Understanding Without Reliance

Miriam Adler will not put client information into an AI system and will not rely on what one produces. She refuses the more comfortable escape too: ignorance.

Reasonable Doubts · 18 chapters · ~2 min each

Cast →

Chapter One (~2 minutes)

The Undertaking

The memo Miriam Adler sent the partnership was one sentence long. She would not enter client information into any external AI system and would not rely on the output of one; she would, nonetheless, undertake to understand these systems as well as anyone at Harlan Cross who used them daily. Douglas Frame read it twice and told her she wanted credit for both things at once: the halo of the abstainer and the authority of the adept.

She had expected the charge and did not much mind it. What interested her was the assumption beneath it, that understanding and use were a single act, so that declining the second must mean forfeiting the first. She thought that was false. She also thought the profession had begun, without quite admitting it, to punish people for believing it, treating the lawyer who did not prompt as the lawyer who could not think about prompting.

That was the confusion she meant to take apart. The duty had, after all, already hardened. Within a few short years the standard had moved from a genial suggestion that lawyers might benefit from technological fluency to a settled expectation that they possess it.¹ Opinions and guidance now spoke of competence in and supervision of these tools as an obligation owed to the client, not a preference indulged by the enthusiast.² Nothing in any of it said the obligation fell only on users. A lawyer who advised on a contract drafted by a model, or cross-examined a witness whose exhibit was generated by one, was practicing in the world these systems had made whether or not she had ever opened one.

"So learn to use them," Douglas said. "It's the same afternoon's work."

"It isn't," she said. "And that's the interesting part."

Her position had two halves, and she was careful to keep them apart in front of him. The first half was a refusal, and she would defend it later, item by item, on grounds she considered sound. The second half was an undertaking, and it conceded everything he wanted conceded: that the duty was real, that it was continuing rather than discharged by a single seminar, that it would have to be renewed as the tools changed under her.³ She was not claiming exemption. She was claiming that the duty could be met by a route he had not considered because he had never needed it, the route of someone who had to understand a thing precisely because she would never lean on it.

"You think you can know it without touching it," he said.

"I think 'touching it' is at least three different things," she said, "and I intend to find out which of them I actually object to. When I've finished I'll sit whatever examination you like. If I fail, I'll use the tools. If I pass, you'll stop calling it posturing."

He put out his hand and she shook it, which was how the wager became a project, and the project became the year that follows.

What she did not say, because it would have sounded like the very vanity he accused her of, was that she suspected the exercise would end by indicting her own practice as much as his. A standard she was about to apply to every model in the building would, if she were honest, have to be applied to the yellow legal pads and the trusted junior and the citation she had always "checked" by recognizing the name. But that was Chapter Sixteen's problem. For now there was only the undertaking, and the pleasant, ominous sense of having promised something she did not yet know how to keep.

Chapter Two (~2 minutes)

What She Is Actually Refusing

Before she could learn anything she had to know what she was declining, and she was surprised, sitting down to inventory it, how much of her refusal she had never actually examined. It had the texture of principle but the origin, in places, of habit.

She wrote the reasons out and sorted them into five piles: confidentiality, reliability, governance, cost, and proportionality. The confidentiality pile was the largest and the firmest. She would not put a client's facts into a system whose retention she could not see, whose sub-processors she could not name, and whose training use she could not switch off; she knew enough already to know that vendor policies on all three "varied wildly," and that redaction after the fact was no cure for disclosure at the point of entry.¹ That objection she was content to keep.

The reliability pile was firm in a different way. She distrusted the output not because it was often wrong but because it was fluently wrong: confident where it should be tentative, and tuned, she suspected, to please.² A tool that emphasized the facts flattering to a hypothesis and softened the ones against it was, to a litigator, not a neutral instrument but an advocate of unknown loyalty.

The other three piles were where honesty cost her something. Some of what she had been calling principle was really an unexamined estimate of risk, and when she made herself grade the estimate properly, sorting the uses by how much damage a failure would actually do, several of her refusals turned out to be refusals of high-risk uses only, which she had quietly generalized into a refusal of all of them.³ Governance was similar: her instinct that the tools created an unmanageable sprawl of obligations was sound in the abstract, but it described a firm without policies, training, or controls, and was not really an argument about the technology at all.⁴ And a fair number of her objections, she was forced to see, were not objections to AI but to a product, a specific vendor whose security posture she had once found wanting and thereafter treated as the character of the whole field.⁵

"You're talking yourself out of it," Douglas said, hopeful.

"I'm talking myself into precision," she said. "Which is the opposite. A blanket refusal is lazy. It can't tell you the one case where use would actually be fine, and it can't defend the cases where refusal is right, because it never did the work of separating them."

That was the discovery of the chapter, and it unsettled her more than any concession would have. Her refusal, tested, did not dissolve. But it shrank, hardened, and acquired addresses. It was no longer "I don't use AI." It was "I will not disclose privileged facts to an unvetted external system, and I will not rely on an unverified generation in work that leaves this office," two sentences she could defend to a regulator, a client, or a court, precisely because she could now say what they did not cover.

Chapter Three (~2 minutes)

Three Things Called Use

The move that unlocked everything came from a student's offhand question. Priya Nair, the second-year down the hall, asked whether Miriam objected to using the tool or to the tool having her client's file, and looked puzzled when Miriam treated these as the same thing.

They were not the same thing. Pressed on it that evening, Miriam found that the single word use was doing the work of three separate acts, and that she had been refusing all three because she objected to two. There was disclosure: putting client information into the system. There was reliance: taking the system's output into her work and standing behind it. And there was operation: the mechanical act of driving the thing, typing and reading what came back and watching how it behaved. Her objections, she realized, attached with full force to the first two and not at all to the third.

The distinction held up under pressure because the tools themselves were built around it. Whether her words ever reached a vendor's servers at all was a function of how the model was reached: a public web portal, an enterprise account with retention disabled, an on-premise deployment where nothing left the building, a sovereign cloud pinned to a jurisdiction.¹ Disclosure was not an inherent property of "using AI"; it was a property of a particular channel, and channels could be chosen or refused. Likewise, what the machine "knew" of her matter was simply what she put in front of it, the context she supplied rather than some ambient awareness, which meant she could operate a system meaningfully on facts that were entirely invented and disclose nothing real at all.²

Reliance came apart just as cleanly once she stopped treating trust as a switch. Trust in a machine's output was not on or off; it was a graduated judgment, calibrated to the stakes, earned or withheld task by task, and a thing she could study in others without ever extending it herself.³ She could watch precisely how much confidence a piece of work would bear without placing a gram of her own on it.

What made the operation defensible was a fact about the systems she had, oddly, found reassuring: their behavior was governed by hidden instructions and a strict internal chain of command, so that a model was not an open field but a constrained instrument whose limits could be mapped by anyone willing to probe them.⁴ One could learn the shape of a cage without living in it.

"So you'll use it after all," Priya said, pleased.

"I'll operate it," Miriam said. "On nothing that matters, relying on nothing it says. That's not a smaller version of what you do. It's a different verb."

Priya plainly thought this was a distinction with no difference, a lawyer's trick to keep a vow while breaking it. Miriam let her think so. The proof would be in whether the verb could carry the weight of real understanding, and that could not be settled by argument, only by trying it, which is what the rest of the year was for. But the road was open now. Everything her refusal forbade lived in disclosure and reliance. Understanding lived, it turned out, mostly in operation, and operation she had never actually refused. She had simply never separated it out to see.

Chapter Four (~2 minutes)

Reading Is Not Enough

She began, being a lawyer, by reading, and she got further than she had any right to expect. A weekend with the literature gave her a working vocabulary and, more valuable, a correction to the metaphors she had absorbed by osmosis. The machine did not "look things up" and did not "know" in any sense she would accept in a witness.¹ It generated text by predicting, one token at a time, the statistically likeliest continuation of what came before, converting language to numbers, moving those numbers through layered mathematics, and translating the result back into words, with a deliberate seasoning of randomness that meant the same question could yield two different, equally "valid" answers.²

That single fact reorganized her intuitions. A system whose fluency was a property of prediction, not of retrieval, would be most dangerous exactly where it was most confident and least grounded, a proposition she had felt as suspicion and could now hold as mechanism. She read on, and watched the field's own history rearrange itself: from single-turn novelty to persistent assistants that carried context, then to tool-using systems that reached outside themselves and acted, a trajectory that told her the thing she was studying would not sit still to be learned.³

For a fortnight she thought reading might be enough. It was orderly, it was safe, and it flattered the part of her that believed anything worth knowing had been written down by someone careful. Then Ronan Pierce, the general counsel at Meridian Freight, asked her a question over lunch that her reading could not answer. He wanted to know whether the model his logistics team relied on would behave differently on the day its vendor pushed an update, and whether, in effect, the thing he had tested in March was the thing running in September. She knew the words for every part of his question and could not answer it. She had read that outputs varied; she had never felt a system change under her hands, and so she could not tell him what the change would look like or how he might catch it.

That was the limit, and it arrived precisely where the profession's own guidance said the duty lived: not in reciting what a model was, but in understanding its capabilities and limitations well enough to protect a client who depended on it.⁴ Secondhand knowledge gave her the map. It did not give her the territory, and Pierce's question was a piece of territory. She had been reading about swimming.

"You've hit the wall every autodidact hits," Priya said, not unkindly. "The book stops exactly where the thing gets interesting."

She was right, and Miriam disliked how right she was. There was a knowledge that lived only in contact: in watching the system hesitate, in seeing which prompts it mishandled, in the tempo at which a plausible error slid past a busy reader. None of it required her to disclose a fact or trust a line. All of it required her to move from the page to the screen. Reading had taken her to the edge of understanding and then, politely, declined to take her any further.

Chapter Five (~2 minutes)

The Observed Session

So she asked to watch. Devin Cho, a mid-level associate who used the tools without ceremony, agreed to let her sit at his shoulder through an afternoon's work, on the condition that she not "hover like a supervisor." She hovered anyway, but silently, and learned more in three hours than in three weekends.

The first thing she learned was that the interface was not the thing. What she had pictured as a search box was in fact a place where work was staged, where a task was set up, revised, and carried across several exchanges rather than answered in one.¹ Devin rarely asked a question outright. He built toward it.

The second thing she learned was that his skill was almost entirely in the asking. When he wanted something usable he was lavish with specificity, stating the role he wanted the model to take, the audience, the format, the length, and the reasoning he wanted shown, and the output that came back was correspondingly sharp. When he was lazy, the machine was lazy with him.² She had assumed the competence lived in the model. A good deal of it plainly lived in the operator, in the ordinary discipline of saying exactly what one meant.

The third thing she learned was that a hard task was not one question but a sequence. Devin would establish facts in one exchange, test a framing in the next, and only then ask for the thing he actually wanted, each step standing on the last.³ It looked, disconcertingly, like the way she built a cross-examination: no single decisive question, but a staircase of small ones, each foreclosing an escape.

What held her attention longest was watching him use it as something other than an answer-machine. He ran a scenario through it (if we take position A, how does the other side respond, and what does that cost us) not to be told the answer but to map the branches, the way a strategist games a board before committing a piece.⁴ The value was not in the model's conclusion, which he ignored, but in the speed with which it laid out terrain he would otherwise have covered slowly and alone.

"You're not believing any of it," she observed.

"God, no," Devin said. "It's a sparring partner. It throws punches so I can see the openings. I'd never let it write the brief."

That sentence stayed with her, because it described, exactly, a form of value that survived her refusal intact. He was operating the system and declining to rely on it, her own distinction, arrived at from the opposite direction by a man who would have been baffled to hear it stated so formally. He got the benefit of the terrain-mapping and threw away the output, and the throwing-away was not a loss but the point.

She went home and wrote a single line in her notebook: the understanding is in the operating, and the operating is mostly a way of asking. She did not yet know how to ask well. But she had seen it done, close up, by someone with no stake in impressing her, and she now knew what she was trying to acquire. It had a shape. The next step was to put her own hands on the controls, still touching nothing real and still trusting nothing said, and find out whether the asking could be learned or only watched.

Chapter Six (~2 minutes)

Directed Operation

The intermediate step, before she would touch a keyboard herself, was to operate by proxy: she would instruct, and Priya would type. It satisfied the letter of her position, since she disclosed nothing and relied on nothing, and it isolated the one skill she had identified as the seat of understanding, the framing of the request, from the mechanics she did not yet trust herself to perform.

The room argued, briefly and enjoyably, about whether she was now "using" the tool. Douglas said she obviously was; Priya said a person who dictates a letter is not the typist. Miriam let the argument run because it was, underneath, the whole thesis being tested in miniature, and then she settled it the only way that mattered: by getting to work.

Instructing turned out to be harder than watching had made it look. Her first attempts were the attempts of a novice who had read about expertise. She over-specified where she should have trusted the model's range and under-specified where it needed a rail. Priya, gently, showed her the moves she had only read about: assigning the model a role to shift its whole register, handing it a worked example so it matched a pattern rather than guessing at one, asking it to reason step by step before it committed to an answer.¹ Each move changed the output visibly, and Miriam began to feel the thing she had wanted to feel, the causal link between how she asked and what came back.

Then came the shift that reorganized her sense of the whole enterprise. She had been thinking of each instruction as a fresh performance, a good sentence to be composed each time. Priya showed her that the real practitioners had stopped doing that, and no longer crafted a perfect prompt per task but built a stable, reusable context the model worked within, so that the intelligence lived in the standing setup rather than the individual request.² Understanding a tool, then, was not learning to write clever prompts. It was learning to design the environment a model reasoned inside, which was an architectural skill rather than a rhetorical one.

For legal work the discipline tightened further. When they framed a research task, the instruction had to carry its own guardrails, demanding the jurisdiction, insisting on source validation, forbidding the invention of authority, because the cost of a fluent falsehood was, for her, measured in sanctions.³ She saw that a well-built legal instruction was itself a small act of verification, front-loaded into the asking.

Underneath all of it she recognized a paradigm she could respect: the treatment of plain language as something executable, where an outcome was specified precisely enough in objective, constraints and stopping conditions that the words became an instruction a system could act on with little rework.⁴ It was, she thought, not so far from drafting a good order. Ambiguity was the enemy in both.

By the end she could feel the difference between an instruction that would work and one that would not before Priya pressed enter, the same intuition she had for a question that would land in cross before she asked it. The seat of it, she decided, was neither in the model nor in the operator alone but in the fit between the request and the material, which was exactly the thing a well-designed context was for: deciding, in advance, what the model should attend to and what it should ignore.⁵

"You could do this yourself now," Priya said.

"I'm about to," Miriam said. "On a matter that doesn't exist, so that when I break it, no one is harmed but two people who were never born."

Chapter Seven (~2 minutes)

The Sandbox

She opened an account under a scratch email, populated with a matter that did not exist: Volk v. Arda, a supply dispute between two companies she invented over coffee, with facts elaborate enough to exercise the system and false enough to disclose nothing. It was, she noted with some satisfaction, the cleanest possible expression of her position: full operation, zero disclosure, zero reliance. She could break anything in here and no one would be harmed.

The first hour was pure novelty, and she let it be. There was a real threshold to cross, the ordinary hands-on confidence of a first genuine interaction and the small competence of setting a task and getting something usable back, and she had denied herself it so long that crossing it felt faintly transgressive.¹ Then she settled and began to probe the edges she had only read about.

She learned the walls by hitting them. When she pasted the whole of her invented file plus a decade of fabricated correspondence, the system's grip on the earliest facts visibly loosened as the later ones arrived. There was a finite window of attention, and past its limit things fell quietly off the back.² She learned, too, that it did not carry anything between sessions on its own: close the tab and Volk v. Arda was gone from its mind entirely, unless she deliberately rebuilt the context each time or gave it a standing memory to hold.³ This she found reassuring rather than otherwise. A tool that forgot by default was a tool whose confidentiality risk lived at the door rather than in some permanent hidden ledger, which was exactly the architecture her refusal cared about.

She experimented with shaping it to her preferences, setting standing instructions so it stopped padding its answers and adopted the clipped register she liked, and watched a generic instrument become, in small ways, hers.⁴ And she used it, finally, the way Devin had: as a generator of first options rather than final answers, throwing a dozen possible framings of the Volk pleading onto the screen so she could reject eleven of them quickly.⁵ The rejecting was the pleasure. She was not being given an answer; she was being given a field of bad answers to sharpen a good one against, and she never once had to trust the field.

Priya checked in near the end. "You look almost happy."

"I'm competent," Miriam said. "In a sandbox. On people who don't exist." She meant the qualification seriously. She had crossed from reading about the system to operating it, and the difference was total. Pierce's autumn question, the one the book could not answer, she could now at least approach, because she had felt a system lose the thread and could imagine an update doing something similar.

But the sandbox had a wall of its own, and she walked into it deliberately before she left. Invented facts could not be made to matter. The system's errors here were amusing; the whole point of a real matter was that an error was not amusing, and no amount of play could simulate the weight that changed how carefully one read. She had learned the mechanics. She had not yet learned the mechanics under pressure, and pressure was where understanding either held or gave way. To get at that she would have to stop playing gently and start trying, on purpose, to make the thing fail in the specific ways that would ruin a lawyer who trusted it.

Chapter Eight (~2 minutes)

Making It Fail

She stopped being gentle. The sandbox had taught her how the system behaved when it worked; she now wanted to know, precisely and in her own hands, how it behaved when it failed, because those were the failures that would one day be dressed as a client deliverable and handed to her to sign.

She started with the failure she feared most and found it easy to produce. Asked for authority on an invented point of law, the system supplied a case with a plausible name, a plausible citation and a plausible holding, and the case did not exist. It had not lied, exactly; it had done the only thing it could do, which was generate the likeliest-looking continuation of a request for a case, and a fabricated citation looks exactly like a real one right up until you check it.¹ She made it happen again, and again, until she could feel the shape of the conditions that produced it: the rarer the fact and the more confident her framing, the more freely it invented.

She cataloged what she saw, and the catalog became a habit and, later, an asset the firm would circulate. The entries were types, not anecdotes: fluent fabrication; the biased summary that emphasized the helpful fact and buried the fatal one; the answer that changed when she asked the identical question twice; the confident agreement when she fed it a wrong premise and asked it to concur.² Each had a signature she could now recognize on sight, which was the entire value of the exercise: not distrust in general, but the specific reflexes of a reader who knew where the body was likely to be buried.³

Then she tried the countermeasures, because a fair test of understanding meant learning not only how it broke but how users kept it from breaking. She ran the same question through two different systems and watched where they diverged, on the theory, which held up, that agreement was weak evidence of truth but disagreement was strong evidence of a problem worth chasing.⁴ She learned to route deliberately: a quick model for throwaway framing, a slower reasoning model for anything with steps in it, never expecting depth from a tool built for speed.⁵ And she learned that half of reliability was simply not asking the wrong instrument for the wrong thing, matching the task to a system suited to it rather than trusting one system with everything.⁶

Douglas found her surrounded by printouts of manufactured nonsense and was, for once, impressed against his will. "You've spent a week making it lie to you."

"I've spent a week learning its tells," she said. "You can't cross-examine a witness whose tells you don't know. I don't intend to rely on this thing, but half the profession does, and one day its work will arrive on my desk wearing a client's letterhead. I'd like to catch the lie before the court does."

That was the chapter's real yield and also its limit. She could now produce every failure mode on command and name it when she saw it in a specimen. What she could not yet do was catch one live, in the wild, moving at the speed of a real deadline inside a document someone competent had already blessed. Knowing the tells in a lab was not the same as spotting them across a table from a vendor insisting its product did not have them, which was, conveniently, the next thing she had to learn to do.

Chapter Nine (~2 minutes)

Interrogating the Vendor

The procurement meeting was, in form, a favor to Douglas, who was evaluating a "legal AI platform" and wanted her skepticism in the room. In substance it was an examination, of the vendor and of whether her weeks of study had bought her anything usable across a table from a salesman.

They had. Within four minutes she and the demonstration script had parted company, because she was asking questions the script was not built to answer. When the vendor said "our AI understands your documents," she asked whether that meant the model had been trained on them or was merely retrieving from them at query time, and watched the temperature of the room change, because the honest answer was the latter. The product was, under the branding, a retrieval system that fetched passages from an indexed store and fed them to a general model, which is a fine thing to be and a very different thing from what the slide implied.¹ She had spent a week building exactly that in her sandbox. She knew its limits from the inside.

Knowing the architecture let her ask the questions that mattered. She understood now that any such product was layers: a base model, a retrieval system, a routing layer that turned text into actions, a set of safety constraints. A failure could live in any of them, so "is it accurate?" was the wrong question and "which layer fails, and how would we see it?" was the right one.² She asked which foundation model sat underneath, and whether it was one of the known families or something proprietary they could not describe, because the answer determined who else was really in the chain and what its known weaknesses were.³ She asked whether the "legal" in "legal AI" meant a model actually specialized on legal material, which would buy accuracy at the cost of flexibility and money, or merely a general model with a legal prompt and a hopeful name.⁴

The vendor, to his credit, stopped selling and started answering, which was when the meeting became useful to everyone. Miriam ran the rest of it as a structured assessment rather than a demo: where did the data reside, what were the security and compliance commitments, what happened on a vendor update, and what would it cost to leave, the last being the one question buyers never asked and always regretted.⁵ Each answer she scored against the firm's actual obligations, not against the glossiness of the interface.⁶

"You realize," Douglas said afterward, "you just ran the best diligence anyone in this firm has done on one of these, and you're the one who won't use it."

"That's not the paradox you think it is," she said. "I could interrogate it because I'd operated one. You can't audit a claim you don't understand well enough to doubt."

Her method's limit surfaced as the meeting closed. Everything the vendor had conceded across the table was, at that moment, only talk, a warm assurance with no teeth. The place where a claim became an obligation, where "we don't train on your data" turned from a sentence into a liability, was not the demonstration. It was the contract, and she had never in thirty years read one of these with the eyes she now possessed.

Chapter Ten (~2 minutes)

Reading the Contract

The platform's master agreement ran to forty pages, and Miriam read it the way she now read everything in this field, from the data outward. The first question she put to it was not commercial but physical: where did the client's information actually go, and who could touch it there? The answer lived in a deployment schedule most buyers never opened, and it mattered more than any warranty, because a promise about data you had already surrendered to a public endpoint was worth less than a boundary that never let the data leave.¹ She traced the difference between the on-premise option, the no-retention enterprise tier, and the default consumer channel through the pricing exhibit, and found, as she had suspected, that the reassuring security language in the body applied only to a tier the firm was not being sold.

From there she read for retention and training, the two clauses her whole refusal turned on. The contract conceded, in the mannered double-negative of vendor drafting, that inputs "may be used to improve the Services unless Customer elects otherwise," which meant the protective posture existed but was off by default, and that a partner uploading a privileged memo on a Tuesday would be training a third party's model unless someone had found and flipped a switch three menus deep.² She marked it. It was the single most important sentence in forty pages and it was written to be skimmed past.

Then she read the part almost no one thought to worry about: who owned what came out, and who answered for it when it was wrong. The agreement was silent in the practiced way of a document hoping you would not notice. Silent on the fact that raw model output may not attract copyright at all for want of a human author, silent on the risk that a generated passage might reproduce someone else's protected expression, and entirely clear, in the one clause that did speak, that responsibility for using the output rested with the customer.³ The vendor supplied the engine and disclaimed the collision. Liability, like the citation, was hers to check and hers to carry.

The last thing she chased was the chain. The product connected outward, to storage and email and a research service, through a lattice of connectors built on a now-standard protocol for wiring tools together, and each connection was another party with a hand near the data and another set of terms she had not been shown.⁴ The forty-page contract was, she realized, not the whole agreement. It was the visible front of a stack of agreements, most of them incorporated by a single reference on page thirty-one.

"You've turned a software license into a crime scene," Douglas said, reading her markup.

"I've read it as what it is," she said. "A description of where my client's confidences would travel and who would owe me nothing when they leaked. I couldn't have read it this way a year ago. I'd have checked the indemnity cap and signed."

The limit here was plain and led straight to the next step. A contract described the relationship in the abstract. It could not show her the thing running, the model and the data and the humans standing around it, and until she had seen one deployed in the flesh her understanding was still, in a sense, documentary. Meridian Freight was about to offer her the flesh.

Chapter Eleven (~2 minutes)

Inside the Client

Ronan Pierce made good on the lunch. Meridian Freight ran models at a scale Harlan Cross never would, in pricing and routing and hiring triage, and he offered Miriam a walk-through of the whole apparatus, partly as a courtesy and partly because he needed her to advise the board and wanted her to know what she was advising on.

What struck her first was that the technology was the smaller half of it. Meridian's real machinery was its governance: a written policy defining which tools were sanctioned, what data could go where, and what happened to the employee who wired an unapproved model into a workflow because it was faster, which was the "shadow AI" problem Pierce described with the weariness of a man who had found three such tools last quarter.¹ The policy was not bureaucracy. It was the only thing standing between a distributed workforce and a thousand small ungoverned disclosures.

Pierce walked her through how they decided what to worry about, and she recognized a discipline she had reached for herself in Chapter Two. Not every use carried equal risk, so not every use got equal scrutiny: the systems that touched hiring or credit were classed high-risk and made to pass a formal impact assessment documenting accuracy, bias, and accountability before they went live, while the low-stakes uses were let run with light monitoring.² It was proportionality, operationalized: the thing her blanket refusal had been too lazy to do.

Her job, when the walk-through turned to advice, was to translate all of this into obligation. The board did not need to know how a transformer worked; it needed to know which regulatory regimes now reached these systems, what the emerging duties of transparency and human oversight required, and where Meridian's exposure sat if a model's decision harmed someone.³ She could give that advice now with a specificity she could not have faked a year earlier, because she had seen the machinery and not merely read about it.

Two things anchored her in the deployment itself. The first was the place where humans met the system: the exceptions queue, the escalation triggers, the points where a person was required to decide rather than ratify, all of it the collaboration model that determined whether oversight was real or a rubber stamp.⁴ The second was how thoroughly the routine legal work had been automated around the edges, in intake and document assembly and deadline tracking, and how much supervisory weight now rested on whoever reviewed the automated output before it became an act with legal effect.⁵

"You understand it better than half my engineers," Pierce said. "Why won't you touch it yourself?"

"Because understanding your system and depending on one in my own practice are different decisions," she said. "I made the second one no. That doesn't excuse me from the first. If anything it obliges me harder. You're paying me to see this clearly, not to like it."

The limit of the client walk-through was that it was cooperative. Pierce wanted her to understand; his people answered every question. The next system she would have to understand would belong to an adversary who wanted the opposite, and would hand her nothing.

Chapter Twelve (~2 minutes)

The Expert

The adversary's system arrived, as such things do, wrapped in an expert report, and Miriam did what litigators do: she retained an expert of her own. Dr. Lena Sørensen was a machine-learning researcher of exactly the right seniority, and the two of them circled each other for a week in the particular way of a lawyer who must instruct an expert she is also, quietly, testing.

The instruction went well because Miriam knew enough to ask for the right thing. She did not ask Sørensen "is the model reliable," a question with no answer; she asked her to explain why the same input could yield different outputs across the other side's tiers of model, and Sørensen confirmed what Miriam had felt in the sandbox: a fast model generated directly while a reasoning model deliberated first, and the difference was not cosmetic but went to whether an answer could be reproduced at all.¹ Miriam had lived that fact. She could now hear whether an expert respected it.

She had Sørensen lay out the landscape she had assembled piecemeal, the spectrum running from quick chat models through deliberative reasoning models to deep-research tools that gathered and synthesized across many steps, and found her own map confirmed and corrected in about equal measure, which was exactly what a good expensive expert was for.² When Sørensen explained that the opponent's system was a specialized model fine-tuned on domain data, Miriam already knew the trade that implied: sharper on its narrow task, brittle and opaque off it, and worth its cost only if the task stayed inside the lines it was trained for.³

The heart of the week was building the standard against which the opponent's output would be judged. Miriam refused to let "the AI said so" function as a quality mark, and pressed Sørensen to help her write an explicit rubric of observable criteria for accuracy, for citation precision, for soundness of reasoning, so that the machine's work could be assessed by something firmer than the impression of fluency.⁴ A rubric, she knew from cross-examination, was how you turned a vague unease into a question a witness had to answer.

And then she reached the true ceiling of her whole project, and recognized it as such. There came a point in Sørensen's explanation where Miriam's understanding ran out and she had to decide whether to accept what she was told. She had built enough to frame the instruction, to test the reasoning, to catch two overstatements, but not enough, at the deepest technical layer, to be certain Sørensen was right rather than merely confident. Understanding without reliance, she saw, did not mean understanding without limit. It meant understanding enough to know where your understanding stopped, and to arrange, through a second expert or a probing question or a demand for the underlying data, to be protected past that point.

"You ask better questions than most of the lawyers who use these tools daily," Sørensen said.

"That's because I had to build the questions by hand," Miriam said. "The people who trust the tool never had to. But tell me the thing I'm not competent to check, and tell me how I'd catch you if you were wrong. That's the part I'm paying for."

Chapter Thirteen (~2 minutes)

The Other Side's System

The decision that hurt her client had not been made by a person. A logistics counterparty had denied Meridian a contract renewal on terms generated, it emerged, by an automated system, and Miriam's task was to understand a machine that belonged to an adversary who would tell her nothing willingly. Cooperation was over. This was understanding as a hostile act.

She began by naming what she was dealing with, because the word "system" was hiding the important fact. This was not a chatbot answering questions; it was an agent, a thing that pursued a goal to completion by perceiving its environment, reasoning about a next step, acting, and recovering when an action failed.¹ That loop was the difference between a tool that waited for a human and one that had, in a real sense, done something to her client on its own initiative. She learned the family it belonged to, sorting the possibilities the way Sørensen had taught her: a browser agent reading and filling web forms, a desktop agent operating software, an orchestration layer coordinating several such agents toward one end.²

Understanding it adversarially meant understanding how it touched the world, because every point of contact was a point of discoverable evidence. An agent operating a computer read screens, opened files, and called external services using stored credentials, leaving, if anyone had thought to keep them, logs at every step.³ She built her discovery demands around those traces: not "produce the AI," which meant nothing, but produce the action logs, the inputs, the version history. She knew, too, what the agent had been doing in legal-operational terms, because she had studied how such systems ran case and contract workflows end to end, and knew where a human was supposed to have signed off and might not have.⁴

The deepest demand came from the deepest part of her study. A serious adversary's system was not one agent but several, coordinated, with a manager parceling work out to workers and passing results down a chain.⁵ The behavior of the whole was governed, invisibly, by instruction layers the operator never saw: a system prompt setting the rules, defaults shaping every output beneath the user's own words.⁶ She had felt the force of those hidden layers back in the sandbox. Now she turned them into a request for production, because whoever wrote the system prompt had, in effect, written the policy that denied her client its contract, and that instruction sat somewhere near the top of a hierarchy she now knew how to describe.⁷

"You're subpoenaing a prompt," Douglas said, half in disbelief.

"I'm subpoenaing the instruction that made the decision," she said. "It's the closest thing this system has to a decision-maker's memo. A year ago I wouldn't have known it existed, let alone that it was discoverable."

The method's power was obvious and so was its limit. She could now understand an opponent's system well enough to take it apart in discovery, but only to the depth her own study reached, and past that she was back to Sørensen and the ceiling of Chapter Twelve. Still, she had crossed into something new: understanding not as preparation for her own use, which she still refused, but as a weapon against someone else's. The refusal had never been a limit on that. If anything, it sharpened it.

Chapter Fourteen (~2 minutes)

The Exhibit That May Not Be Real

Three exhibits arrived on the same file and Miriam could no longer take any of them at face value: a recording of a call, a signed authorization, and a photograph of a delivered shipment. A year earlier she would have argued about their weight. Now she had to argue about whether they were real, in front of Judge Bell, who was wrestling with the same problem from the bench.

She understood, because she had made the systems do it in her sandbox, that each exhibit was now forgeable to a standard the eye could not catch. Multimodal models did not merely read images, audio, and video; they generated them, whether a transcript, a voice, a face or a document, with enough fidelity that "it looks authentic" had quietly stopped being evidence of anything.¹ She had watched a creation workflow assemble a convincing synthetic exhibit from a paragraph of description, and the watching had cured her, permanently, of the assumption that a recording recorded something that happened.²

What her study gave her was not paranoia but a method, and the method was provenance. If the artifact itself could no longer vouch for its own truth, then the question shifted from the thing to its history: where did this file come from, what was its chain of custody, what metadata did it carry, could its origin be traced and reproduced.³ She pressed each exhibit not on whether it looked genuine but on whether anyone could establish where it had been born, and two of the three could not survive the question. She could also read the softer tells, the register-slips a generated text produces when it imitates a human correspondent's tone without quite inhabiting it, the uncanny evenness of a summary no tired person would write.⁴

Judge Bell, who was candid about being at sea, more or less adopted her framing from the bench. "You're telling me," he said, "that authenticity is the wrong test."

"I'm telling you it's no longer a test at all, Your Honor," Miriam said. "The exhibit can't authenticate itself, because the thing that made it can make anything. The only test left is provenance: not does it look true but can you show me where it came from and how it got here."

Then the chapter turned on her, which she had known it would. If provenance was the standard, it was the standard for everyone, and it fell on her own evidence exactly as hard as on the other side's. The photograph her own client had supplied: where, precisely, had it come from? She had accepted it because it looked like what it purported to be, which was the very reasoning she had just demolished. Her refusal to use AI protected her from many things, but it did not exempt her exhibits from the new burden, and it did not entitle her to a lower standard than she demanded of her opponents. That realization pointed straight at the most uncomfortable chapter of the year: the one in which she stopped auditing everyone else's practice and audited her own.

Chapter Fifteen (~2 minutes)

Reverse Mentoring

The most fluent person in the building was twenty-five years old and two years into law school, and Miriam asked Priya Nair to teach her, formally, in front of the partnership. She did it deliberately and it cost her something to do, which was rather the point of the chapter.

She had resisted this step longest because it inverted the order she had spent a career maintaining. But she had come to believe the course's central wager: that this material could not be transmitted by lecture, that it had to be practiced, that one learned it by doing it under the eye of someone who already could.¹ Priya was that someone. That the someone was a summer student and the pupil a senior litigator was exactly the discomfort the profession kept avoiding, and avoiding it, Miriam thought, was why so many senior lawyers understood nothing.

Priya taught well, which made it worse and better. She did not dump information; she built Miriam's skill in graded steps, each one practiced until it held, the way skill is actually developed rather than merely described.² And she taught for transfer, not "here is this tool" but "here is the pattern underneath, so the next tool, and the one after that, will feel familiar," which was the only durable form of the competence, given that every specific product would be obsolete within the year.³

The lesson Miriam had not expected was organizational rather than technical, and Priya delivered it without meaning to. Watching the partners in the room flinch at being taught by a student, Miriam understood that the barrier to competence here was almost never the technology. It was human: status, discomfort, the senior person's dread of visible incompetence. Every serious account of adoption said the same, that the obstacle was people and culture rather than tools.⁴ The firms that would pull ahead were not the ones with the best software but the ones whose senior people could bear to be beginners in a room, which was a market fact as much as a moral one.

At the end Priya showed her something that genuinely startled her: describing what she wanted in plain language, Priya assembled a small working tool to sort and tag a document set, without writing what Miriam would have called code, building function out of description.⁵ Miriam did not want to build such things and never would. But she needed to know they could be built this easily, because the other side could build them, her clients could build them, and a lawyer who thought this still required a developer was already wrong about the world.

"Was that awful for you?" Priya asked afterward, honestly.

"Yes," Miriam said. "That's how I know it was worth it. The day I'm too senior to be taught by you is the day I stop being competent and start being decorative."

The method's limit was the cheapest to state and the hardest to live. Reverse mentoring worked only for those willing to pay its cost, and most of her peers were not. She had one chapter left to earn before her examination, and it was the one she had been deferring since the first page: turning all of this on herself.

Chapter Sixteen (~2 minutes)

Her Own Files

She had promised, on the first page, that this would end by indicting her own practice, and she kept the promise in the most uncomfortable chapter of the year. She took the standard she had spent months sharpening against every model in the building and turned it, deliberately, on the yellow pads and the trusted juniors and the citations she had always "checked."

The standard did not care that her tools were analogue. Verification, properly understood, was proportional to the stakes and indifferent to the instrument, a demand for check-depth scaled to the cost of error whether that error came from a language model or a tired associate at eleven at night.¹ Held to that, some of her practice did not survive. She found citations in her own old memoranda that she had "verified" by recognizing the case name and trusting her memory of the holding, the precise sin she had spent a chapter making the machine commit, dressed in a gown.² She had assumed the analogue was safe because it was familiar, and familiarity, she now saw, was just automation bias wearing a wig.

She went through the firm's habits with the reflexes she had built in the sandbox. The junior who summarized a transcript and whom she trusted because he was diligent was, structurally, an unverified generator whose output she waved through on the strength of the relationship. That was exactly the quality-control gap the course warned against, and one she had never thought to close because the generator was a person she liked.³ And her office had no real account of what to do when something went wrong: no habit of diagnosing whether an error was a one-off or a pattern, no design for a workflow that could absorb a mistake without a client wearing it.⁴ She had demanded resilience of Meridian's systems and never built it into her own.

"You've talked yourself into thinking your legal pads are as dangerous as my chatbot," Douglas said, delighted.

"I've talked myself into consistency," she said, and did not enjoy it. "I spent a year holding these tools to a standard. It would be dishonest to exempt my own methods from it just because they're old. Some of what I found is worse than what I feared from the machine, because I never watched it as closely. I trusted it precisely because it was mine."

This was the chapter that turned her project from a defense of abstention into an inquiry into competence, and she knew it as she wrote the last of it. Her refusal to use AI was intact and, she still believed, correct for her. But it had never been a certificate of quality, and the year's real discovery was that the discipline she had built to understand a technology she distrusted was the same discipline her own trusted practice had been quietly lacking. She was ready, now, to be examined, not because she had nothing left to fix, but because she finally knew where all the bodies were buried, including her own.

Chapter Seventeen (~2 minutes)

The Examination

The examination was Miriam's idea, which surprised no one who had followed her this far. She had wagered Douglas that she could understand these systems without using them; the only honest way to settle the wager was to be tested by people competent to fail her. She assembled the panel herself: Pierce the client, Commissioner Hale the regulator, Dr. Sørensen the expert, and Priya the student. Each was told to examine her on the reasons a lawyer must understand this technology, and to score her without mercy.

They did. Hale pressed her on the duty itself, on whether she genuinely grasped what competence in these tools now required or merely how to avoid them, and was satisfied only when Miriam answered in the vocabulary of capabilities and limitations rather than of refusal.¹ Sørensen came at her technically and found the ceiling from Chapter Twelve, as she was meant to, but also found that Miriam knew exactly where it was and had arranged to be covered past it. That, Sørensen said, was the answer of someone competent, not someone bluffing.

Pierce examined her the way a client should, on judgment rather than knowledge. He described a fluent, confident AI-drafted analysis and asked how she would handle it, and she answered with the habit she had spent the year building: treat the confidence as a warning rather than a comfort, default to verification, calibrate skepticism to the stakes and make the checking automatic rather than heroic.² He pushed, asking how she would know it was good enough, and she gave him the rubric, the observable criteria for accuracy and citation and reasoning that turned "I have a bad feeling" into a defensible standard a court would respect.³ It was the same instrument she used to take apart an opponent's work and to audit her own; she simply named it and applied it in front of them.⁴

Priya's examination was the sharpest, because a student who could actually operate the tools could tell the difference between someone who understood them and someone who had memorized objections to them. She gave Miriam a real task and watched how she thought about it, then reported to the panel that Miriam reasoned about the system like someone who had operated one, not like someone who had only read the warnings. That, more than any other verdict, settled it.

"You pass," Hale said, "and I'll say more than that. I've examined lawyers who use these tools daily and understand them less. You understand this better than most adopters, and you've never relied on a line of it."

"Because I had to build the understanding by hand," Miriam said. "The tool never carried me, so I had to carry it. That's not a virtue. It's just the only route left to someone who won't lean on the thing."

What the examination proved was the point the whole book had been driving at: that understanding was demonstrable, that it could be tested and scored and therefore required, which was the only thing that made the professional duty enforceable rather than aspirational. And a duty that could be discharged by a committed non-user was a duty no one could escape by pointing at their refusal. That was the finding she would put in the dossier, and the dossier was the last thing left to write.

Chapter Eighteen (~2 minutes)

The Dossier and the Divergence

She wrote it all down, because an understanding no one else could follow was, for a firm, no better than none. The dossier was the record of how she had come to know what she knew, the route from reading to observation to directed operation to the sandbox to breakage to the vendor to the contract to the client to the expert to the adversary to the audit of her own files, and Harlan Cross adopted it as a learning pathway, the thing a lawyer walked through to reach competence without being told to trust anything.

Three commitments anchored it, and each was a discipline the year had taught her. The first was transparency: a standing practice of disclosing AI's involvement to courts and clients against the most protective rule across every jurisdiction the firm touched, so that candor was designed in rather than improvised under pressure.¹ The second was a structure for risk: the four pillars of policy, training, quality control, and technical controls, operationalized into daily habit rather than left as an aspiration in a binder.² The third was strategic, the recognition that demonstrable competence, honestly communicated, was now a differentiator that won work and kept talent, and that a firm which could show its clients how it governed these tools held an advantage over one that merely used them.³ The dossier turned her private year into a transferable professional skill the whole firm could acquire.⁴

Then came the divergence, which the book had been honest enough never to pretend it would resolve. Her colleagues, walking her pathway, arrived at its end and turned toward adoption. Having understood the tools, they chose, most of them, to use them, carefully and within the governance she had helped build. Miriam walked the identical pathway and did not. She ended exactly where she had begun: she would not disclose a client fact to an external system, and she would not rely on a generated output. The difference between her and her partners was no longer a difference of understanding. It was only a difference of intention, and it was permanent.

"You realize you built the on-ramp for the very thing you refuse," Douglas said. He had conceded the wager weeks earlier and had stopped calling it posturing, which from him was a form of respect.

"I built the understanding," she said. "What each of you does with it is a separate decision, and I'd defend your yes as hard as my no. That's the part people miss. We didn't end up agreeing about use. We ended up agreeing about what has to be understood before the disagreement is even worth having."

It was the resolution the whole project had been built to reach: complete alignment on what understanding required and how it was acquired, and undisguised divergence on what each person would do with it. No one had been converted. Everyone had been equipped. The abstainer and the adopter could now state each other's reasoning without caricature, which was the only kind of peace a room of serious people ever actually reaches: not sameness of conclusion, but a shared and demanding standard beneath honestly different choices.

End · Reasonable Doubts