Regulating tech companies like banks

Banks have to monitor the transactions that go through them. If they see something suspicious, they report it to the government. If a customer is a known bad actor, they stop doing business with them. And if a bank does not have a working program to do these things, the fines from the Financial Crimes Enforcement Network (FinCEN) are routinely nine digits.

I think this model should be applied to other industries, in particular social media, AI providers, and healthcare.¹

The FinCEN model

Under the Bank Secrecy Act, financial institutions must:

Monitor: Know who your customers are, and run transaction monitoring to flag suspicious activity.
Stop bad actors: Block transactions with sanctioned individuals, and close accounts clearly being used for crime.
Report: File a Suspicious Activity Report (SAR) within 30 days for suspicious activity of $5,000 or more. Institutions file over 4 million SARs a year.

The important part is what gets penalized. Banks are not fined because one launderer slipped through (it is hard to catch everything). They are fined because their program to catch launderers was inadequate. FinCEN fined TD Bank $1.3 billion in 2024 because its anti-money laundering program was not designed or staffed for the risks it faced, including not monitoring Zelle and Venmo transactions indicative of human trafficking. Binance paid $3.4 billion after never filing a single SAR.

This makes the regulation preventative. No victim needs to come forward. Examiners check whether the program works, and the fine scales with how badly it fails.

Social media platforms are marketplaces for drugs, guns, and counterfeit pills, and are used for sex trafficking and child sexual exploitation.

The reporting piece already exists. Under 18 U.S.C. 2258A, platforms must report child sexual abuse material to NCMEC, and the 2024 REPORT Act added child sex trafficking and enticement. But the law explicitly says platforms are not required to monitor their users, and fines for failing to report max out at $1 million for the largest platforms.

Compare that to testimony unsealed in 2025 that Meta had a 17-strike policy for accounts engaged in sex trafficking (Meta says it now uses one strike). That is the same failure as Binance – a policy decision not to act on known bad actors. Under a FinCEN style regime, that is a nine or ten digit fine.

I would keep a US version narrow: clearly illegal transactions, not misinformation or other speech.

AI providers: monitor usage, not training

Much of the AI safety debate is about how models are trained. I think it is easier to regulate how the tools are used. We do not know how to write a rule that specifies “safe training”. We do know how to write a rule that says “monitor usage for these behaviors, ban the accounts, and report them.”

The frontier labs already do this voluntarily. In August 2025 Anthropic reported a criminal who used Claude Code to automate data theft and extortion against at least 17 organizations. Anthropic banned the accounts, built a classifier to detect similar behavior, and shared indicators with authorities. Monitor, stop, report. The problem is it is voluntary. California’s SB 53 requires reporting critical safety incidents, but the scope is narrow (catastrophic events) and fines are capped at $1 million.

A FinCEN version would require know-your-customer for API usage at scale, monitoring for specific misuse categories (malware, weapons, CSAM, fraud), suspicious usage reports to a central agency, and examiners who test the monitoring. The limitation is open weight models – if someone runs a model locally, no one is watching. Given that most of the frontier open weight models need to be hosted though, this is not a serious concern (although it can only be relevant for frontier models hosted by US companies).

Healthcare and personal data

I am getting tired of receiving another 12 months of free credit monitoring every time my data is breached. And I would bet each of those companies had a clean SOC 2 report. This applies to any company holding personal information, but healthcare is the clearest example.

Healthcare organizations reported breaches exposing around 275 million records in 2024. Most of that was Change Healthcare, where attackers logged into a remote access portal with stolen credentials (no multi-factor authentication), moved around the network for over a week, and took data on roughly 190 million people. That was not a sophisticated attack. It was a missing basic control, and no one noticing the intruders.

HIPAA requires audit controls and review of system activity, but in practice compliance is a documentation exercise. The largest HIPAA settlement, as far as I can tell, is still Anthem’s $16 million for a breach of nearly 79 million people. TD Bank paid $1.3 billion.

The FinCEN analog is to require active monitoring for data exfiltration. Flag anomalous access and egress – a customer support account pulling millions of records, a large outbound transfer at 2 AM – and respond in hours, not the 241 days it takes on average to identify and contain a breach. Fine companies for not having a working program, whether or not a breach happened.

Why not just use lawsuits?

Judge Glock recently argued in City Journal that tort liability is a better way to manage AI risk than new bureaucracies, since it would “put the onus on companies themselves to game out dangerous scenarios.” I agree with the goal, but I think regulation is the better mechanism:

Lawsuits are after the fact. You need a victim. The FinCEN model penalizes the missing program before anyone is harmed.
Juries are fickle. Peter Huber’s Galileo’s Revenge documents juries rewarding junk science. Bendectin, a morning sickness drug, was pulled from the market in 1983 due to litigation costs, despite the scientific consensus that it did not cause birth defects.
You cannot choose which lawsuits get filed. Glock himself criticizes suits against social media companies over addictiveness. A tort regime invites exactly those suits. A regulator can specify what to monitor.
Social media is mostly immune anyway. Section 230 shields platforms from most claims based on user content.

Limitations

The Bank Secrecy Act is not perfect. Compliance is expensive, banks file defensive SARs no one reads, and there is little evidence on how much it reduces crime. Monitoring messages or AI prompts is also more invasive than monitoring bank transactions. So the scope needs to be narrow – illegal sales and trafficking for social media, specific misuse categories for AI, and data access (not content) for healthcare – with size thresholds so small companies are not crushed by compliance costs.

Even so, I would rather have a known set of rules companies are examined against than wait for the next breach or AI enabled attack, and hope a jury gets it right.

If your organization needs help building monitoring or anomaly detection systems, get in touch.