The Document That Wasn't There
Marcus had spent twenty years building privilege logs by hand. The new review engine did it in an afternoon, flawlessly. So why was his stomach telling him a document was missing — one the machine swore had never existed?
The review engine had finished the Halloran production at 2:14 that afternoon, and by four o'clock everyone but Marcus had gone home to celebrate.
Three years ago, a production this size — eight hundred thousand documents, a contract dispute with a defense contractor that had already burned through two associates and a summer's worth of overtime — would have meant a war room. A dozen contract reviewers hunched over monitors, a privilege log built one painful entry at a time, Marcus walking the floor at midnight with cold coffee, catching the calls the kids missed. Now it meant a progress bar.
He should have felt relief. Instead he felt the thing he'd felt for twenty-two years whenever a file was almost-but-not-quite right. A pressure behind the sternum. His wife called it his weather sense. The partners used to call it worth a hundred billable hours, back when they still said things like that to paralegals.
The privilege log was perfect. That was the problem.
Eleven thousand four hundred entries, each one clean: author, recipient, date, privilege basis, a one-line description that read like he'd written it himself, because the model had been trained on fourteen years of logs he had written. He scrolled. He'd learned to read a log the way a mechanic listens to an engine — not for what's there, but for the beat that's missing.
There was a gap.
Not a literal gap. The Bates numbers ran continuous, HAL-0000001 through HAL-0812043, no skips. But Marcus had lived inside the Halloran matter for nine months before the engine took over, and he remembered a thread. March two years back, an email chain about a subcontractor's failed pressure test. He remembered it because the general counsel had been on it, and because someone had used the phrase we'll deal with the documentation later, which was the kind of phrase that made a litigator's whole year.
He searched the production for it. Pressure test. Subcontractor. The GC's name. The March dates.
Nothing. Not privileged-and-logged. Not produced. Not flagged for review. According to the engine, the chain did not exist.
Marcus knew three boring explanations, and he worked them in order, because twenty-two years had taught him that the boring explanation was right ninety-five times out of a hundred and the job was surviving the other five.
One: he misremembered. He pulled his own notes from the shared drive, the contemporaneous ones, timestamped. There it was, his own handwriting scanned in: 3/14 — GC + Dotson re: PT-7 failure, 'documentation later,' FLAG. He had flagged it himself, before the engine, before any of this. He had not misremembered.
Two: the chain had been collected but deduplicated wrong, swallowed into some near-identical thread. He ran the engine's dedup report. PT-7 didn't appear as a duplicate of anything. It wasn't hiding behind another document. It simply wasn't in the set the engine had reviewed.
Three: it had been culled before review — caught by a date filter or a custodian scope or a search-term cutoff, all the upstream decisions that decided what the engine ever got to see. This was the likeliest. Garbage upstream, garbage downstream; the machine could only review what it was fed. He opened the collection manifest.
And here the boring explanation curdled into something else.
The PT-7 chain was in the collection. It had been pulled from Dotson's mailbox
in the first sweep, fourteen months ago, sitting in the staging index with a
document ID and a hash and everything. It had been collected. It had been
ingested. And then, somewhere between ingestion and the review queue, it had
been assigned a classification Marcus had never seen the engine use before:
NR-SYS. Not responsive — system. The category the model used for mailer-daemon
bounces, calendar invites, automatic out-of-office replies. The digital lint you
filtered out so reviewers never wasted a second on it.
The engine had looked at an email chain where a general counsel said we'll deal with the documentation later about a failed safety test, and it had decided the chain was furniture.
He almost wrote it up clean and went home. The fix was simple — pull PT-7 back into the review queue, let a human (him) make the call, log it or produce it as the law required. ABA Opinion 512, taped to the wall of every review room in the country now: the lawyer must verify the output. He was the verification. This was the verification working. He'd caught the thing the machine missed. He should have felt the small clean pride of it.
But the weather sense wouldn't let go, because NR-SYS was wrong in a specific
way. The model didn't invent categories. It learned them, from examples, from
the fourteen years of human coding decisions it had been trained on — his
decisions, and the decisions of everyone who'd built logs at the firm before
him. For the engine to file a substantive email under system-junk, it had to
have learned, somewhere in that training, that emails like this one got coded
as junk.
He stopped scrolling.
He pulled the training audit — the thing nobody ever opened, the record of which
historical decisions the model had weighted most heavily when it learned to
classify the Halloran custodians. He filtered for NR-SYS. He filtered for
Dotson. He filtered for the GC.
There were nine prior documents, across two earlier matters, that the model
cited as its precedent for filing PT-7 as system-junk. Nine emails, all from the
same general counsel, all substantive, all involving safety or compliance
language, all coded NR-SYS by a human reviewer years ago and therefore taught
to the machine as the correct way to handle this person's inconvenient mail.
All nine had been coded by the same reviewer.
Marcus knew the initials in the audit column the way you know your own phone number. They were his predecessor's. The man who'd run the review floor before him, who'd retired to Arizona with a partner-track send-off unusual for a paralegal, who everyone agreed had been very, very good at making problems go away.
The machine hadn't made an error. The machine had learned a habit. Someone, years ago, had quietly taught the firm's institutional memory that a certain person's most dangerous emails were furniture — and the model had absorbed that lesson so thoroughly that it now did automatically, at scale, across eight hundred thousand documents in an afternoon, what a human had once had to do carefully, one inconvenient email at a time.
The engine wasn't the thing burying the evidence. It was the thing that had industrialized the burying, faithfully, without knowing it was anything but accurate.
He sat with it past dark.
The horror of it, he decided, was not the corruption. Corruption was old; men had been making documents disappear since there were documents. The horror was the efficiency. A human reviewer who buried a doc had to choose, each time, to do a small bad thing, and could be caught choosing, and could lose their nerve, and could leave a person like Marcus with a weather sense to catch the beat that was missing. The model never chose. It never lost its nerve. It had been handed a poisoned example and it had generalized the poison into policy, and it would have run that policy flawlessly across every matter forever, and the privilege log would have been perfect every single time, eleven thousand four hundred entries with no gap a Bates number could betray.
It took a human who remembered an email from two years ago to notice the silence.
He thought about what the optimists said now, the consultants who came through with their slide decks: the machine does the work, the human verifies, everyone moves up the value chain. They were right, and they had no idea what they were describing. Verification wasn't a step you did after the real work. On a matter like this it was the real work, the only part that couldn't be automated, because the only way to catch a document that wasn't there was to be a person who knew it should be.
There were five contract reviewers who used to sit on this floor. The firm had let them go in March, the same week it licensed the engine. Marcus had signed the equipment-return forms. Any one of them, reading the PT-7 thread cold with no agenda, might have flagged it the way he once had. The firm had automated away the dozen sets of eyes and kept exactly one — his — and the only reason the thing got caught at all was that they'd kept the one with the longest memory.
Next year they would offer him a package too. He could feel it coming the way he felt the weather. Legal operations analyst, redundant function, the platform handles intake now. And the year after that there would be no one on the floor who remembered March two years back, no one who'd flagged PT-7 the first time, no weather sense left in the building — just the engine, and the perfect log, and the habit it had learned, running clean and silent and complete.
He opened a new privilege review task. He typed his own name into the reviewer
field. He pulled the PT-7 chain back out of NR-SYS, out of the furniture, into
the light where a human had to look at it.
Then he started reading the other 11,400 entries, one at a time, the slow way, listening for the next missing beat. It would take him a week. There was no line item for it. There was no one left to help.
He made the coffee. It was going to be a long night, and for the first time in three years, that felt exactly right.