of the effect, not necessarily involving cobras, specifically.I'm not sure how many high intensity "you're stealing!" confrontations really occur.
The other issue I have is that the barcode scanners are not activated until the weight sensor settles out, which can take several seconds up to a minute, so it's way slower than going to a checker. If the weight sensor didn't gate barcode scanner activation I would have a much better experience.
All they really need to do is verify the weight of the total transaction and have the associate check on any discrepancies. And they can probably use a very noisy weight signal with a low-pass filter. They don't need to blow a ton of money on an AI solution that is about the same level of inconvenient.
If I had 10 or 20 million dollars I'd build a better self-checkout.
It's so much nicer of an experience compared to machines that just stop functioning until an employee can "help" you.
I know when the shops near me (UK) started rolling out automated checkouts I'd avoid the shops that had poorly calibrated systems that required frequent interventions because it was extremely annoying. False positives certainly had revenue implications for those shops.
Their job isn't to "identify" the shoplifting, but deal with the consequences and shoplifter after it has been identified.
So if the A.I. get's better, then they harass less innocent people, and the humans job gets better.
The employees job will be at greater risk when Walmart goes full ROBOCOP.
[1] https://www.nytimes.com/2018/08/17/business/falsely-accused-...