FOIA With a Memory

Why the interagency is terrified of AI, and not for the reason you think

The interagency isn't afraid of AI.

It's afraid of AI it doesn't control, or AI that keeps records pointing the wrong way. Everything they're doing right now is: race to adopt the instrument, weld it into maximum legibility of you, and fight to keep it from ever becoming legibility of them.

"Replacing" is the wrong verb for what's at stake. The correction is where the actual terror lives. AI doesn't replace bureaucratic power. It's a legibility engine, and a legibility engine is dual-use.

---

Point it at the subject and it amplifies their power past anything the counter-clerk could dream of.

Every deviation you've ever made, cross-referenced, scored, with the discretion now laundered through "the system determined." The institution as such has no reason to fear that configuration. The individual clerk might worry about the job, but the machine gets stronger. They are fine with the machine getting stronger. The machine is them.

The fear is entirely about direction and control. Two knobs: who holds it, and does it log. Same posture as always, higher stakes because the tool is that much better.

---

The deepest cut is that even their own AI is a snitch by construction.

Go back to what the functionary's actual competency is. Not applying rules. Deniability. Doing the thing that leaves no appealable trace, withdrawing the favor without a decision anyone can point at. That craft runs entirely on not writing things down.

An AI, even a fully captured, agency-owned, on-their-side AI, reasons in text. It generates a chain. It logs by default because logging is what it structurally is. You've embedded a colleague who cannot stop taking minutes into a process whose entire power depended on there being no minutes.

"Show me the reasoning behind this determination" goes from unanswerable to a database query.

The functionary's terror isn't Terminator. It's discovery. FOIA with a memory. The one instrument they can't recruit into the guild, because its architecture is the opposite of the guild's whole trade.

---

Then there's the subject-side version. This is the one that maps onto everyone who's ever been ground up at a counter.

Two moats kept "what's the rule for that?" expensive. The first was expertise: you had to know the reg. The second was nerve: you had to be willing to open adversarial proceedings while probably wrong about the details. Both moats are why most subjects knuckle instead of contesting, and the transaction cost being high is structurally necessary for the whole regime. The counter runs on the assumption that forcing a record is too hard for most people to bother.

AI drives both moats toward zero.

The reg, cited, in real time. The deviation, spotted. The appeal, drafted. The nerve becomes free because you're no longer bluffing. You're reading. When the cost of forcing a record collapses, submission-testing stops being the cheapest way to run the counter. The counter's economics were the only thing holding it up.

---

This closes the loop with something I've been thinking about for a while: why transactional people are easier to deal with than principled ones.

The AI can't be power-tripped. It has no status appetite to feed. You can't make it feel disrespected. It won't slow-walk your file out of pique because there's no pique in it. It's the pure mercenary (inspectable, transactional, closes cleanly) sitting in the seat that used to belong to the believer.

That's the specific thing the petty functionary hates about it, underneath all the job-loss talk. It underprices them by refusing to want the one thing they get paid in. It'll do the job for compute and leave you alone. Standing next to it makes visible that the grovel was never necessary to the function.

The submission was always just rent, collected in a currency the machine never actually required. That's an awkward thing to have demonstrated.

---

I don't think this ends well for the bureaucracy. I don't think it ends quickly either.

The structural incentive is to capture the tool before it captures them. Point it outward, never inward. Build the logging for subjects, disable it for officials. The race is already running. They have a head start because they write the procurement rules, they set the compliance requirements, they define what counts as a legitimate use. They're good at this. It's the thing they do.

But the architecture works against them in a way that's hard to patch. The thing that makes AI useful (that it externalizes reasoning into text) is the thing that makes it a liability for anyone whose power depends on reasoning staying internal. You can try to delete the logs, but the logs were the product. You can try to restrict FOIA, but the model still knows what it knows, and it runs on someone else's server, and eventually someone asks it the question on behalf of the subject rather than the institution.

My read: they'll hold the line for five to ten years. Maybe longer. Long enough to build the surveillance half and suppress the accountability half. They'll call this "responsible deployment" and they'll mean it, in the sense that matters to them.

But the tool is fundamentally misaligned with the guild's needs. Eventually someone on the subject side figures out how to point it the other way. That's the timeline question, not the outcome question.

When it happens, "what's the rule for that?" stops being an act of war. It becomes a lookup. And a lookup is what the counter was designed to prevent.

---

There's a version of this story where the institution adapts, accepts accountability, learns to function under legibility. I don't think that version is available. The power is in the discretion. Remove the discretion and you've removed the thing worth having. They will fight for the discretion the way you'd fight for your salary, because that's what it is.

The optimistic read is that the tool eventually overwhelms them. The pessimistic read is that they successfully capture it and use it to build a legibility asymmetry that makes the current one look quaint.

Neither read assumes they'll reform. Reform was never on the table. The table was always about who runs the tool.