The article mentions machine learning. So, it triggers, they search you, they find a thing that is not a weapon, they update their database and comp you something for disturbing you.

Of course, they could as tgb suggests just ban anything that triggers their limited tests but that would impact income once word got around

But until that happens, a casino is consistently getting false negatives. What happens when the trinket for sale as you're exiting Cirque du Soleil looks like a machine gun to the system? Suddenly, you've got dozens of people held up at the entrance. At best, that's a customer service annoyance; at worst, a bad actor could use this to their advantage.

Then they put a sign up saying "please remove trinket from pocket". This really doesn't seem like a huge problem.

Sure it does. Who's going to go because they constantly get stopped?

I think the false-negatives of such a system, which weren't entirely addressed ("will get better with machine learning"), would be worth more to look into than things not guns being seen as guns.

False negative and false positive rates are correlated, you can generally trade off one for the other - although I don't know how easy that is with machine learning. The false negative rate doesn't have to be perfect to be a deterrent - would you try to sneak a gun into a casino if you knew you had a 50/50 chance of being caught?

People sneak and accidentally take guns on a plane all the time and don't get caught, but I'm not sure what the 'catch-rate' is for that.

Well, that's a concern (for the profitability of the casino) but isn't what I was replying to. The parent post was suggesting that there could be objects designed to look like guns to this system. I have a hard time imagining why anyone would do that.

Which implies it shouldn't be too hard to sneak in a gun if it doesn't look like a gun to the scanner.