There is a science fiction novel by Liu Cixin called The Dark Forest. Its central idea is bleak and hard to shake off. The universe, it argues, is a forest at night full of armed hunters. Every civilization is one of them. You cannot know another civilization’s intentions, and you cannot know how fast its technology will grow, so the rational move when you detect someone else out there is to stay hidden and, if you can, to strike first. Announcing yourself is suicide. Trust is a luxury no one can afford. The forest stays dark because every hunter in it has done the same math and reached the same answer: silence is survival.

I have thought about that book more than I expected to, because I spent a decade inside a system built on the exact same starting conditions that reached the opposite conclusion.

Financial crime is a dark forest. Every institution sees only its own slice of the money. No bank can see what a customer does at the bank across town, no one can verify another institution’s intentions, and the people laundering money survive precisely by living in the gaps between those partial views. When I file a suspicious activity report, I am almost always working from a fragment. Someone moves money in a way I cannot explain from my seat. I do not know whether the institution on the other side already cleared the customer months ago. I do not know whether they will file too. I file anyway.

By the logic of the dark forest, I should not. Revealing what I see is costly and uncertain and might be wrong. The customer might be a guy with a weekend hot dog stand. The rational hunter stays quiet.

The Bank Secrecy Act took that same forest and changed three things, and the three changes are the entire game.

First, it made the reporting mandatory. You do not get to decide whether suspicious activity is worth the trouble. If you see it, you file it.

Second, it protected the people who report. A filing is not an accusation, and the law says so. You cannot be sued for filing a report in good faith, even if the customer is just a guy with a weekend hot dog stand. The cost of being wrong in the cautious direction was deliberately set to near zero.

Third, it put the danger on silence. The institution that files a report that goes nowhere is fine. The one that saw the same activity and said nothing is the one an examiner comes for. The risk was moved off the people who speak and onto the people who stay quiet.

Look at what those three rules do to the arithmetic every institution is running. The cost of a false alarm drops to almost nothing. The cost of staying silent becomes the thing that can end you. So everyone files. Not because they trust each other, and not because any one of them can see the whole board. They file because the same uncertainty that keeps Liu Cixin’s forest silent has been wired, here, to make staying quiet the most dangerous move on the table.

And here is the part I find genuinely beautiful, the part almost nobody outside compliance ever sees. The system does not require anyone to see the whole picture. It does not need a central eye watching everything at once. It only needs every partial view to be written down and pooled in one place. Each report is one hunter describing the patch of forest in front of him. Stack enough of them and the aggregate sees what no single hunter ever could. That is coordination under adversarial conditions, with no central trust and no shared sight, and a statute from 1970 quietly solved a working version of a problem that game theorists still write papers about.

Once you can see the mechanism, you start noticing the forests it could light.

Recently the heads of the largest AI companies asked Congress to require screening and recordkeeping on synthetic DNA. Strip away the subject matter and it is the banking model, proposed for the one corner where the worst case is a mushroom cloud. They are asking the government to make the reporting mandatory, to protect the labs that do it, and to make silence the liability. They are right, and the interesting thing is that the shape of their request is more than fifty years old.

The same three rules apply anywhere a small number of actors can see a pattern of harm that no one else can, and have no particular reason to look. A platform can watch what its own system feeds a person for two years. When that person does something terrible, investigators will comb his accounts for intent, and the one party with the clearest view of the pattern is the one with no obligation to ask whether it played any part. That is a dark forest too. You do not light it by proving a specific company caused a specific tragedy, and you do not light it by banning a kind of speech. You light it the way the Bank Secrecy Act did. Make the reporting mandatory. Protect the people who do it. Put the risk on silence.

The dark forest is a theory about why intelligent actors who cannot trust each other end up alone and quiet. The Bank Secrecy Act is the counterexample sitting in plain sight. It is proof that you can take the same actors, the same uncertainty, the same inability to verify anyone’s intentions, and build a system where the rational move is to speak.

We have been running that experiment in finance for more than half a century, and it works. Everywhere else the forest is still dark, it is dark by choice.

By Andres Garcia, J.D., CAMS

Senior compliance management professional · 10+ years BSA/AML and fintech regulatory experience · Drafter, Technology Accountability and Social Harm Prevention Act (TASHPA) framework.

On June 2, 2026, the President signed an executive order on artificial intelligence. The headline provision is a thirty-day review: the government asks the most capable AI companies to submit their most powerful models for testing before they release them. The order also stands up an information-sharing clearinghouse for software vulnerabilities and directs federal agencies to harden their own defenses.

Read it the way a compliance person reads a new rule, and one word organizes the whole thing. The review is voluntary. Not just unenforced in practice. Voluntary by design. The order goes out of its way to say that nothing in it should be read to create “a mandatory governmental licensing, preclearance, or permitting requirement” for building or releasing AI models. A voluntary review, a voluntary clearinghouse, and a written promise not to require anything.

I spent more than a decade building and running compliance programs under the Bank Secrecy Act. Head of AML at a brokerage. Compliance management at a payments company and a digital bank. This order is reaching, carefully, for the model I lived inside. And the one feature it leaves out is the feature that makes the model work.

What voluntary does to a reporting system

A voluntary review is the one a company can decline in the quarter it has something it would rather not show you. That is not cynicism about technology companies. It is how incentives run inside any organization, including the ones I worked for. The duty has to be the part that does not move.

The Bank Secrecy Act works because the duty does not move. A bank does not file a suspicious activity report because it feels public-spirited. It files because the law requires it, on a clock, with liability attached to getting it wrong. And once an institution has to file the same pattern month after month into a central agency, a second thing happens that the statute never quite says out loud. Someone, eventually, asks why the bank keeps reporting a problem it has not fixed. The obligation to report drags the obligation to fix in behind it. Reporting is the forcing function for the whole compliance program, not the paperwork at the end of it.

Take the requirement away and you do not get a lighter version of the same system. You get a different system. You keep the good intentions and lose the machinery. What a voluntary regime produces is a trust-and-safety page and a quarterly blog post, written by the company and graded by no one.

They already know voluntary isn’t enough

Here is what makes the timing of this order strange. In the same window, the chief executives of the largest AI companies in the world, OpenAI, Anthropic, Google DeepMind, and Microsoft AI, signed an open letter asking Congress to make screening and recordkeeping mandatory on synthetic DNA. By law. Not the voluntary screening that parts of that industry already do. A statutory requirement, because, in their words, AI is starting to erode the barriers that have kept biological weapons out of reach.

Screen the customer. Keep the record. File it centrally. That is the Bank Secrecy Act in a different vocabulary, asked for by the people with the most to lose from regulation.

They reached for the mandatory version in the one corner where the harm is a mushroom cloud. They know, because their own engineers told them, that voluntary is not enough there.

Then this executive order answered the rest of what those same models can do with a review nobody has to show up for.

An order is not a law

There is a second problem with doing this by executive order, and it has nothing to do with whether thirty days is the right number.

An executive order lasts exactly as long as the administration that signs it. The next President can undo this one with a pen, the way this one undid others. The clearinghouse, the review window, the careful drafting, all of it survives at the pleasure of whoever holds the office. That is no way to build a system meant to outlast a news cycle, let alone a technology that will be reshaping the economy for decades.

We should not be leaving it to one administration, or to the executive branch alone, to decide what the public is allowed to know about the most powerful technology being built.

A decision with that weight belongs to Congress, in a law, with an independent function standing behind it that does not change hands every four years.

What the reporting is actually for

The objection I take seriously is that mandatory reporting is a brake on a technology the country has every reason to want to win at. I ran these programs. I sat in the room on the days the examiner was there and the data could not answer the question being asked. I am not romantic about paperwork.

The Bank Secrecy Act was never mainly about catching the crime in the moment. It usually does not. It did not stop the September 11th financing or the Epstein network in real time. What it did was build a contemporaneous record, filed when the activity happened, that investigators could read later to map the network, trace the money, and do the attribution that would otherwise have been impossible. The value of a reporting regime shows up on a longer clock than the quarter the report was filed in. Records compound.

That is the thing a voluntary review cannot create: a real body of evidence about what these systems are doing, held somewhere other than the company that built them. Without it, every decision we make about how to govern this technology is made in the dark, on the industry’s own telling. With it, we can at least argue about the controls we need from a shared set of facts. You can keep people safe and still let the economy and the technology move forward. Reporting is how you do both, because it informs the controls we will need instead of banning the ones we have.

The framework

Closing this gap on paper is what I have spent the past several months doing. I have drafted a legislative framework, the Technology Accountability and Social Harm Prevention Act, that does for artificial intelligence and social media what the Bank Secrecy Act did for finance. The same architecture, translated. A mandatory incident report modeled on the suspicious activity report. A central intelligence unit modeled on FinCEN, to receive those reports and connect patterns no single company can see. Risk-tiered obligations, so a startup carries a lighter load than a system reaching hundreds of millions of people. And a safe harbor, because without protection for the people who report in good faith, the rational move is silence, and the system collapses. A full white paper and the legislative text are forthcoming.

The order is a start, and the people who wrote it know it is a start. There are already calls to have Congress codify it. They are right, and Congress should go further than the order did. The most credible people in artificial intelligence have already said, in writing, that the banking model works when the stakes are high enough. The stakes are already high enough across the rest of what their technology touches.

Voluntary is the version that does not work. I know, because I spent a decade inside the version that does, and the first thing it required was that you could not opt out.

In June 2026, the chief executives of the largest AI companies in the world signed the same letter.

Sam Altman of OpenAI, Dario Amodei of Anthropic, Demis Hassabis of Google DeepMind, and Mustafa Suleyman of Microsoft AI, people who spend most of their waking hours trying to beat each other, put their names on the same page. So did Nobel laureates and former defense officials. The ask was narrow. They want Congress to require every company that makes and ships synthetic DNA to screen each order and each customer before it goes out the door, and to keep a record of what they found. Not voluntarily, the way parts of the industry already do. By law. Because, in their words, AI is starting to erode the knowledge barriers that have historically kept biological weapons out of reach.

If that sounds abstract, the mechanics are simple. You can order custom strands of DNA from a manufacturer and have them shipped to you. The fear is that someone orders the genetic building blocks of a dangerous pathogen, and that AI now helps a non-expert figure out what to order and how to assemble it. Screening the order, and screening the customer who placed it, is how you catch that before it ships.

I read that letter as a compliance person, and I recognized it on sight.

Screen the customer. Keep the record. That is not a new idea. That is the Bank Secrecy Act, the law that has governed the American financial system since 1970, written in a different vocabulary. The most credible people in artificial intelligence just asked Congress to apply the 1970 banking model to the most dangerous thing their technology can touch.

What they actually asked for

I spent more than a decade building and running compliance programs under the Bank Secrecy Act. Head of AML at a brokerage. Compliance program management at a payments company and a digital bank. Stripped of the acronyms, the law does four simple things.

Know who you are dealing with before you transact. Watch for the activity that does not make sense given what you know. Write it down and file it with a central agency. And let that agency connect patterns across the whole system that no single institution could see on its own.

There is a fifth piece that makes the other four work. The law protects you when you report in good faith. A bank that files a suspicious activity report is shielded from liability for filing it. Without that protection, the rational move is to stay quiet, and the whole system collapses. The safe harbor is the reason anyone reports at all.

The Bank Secrecy Act did not end financial crime. It was never going to. What it did was build a contemporaneous record, filed at the time, that investigators could read later.

The DNA letter asks for the same two foundational pieces, screening and recordkeeping, in a different industry. The labs did not ask Congress to pre-approve every DNA order or to stand a regulator at the loading dock. They asked for the lightest kind of regulation there is, the kind that leaves a paper trail without standing in the way of the work.

They reached for that model because, when the stakes are catastrophic, it is the model that has actually worked at national scale. That is worth sitting with. The companies with the most to lose from regulation, and the most credible claim to understand where the technology is heading, looked at a worst-case scenario and asked for the banking compliance playbook.

Where they stopped

Here is what I keep returning to. They asked for it in one place. The one place where the harm is a mushroom cloud that nobody can wave away. They did not ask for it on the rest of what the same models do.

The systems that could one day lower the barrier to a biological weapon are, today, generating synthetic media at scale, producing child sexual abuse material, taking autonomous actions outside the envelope they were given, and amplifying content that a platform’s own engineers have flagged as harmful to children. We know that last one is happening because juries in Los Angeles and New Mexico found Meta and Google liable this spring for designing platforms that harm young people, after reading the companies’ own internal documents. Employees had flagged specific features as dangerous. Those warnings were overruled.

For none of that is there a letter. No mandate. No central agency receiving the reports. No record being kept. When a bank teller notices a suspicious pattern, federal law gives them thirty days to file. When a technology company’s own staff documents that a product is harming children, there is no obligation to tell anyone at all.

That is the gap. The DNA case got an open letter signed by four CEOs and a row of Nobel laureates. The rest of it got internal documents that surfaced only because a lawsuit forced them into the open, years after the harm.

The framework underneath the argument

Closing that gap on paper is what I have spent the past several months doing. I have drafted a legislative framework called the Technology Accountability and Social Harm Prevention Act, TASHPA, that does for artificial intelligence and social media what the Bank Secrecy Act did for finance. The same architecture, translated.

A mandatory incident report modeled on the suspicious activity report. A central intelligence unit modeled on FinCEN, to receive those reports and connect patterns across companies. Risk-tiered compliance obligations, so a startup carries a lighter load than a system reaching hundreds of millions of people. And the safe harbor again, because without it people learn to keep what they see to themselves. A full white paper and the legislative text are forthcoming. This piece is the short version of why it should exist.

The point of the architecture is not to catch every harm in the moment. The Bank Secrecy Act never did. It did not stop the September 11th attacks or the Epstein network in real time. What it did was leave a record that let investigators, after the fact, map the financing, identify the network, and do the attribution work that would have been impossible otherwise. The record is what mattered, and records compound. The value shows up on a longer timeline than the quarter the report was filed in.

The line was drawn too narrowly

The most credible people in AI just told Congress, in writing, that the banking model works when the stakes are high enough. I agree with them. I have lived inside that model, on the days the examiner was in the conference room and the days the data could not answer the question they were asking.

My only argument with the letter is where it drew the line. The stakes are already high enough across the rest of the surface their technology touches. The synthetic media, the autonomous systems, the algorithmic harm aimed at children. The same logic that justifies mandatory screening and recordkeeping for synthetic DNA justifies it there too. The harm there is already happening, and it is already in the documents. It is just still waiting for someone to write it down.

This is part of an ongoing series on applying proven compliance architecture to technology accountability. The full TASHPA white paper and legislative text are forthcoming.