Background: I spent 40 minutes typing up a reply to a different post, but decided that it ran on for too long. I'll include it at the bottom, but I'm curious to know how much cash is still used in this country.
Certainly, a like-for-like Giro (Europe) system doesn't exist in the USA, with ACH, checks, and Zelle almost filling the void -- albeit incompletely -- which I suspect is responsible for the remaining cash utilization. But is that right? Is cash only used for when there isn't another option? Or is it a matter of consumer preference?
I can understand tipping in cash, or paying for a Craigslist purchase in cash. But maybe I'm missing another dimension? Do some folks pay rent in cash? Or taxes? I'm genuinely curious, but please make sure not to dox your finances in the comments.
My original comment
It's annoying when they get suspicious of a 25k USD withdrawal for instance (even if you managed to prove the purpose of such a withdrawal, it remains at the banks discretion whether they'll approve the transaction).
Let's break this down into multiple points:
- Suspiciousness of a 25k USD cash withdrawal
- Suspiciousness of a $25k USD electronic or check withdrawal
- Necessity to "prove the purpose" of any withdrawal
- Bank discretion and considerations regarding withdrawals
- Necessity of approval by the bank
I don't believe any of these five points are actually issues. As background, cash withdrawals within the USA are still very commonplace, as the country is fairly rather cash-centric when it comes to businesses, due in part to the lack of a system like Giro (Europe) that has both low, fixed transfer costs and can be sent or received by third-parties. The Federal Reserve's ACH system requires established relationships between accounts, whereas Giro does not. Debit card systems aren't a replacement for Giro either. Zelle (USA) is closer, but still isn't quite as full-fledged. Hence, businesses often deal in cash, pay employees in cash, and consumers pay other individuals in cash (eg buying an automobile).
To that end, for point 1, $25k as a cash withdrawal is not a daily occurrence but it does happen. I can't really think of ever paying for a private party used car by check, and such a cash-heavy transaction is often performed at the buyer's bank, so the seller is assured that the cash is good. In this setting, requesting to withdraw $25k cash is ordinary and mundane, if done very rarely. I doubt even prolific car buyers have this problem, but would be open to hearing evidence otherwise.
For point 2, electronic and check withdrawals have even less suspicion than cash, because they always leave traceable evidence. Money laundering concerns are reduced because the entire money trail can be reestablished later, whereas as cash can easily disappear or be "forgotten". To that end, the suspicion isn't about the cash amount but the source and destination. Even a $1 million check is not suspicious, if it's coming from a law firm's client account to a client's personal bank account. That is, again, a thing that happens fairly regularly. More down to earth, people can and do pay housing deposits by check, and property taxes are often drawn electronically. When one or both accounts to a transaction is prominent and established, there is a low probability of money laundering.
Point 3 is often though to be an issue, due to confusion about regulations for bank clerks on when to file a Suspicious Activity Report (SAR). Bank tellers are required to follow Federal Reserve regulations that aim to prevent abuse of the American financial system for money laundering. An SAR must be filled in whenever the teller: a) thinks money may be laundered, or b) the transaction is above the bank's or regulation's fixed amounts. The latter is often pegged at $10k, so this is where people think that it's disallowed to withdraw over $10k. This is not correct.
An SAR is something the teller fills in, and to do that, they might ask the customer some questions about the transaction. For the grand majority of people, the purpose is quite simple: cash purchase of a car, housing down payment, loan for a friend. Would the teller know if the customer is lying? Nope, not at all. But the SAR forms part of a trail of records, so that money laundering investigators can trace funds in the future. But note that the clerk can fill in an SAR for any type of transaction, including checks, and don't strictly need the customer's truthful answers (or any answers) anyway. An obligation to fill in an SAR does not prevent the transaction from going through. It's a speed bump, not a stop sign.
As for the actual stop signs, that's what point 4 covers. A bank obviously cannot allow a withdrawal if it would exceed the customer's balance, or if they don't physically have enough cash, or if the withdrawal is not authorized (ie not named on the account, or PIN not known), full stop. But other situations may arise where the withdrawal must be delayed, either for the bank's own convenience or because the account agreement specifically requires certain holdings times.
I quickly perused a random account agreement for Wells Fargo and the Available of Funds section describes that new accounts (less than 30 days old) will have elongated hold times for withdrawal against newly-deposited funds. This is applied in a first-in-first-out fashion, so only fully-draining the account would incur the longer hold time. In other cases, the bank may take more time but is required to inform you of that, and provide a definite date for when the withdrawal will clear. This verbiage does not distinguish cash vs non-cash, so they're within their rights to delay a check, as long as they obey their own agreement. If this is not tolerable, find a different bank.
Finally, this also gives us some insight into the default behavior for banks subject to Federal Reserve regulations, which is point 5. A bank may not deny a withdrawal of unencumbered, unheld funds (cash or otherwise), except when the bank has actual knowledge that the withdrawal definitely is for laundering. It is, after all, not their money: it belongs to the customer and they are just the regulated custodian of it. A bank can certainly advise a customer not to fall for a pig-butcherint scam, but they cannot block the customer from obtaining their own money back out. They can, as described earlier, apply a temporary, finite-time hold on the funds, but that's it.
To my knowledge, there is no Fed-regulated, FDIC/NCUA bank or credit union that requires pre-authorized approval to access a customer's own funds. I am open to hearing evidence to the contrary, but I don't believe such a thing exists. How would they even stay in business? To be clear from point 4, a bank can certainly ask for a few day's notice to prepare $50k in new $2 bills. But that's easy enough: just call the bank and verbally request the withdrawal, then collect it in-person days later.
Who is disadvantaged by this? Mostly money launderers and con artists trying to abscond with their scam proceeds. But I'd be remiss if I didn't also mention rich people that prefer to suddenly go on vacation and pay for everything in cash. But the system is designed to be no obstruction to those that plan ahead, or are dealing in such small amounts that it's not a big issue. Normal everyday people all share the costs of money laundering, so it's not fair to disadvantage them just so rich people and scammers aren't inconvenienced by their inability to plan ahead. They don't even have to plan ahead: just keep a few racks in the safe.
It is to me, frankly, a non-issue to withdraw money for me or anyone in the working or middle class, because the very issue of being "flagged by US banks" just rarely even a speed bump. And the rich folks have private banks that will gladly give them inordinate amounts of cash to spend.
What exactly is the problem here, specifically?
I'm deeply skeptical. If the whole premise is that natural language (which is how prompts to AI generation are given) is insufficiently precise to constrain output, and that formal specification is precise enough, then why are we bothering to use natural language generation? If this works as stated, then it's a tactic admission that the thing being checked is inherently flawed. Why not just use the formal specification to generate code then?
I suspect that two things are true: natural language is inherently insufficiently precise for nontrivial code generation, and also that formal specification is not broad enough to describe all the code which is being generated today using AI/LLMs.
Reliability of generated output was never an engineered goal for LLMs, and no amount of "reasoning by Lego" can fully compensate for this no matter how complex the mitigations in post. It's the same reason why safety (under any definition) cannot be "bolted on" to an LLM after the fact.