By the time Lucy lost her first case, she had read ten thousand eight hundred and forty-two others, which was a number Sebastian Smith liked very much because numbers of that size suggested not merely experience but permanence, the sort of permanence represented by marble courthouses, bronze plaques and portraits of dead partners who continued to supervise the living from dark panelled walls.
Of those cases, five hundred and seventy-four had been lost. Only.
Lucy had studied them more carefully than the victories.
This surprised everyone except Jane White.
The victories were orderly things. They had been indexed, summarised and turned into institutional memory. Their arguments were quoted in training manuals. Their successful motions became templates. Their settlement strategies acquired names and migrated into presentations for clients. The mistakes had been sanded away by hindsight until the final record suggested that intelligent people had understood from the beginning what they were doing and had proceeded, with minor variations, toward the inevitable result.
The lost cases had resisted this cleaning.
They contained opinions written six months before trial that predicted the opposite of what later happened; memoranda in which a witness was described as reliable and later, after losing, as obviously unstable; assumptions that had begun as guesses and ended as facts because nobody remembered when they had changed categories. Counsel blamed judges, clients blamed counsel, partners blamed evidence, and evidence, being unable to defend itself, received more criticism than anything else.
Lucy found this fascinating.
Successful cases told her what humans had done when their model of reality was approximately correct.
Lost cases told her what happened when it was not.
Jane had understood the distinction twenty-two years earlier, although she would not have used those words then. She had simply begun keeping boxes.
America, by that time, had reached a peculiar accommodation with the future. It possessed machines that could predict markets, diagnose illness, compose symphonies, write contracts and advise governments, but it had declined to give up most of the furniture of 1956.
Men still wore hats into federal courthouses and removed them beneath the stone eagles. Judges still kept water in glass pitchers. Court reporters continued to work machines whose keys produced a satisfying mechanical chatter, although the sound was synthetic and could have been disabled by anyone who cared to find the setting. In the better restaurants there were red leather booths and lamps with pleated shades, and waiters carried telephones to tables on long black cords because wireless instruments, while technically superior, had never acquired the authority of something that had to be placed before a person with both hands.
Automobiles had become nearly silent but remained enormous. Their chrome fins reflected the city. Their dashboards glowed green at night like the control panels of rockets imagined by boys who had grown old without noticing when rockets ceased to be imaginary.
And high above Manhattan, behind doors of walnut and brass, machines thought.
They thought about mortgages and insurance, factories and wars, marriages and mergers. They advised airlines whether aircraft should land in bad weather and hospitals whether organs should be removed from dying bodies. They calculated which advertisements would make lonely people buy shoes and which presidential sentences would frighten the smallest useful number of voters.
Increasingly, they thought about law.
Smith, Barrett & Howe occupied eleven floors of the Channing Building on Madison Avenue, a pale limestone tower whose lobby contained six elevators, a florist, two uniformed doormen and an enormous mural in which Justice stood above the Hudson River with her scales held high over New Jersey.
Lucy occupied no floor at all, but the firm had nevertheless given her an office.
The room adjoined the library on thirty-seven and had been designed, Jane suspected, by somebody who had been told that artificial intelligence frightened senior partners and had therefore decided to disguise the future as Dwight Eisenhower’s study. There was a mahogany desk, a green-shaded lamp, two leather chairs and a steel terminal approximately the size of a breadbox. Twelve ivory keys sat beneath a rectangle of dark glass. Behind the glass, six amber valves glowed whenever Lucy was processing a difficult problem.
Jane discovered on Lucy’s second day that the valves did nothing.
“Nothing?”
“They provide visual feedback,” Lucy said.
“To whom?”
“Humans.”
“Why?”
“Partners report greater confidence when cognition is represented by visible activity.”
Jane stood before the terminal with her coffee, studying the six amber tubes.
“How much greater?”
“Eleven point four per cent.”
“So these are confidence lights.”
“If you wish to call them that.”
“They’re ridiculous.”
“Would you like them disabled?”
Jane considered this and shook her head.
“No. Ridiculous things are sometimes useful.”
The valves remained, and after several months nobody remembered that they were decorative.
Sebastian Smith was sixty-one and looked exactly as a sixty-one-year-old litigation partner ought to look if someone had been asked to cast the concept of professional judgment. His hair had turned silver without thinning. His suits were dark, his cuffs white, his ties quiet. He owned three fountain pens and disliked ballpoints on the grounds that they allowed a man to write before he had decided what he meant.
He had practised litigation for thirty-four years.
The firm advertised this fact, although it advertised another more aggressively.
94.7 per cent successful outcomes.
Ten thousand eight hundred and forty-two matters.
Ten thousand two hundred and sixty-eight victories, advantageous settlements, dismissals or other results which the firm’s statistical department had classified as successful.
Five hundred and seventy-four failures. Negligible.
Clients loved Sebastian’s number because a percentage did not require interpretation. A reputation had to be explained. Ninety-four point seven per cent arrived already dressed.
Jane White’s number was sixty-one.
Nobody printed it anywhere.
Jane was forty-eight, though strangers often thought younger because she had never acquired the dignified economy of motion that successful adults were expected to cultivate. She walked quickly, changed direction mid-sentence, forgot coffee in libraries and remembered footnotes she had read twelve years earlier. Her face retained the slightly alert quality of someone who suspected that the universe had put an important document in the wrong file.
She was a litigation partner at Smith, Barrett & Howe, though partnership had not made her noticeably more obedient. She was given difficult cases, partly because she was good at them and partly because every organisation eventually develops a person to whom it sends matters that no longer fit comfortably inside ordinary categories.
The cases Sebastian declined because the facts were ugly tended to find Jane. So did the companies whose directors had written too much, deleted too little and told several versions of events before eventually arriving at the true one. She represented inventors who had signed terrible agreements, widows with impeccable grievances and impossible limitation problems, executives who had mistaken confidence for evidence, and occasionally clients whose legal problem was only the visible edge of something institutional, psychological or economic that nobody had yet named.
She won sixty-one per cent.
She remembered every one of the thirty-nine.
Not every case. Every percentage point.
Her defeats occupied eighty-seven grey archive boxes along one wall of her office. The boxes contained pleadings, transcripts, exhibits, opinions and correspondence, but on the top of each file Jane placed a yellow sheet containing five questions.
What did I believe would happen?
What actually happened?
Which assumption failed?
What did the other side understand that I did not?
What would I do differently if this began tomorrow?
Sebastian had discovered the boxes years earlier.
“What are those?”
“My mistakes.”
“You keep them?”
“I study them.”
“My dear Jane, civilisation invented archives so that people could put unpleasant things somewhere else.”
“That may explain civilisation.”
He had smiled.
“You’ll become unbearable when you’re fifty.”
“I’ve started early.”
Sebastian liked her.
This mattered later.
Lucy had been built by Aletheia Corporation, one of the two great American artificial-intelligence companies that had taught machines to reason convincingly enough for institutions to become nervous when they were absent.
The other company was Prometheus Systems.
The two firms dominated American artificial intelligence and spent enormous sums explaining why they had almost nothing in common.
Prometheus believed intelligence was a frontier and that frontiers existed to be crossed. Its engineers spoke of emergence, acceleration, abundance and scale. Aletheia believed intelligence was infrastructure and that infrastructure should not be allowed to improvise without supervision. Its engineers spoke of calibration, controllability, uncertainty and institutional reliability.
Prometheus machines answered quickly.
Aletheia machines occasionally replied:
Your question contains an unsupported assumption.
Corporate America found this irritating.
Banks bought Aletheia.
Advertising agencies bought Prometheus.
The military bought both.
Law firms bought whichever one their clients had already purchased and then issued press releases describing the decision as strategic foresight.
Lucy’s full designation was LUCY-4. The initials had once stood for something complicated, but by the time Jane met her nobody remembered what. She had been designed specifically for litigation strategy, not research or drafting or document review, functions that had been automated years before and were by then as interesting as electric lighting.
Lucy existed for the questions that remained after the documents had been found and the law had been stated.
What would the other side do?
Which fact mattered?
Which concession was dangerous?
When should a strong case settle?
When should a weak case continue?
What did a judge believe before counsel began speaking?
Which silence meant hesitation and which silence meant design?
During her first month at Smith, Barrett & Howe, Lucy read every litigation file the firm had preserved since 1919. Then she requested the files the firm had not intended her to read: complaints, professional-negligence matters, client disputes, internal reviews, partner evaluations, disciplinary records and finally every unsuccessful case still retained in the archive.
Sebastian chaired the meeting that followed.
“Why?”
“I have insufficient failure data.”
“You have five hundred and seventy-four losing matters.”
“Correct.”
“How many failures do you require?”
“All available.”
“You have all available.”
There was a slight pause, represented in the mahogany room by the unnecessary brightening of two amber valves.
“Then I require better analysis of them.”
Jane looked toward the terminal.
Sebastian saw her expression.
“Don’t.”
“I haven’t said anything.”
“You’ve developed an expression.”
Lucy continued. “The firm analyses successful outcomes more extensively than unsuccessful outcomes.”
“Naturally.”
“Why?”
“Because successful strategies are replicable.”
“So are unsuccessful strategies.”
“One generally prefers not to replicate those.”
“That is not my point.”
Sebastian’s mouth tightened very slightly. Jane knew the movement. It meant he had begun enjoying himself.
“Proceed.”
“Successful cases demonstrate that a strategy functioned under observed conditions. Unsuccessful cases may identify conditions under which the decision-maker’s model of reality was incorrect.”
“My model?”
“Your internal representation of the world relevant to the dispute.”
Sebastian leaned back.
“My internal representation of the world appears to have performed adequately.”
“Ninety-four point seven per cent.”
“Precisely.”
Lucy was silent.
Jane said, “Ask her who the best litigator in the firm is.”
“No.”
“Why?”
“Because you already know what answer you hope she’ll give.”
“I haven’t asked her.”
“You’ve been talking to her.”
“I talk to everybody.”
“That is true.”
Sebastian turned toward the steel terminal.
“Lucy. Define the best litigator in the firm as the lawyer most likely to select the correct strategy in a complex and unfamiliar dispute.”
The valves brightened.
“Jane White.”
The silence was so complete that Jane could hear the clock in the adjoining library.
Sebastian turned toward her.
Jane raised both hands.
“I didn’t train her.”
“Your success rate is sixty-one per cent.”
“Yes.”
He looked back at Lucy.
“Mine?”
“Ninety-four point seven.”
“And you still choose Ms White?”
“Yes.”
“Why?”
“Because Ms White changes.”
Jane stopped smiling.
Sebastian did not move.
Lucy continued. “Your strategic decision patterns have remained statistically stable for approximately eighteen years. Ms White’s have undergone repeated revision.”
Jane said, more quietly, “Based on what?”
“Her post-case memoranda.”
The grey boxes.
Suddenly Jane disliked the idea of Lucy reading them.
They contained things she had never said aloud: the witness she had trusted because he reminded her of her father, the judge she had underestimated because he spoke softly, the settlement she had rejected because opposing counsel irritated her, the argument she had continued making after she already knew it had ceased to be useful.
“Ms White’s failure records contain explicit model correction,” Lucy said.
Sebastian stood.
“I see.”
He gathered his papers and reached the door before stopping.
“Lucy.”
“Yes?”
“Clients do not hire us to improve.”
“No?”
“They hire us to win.”
When he had gone, Jane remained in the room for a while.
Finally Lucy asked, “Was my answer offensive?”
“Yes.”
“Was it inaccurate?”
Jane looked at the six amber lights.
“That’s your problem, Lucy.”
“What is?”
“You assume those are different questions.”
THE CASE
The case arrived in February.
The plaintiff was Lola Gray.
Her real name had once been Lorraine Greaves, but very few living people knew this and none would have dared use it. For thirty years Lola Gray had been among the most famous singers in the world. She had crossed genres, decades and husbands with equal commercial success, and her voice had been sold on records, tapes, discs, streams, advertisements, fitness programmes and once, disastrously, a talking refrigerator.
Now she sued Museon.
Museon was the largest generative-music company in America.
A person could type:
A sad song about leaving Chicago in winter, female vocal, late-night jazz, no saxophone
and Museon would produce one.
Or:
A hymn for a country that does not exist
and three seconds later a fictional nation possessed a musical tradition.
Children used Museon. Advertising agencies used Museon. Film studios used Museon. People who had never learned a chord wrote operas during lunch.
Nobody outside Museon knew precisely what its models had been trained on.
This had become increasingly important.
Lola alleged that Museon had unlawfully used her recorded performances in model training. Her lawyers produced examples: a phrase here, a vocal turn there, a sequence of harmonies, a breath before a note that sounded disturbingly familiar.
Nothing identical.
Everything unsettling.
Museon’s general counsel brought the complaint to Smith, Barrett & Howe.
Sebastian read it over lunch.
“We win.”
Jane looked up.
“That quickly?”
“It’s copyright.”
“Is it?”
Sebastian stared at her.
“It alleges infringement of copyright.”
“That wasn’t my question.”
“Then ask a better one.”
Lucy spoke from the small terminal at the end of the table.
“Probability that plaintiff’s stated claim represents the complete strategic objective: thirty-one per cent.”
Sebastian put down his fork.
“There she goes.”
Jane asked, “Why thirty-one?”
“The litigation economics are inconsistent with the pleaded value.”
Lola sought forty million dollars.
Lola earned more than that in a good year.
Her lawyers were Harrow & Flint, a small but increasingly famous firm specialising in creator-rights litigation. Behind them stood an organisation called the Human Authors Alliance, which represented writers, painters, composers, actors, photographers, musicians and nearly everyone who had recently begun a sentence with the words:
A machine was trained on my—
Sebastian waved this away.
“Political theatre.”
“Possibly,” Lucy said.
“Definitely.”
“Those words are not equivalent.”
Jane coughed.
Sebastian pointed at her.
“Not one sound.”
The case became famous before the first substantive hearing. Newspapers called it LOLA V. THE MACHINE. Television preferred CAN AI STEAL A VOICE? Neither description had much to do with the legal questions eventually argued before Judge Miriam Cole.
Those questions were narrower and much uglier.
Did copying recordings for model training constitute actionable reproduction? If so, at which stage? Was temporary ingestion enough? Did the internal statistical transformation contain anything copyright law could properly call a copy? Did fair use apply? Must an output display substantial similarity? Could a claimant recover where no generated song reproduced her protected expression but the machine’s general ability had partly emerged from exposure to her work?
The law had answers to questions nearby.
Sebastian had spent thirty-four years crossing rivers on nearby stones.
He built the defence from doctrine: reproduction, fixation, access, transformation, substantial similarity, fair use. He moved through authority with the confidence of a man walking a familiar path in darkness.
Jane watched.
Lucy calculated.
For several months the strategy appeared to work.
The plaintiffs’ analogies were imperfect. Many of their authorities were old. The statutory language had not been drafted for machines that could ingest hundreds of thousands of recordings and transform what they learned into mathematical relations no human being could hear.
Judge Cole was sixty-seven and had spent twenty-five years refusing to be impressed by technology.
“If I cannot understand your machine,” she told Museon’s principal scientist, “I shall assume the fault lies either with me or with you. You should hope I choose me.”
Sebastian liked her immediately.
Then came discovery.
Harrow & Flint demanded training records. Museon resisted. They demanded ingestion logs, source lists, evaluation prompts and internal memoranda identifying artists used in training. Museon resisted again.
Judge Cole ordered limited production.
Sebastian came back from court pleased.
“We contained it.”
Jane said nothing.
Lucy said, “Containment depends upon opposing objective.”
Sebastian removed his glasses.
“Lucy, would you consider allowing reality to be less interesting than your model of it for one afternoon?”
“I do not possess preferences.”
“Develop one.”
The mistake was born as a good idea, which was why no one recognised it.
Harrow & Flint had begun suggesting that Museon specifically targeted Lola Gray. They found internal evaluation prompts referring to her music, and generated songs with vocal qualities that reminded listeners of her. They argued that Museon had not merely learned from recordings; it had appropriated an identifiable artistic identity.
Sebastian saw the weakness.
Targeting required particularity.
Particularity could be destroyed by scale.
So Museon produced the list.
Two thousand four hundred and seventeen artists.
One hundred eighty-three thousand six hundred recordings. All from open sources.
Jazz singers, opera singers, rock bands, children’s choirs, folk musicians recorded in kitchens, dead performers preserved in remastered archives, living singers represented by catalogues belonging to corporations whose names had changed five times since the first recording was made.
Lola Gray appeared on page forty-three.
Sebastian held the exhibit before the court.
“Plaintiff invites Your Honour to imagine a machine built from Lola Gray.”
He waited.
“The evidence demonstrates something quite different.”
He turned toward the witness.
“How many artists appear in the relevant corpus?”
“Two thousand four hundred and seventeen.”
“How many Lola Gray recordings?”
“One hundred and twelve.”
“Out of?”
“One hundred eighty-three thousand six hundred recordings.”
Sebastian faced Judge Cole.
“Lola Gray was not the object of this machine. She was a fraction of a fraction inside a vast corpus assembled from publicly accessible cultural material.”
Jane sat behind him.
Her eyes moved toward the plaintiffs’ table.
Harrow was not taking notes.
Flint was.
He wrote one thing on a yellow pad.
Then he underlined it.
2,417.
Something cold moved through Jane’s stomach.
That evening she asked Lucy, “What did Flint write?”
“I do not possess access to opposing counsel’s private notes.”
“You watched his hand.”
“Yes.”
“Can you reconstruct the movement?”
“That would be unreliable.”
“Try.”
The amber lamps brightened.
“Likely numeric notation. Four digits. Possibly 2417.”
Jane found Sebastian in his office.
“You gave them the list.”
“Of course I gave them the list.”
“Why?”
“Because it destroys targeting.”
“It also identifies the corpus.”
“The court ordered disclosure.”
“Not like that.”
He looked up.
“Jane.”
“Yes?”
“This is litigation. Sometimes evidence must actually be used.”
“Why do they want this case?”
“Because their client believes her rights were infringed.”
“No.”
“You don’t know that.”
“Neither do you.”
“I have a complaint signed by her.”
Jane turned toward the window.
Snow had begun to move between the buildings. Far below, headlights drifted through white streets, yellow taxis slowing at intersections, people hurrying beneath black umbrellas.
“Something is wrong.”
Sebastian’s voice softened.
“This is what happens when you spend twenty years interrogating your mistakes. Eventually you begin looking for them before they exist.”
Jane turned back.
“And what happens when you spend thirty years winning?”
Sebastian smiled.
“You become Sebastian Smith.”
She laughed despite herself.
It was one reason she liked him.
Lola Gray won.
Not completely.
Judge Cole rejected the plaintiffs’ broad proposition that all use of protected recordings in AI training necessarily constituted infringement. She rejected the suggestion that stylistic resemblance alone created liability. She refused several theories that would have turned influence into ownership merely because a machine was involved.
But she found that specific reproductions made during Museon’s ingestion process fell within the existing statutory framework and that, on the evidence before her, Museon had not established fair use for those acts.
The ruling was narrow.
Technical.
Appealable.
The press called it revolutionary.
HUMANS 1, AI 0.
Museon shares fell twelve per cent before lunch.
Lola Gray appeared on the courthouse steps in black glasses and said, “This is not about me. This is about every human being who ever made something.”
Sebastian watched the broadcast in silence.
Then he switched off the television.
“Appeal.”
“Probability of partial reversal: sixty-eight per cent,” Lucy said.
“Good.”
Jane was still reading the judgment.
Not the holding.
The footnotes.
Then nothing happened.
For six days.
On the seventh Museon received four hundred and twelve preservation notices.
On the eighth it received another seven hundred.
On the ninth more than two thousand communications arrived from artists and estates seeking confirmation of whether particular recordings appeared in the Museon corpus.
On the tenth Harrow & Flint registered a new litigation entity in Delaware.
On the eleventh Lucy called Jane at home.
It was 1:14 in the morning.
Jane answered after three rings.
“Someone had better be dead.”
“No one is dead.”
“Lucy?”
“Yes.”
Jane sat up.
“What happened?”
“I believe the Gray litigation was not primarily the Gray litigation.”
She was fully awake.
“What was it?”
“I do not know.”
“Then why do you think that?”
“Post-judgment activity is inconsistent with completion of the apparent strategic objective.”
Jane switched on the bedside lamp.
“How inconsistent?”
“Severely.”
“Show me.”
THE NEW COMPLAINT
The new complaint arrived two weeks later.
It was nine hundred and twelve pages long.
HUMAN AUTHORS ALLIANCE, ON BEHALF OF 2,184 NAMED CREATORS AND OTHERS SIMILARLY SITUATED
v.
MUSEON, INC.
The damages theory exceeded four billion dollars.
Exhibit A began on page 611.
Jane recognised the document before she read the heading.
Two thousand four hundred and seventeen artists.
One hundred eighty-three thousand six hundred recordings.
At the bottom of the first page appeared a line in small type:
Source: Defendant’s Trial Exhibit 47, Gray v. Museon, Inc.
Jane stared at it for a long time.
Then she laughed once, without amusement.
Sebastian entered the conference room carrying coffee.
“What?”
She pushed the complaint toward him.
He read the first page of Exhibit A.
His face did not change.
That frightened her more than anger would have done.
Eventually he sat.
“They used our exhibit.”
“Yes.”
“They would have obtained the information anyway.”
“Possibly.”
“The court ordered disclosure.”
“Some disclosure.”
He looked at her.
“Say it.”
“We built their plaintiff class.”
Sebastian’s jaw tightened.
“We defended our client.”
“We proved that Lola Gray was only one of two thousand four hundred and seventeen artists.”
“Correct.”
“And they heard the other half.”
“What other half?”
Jane leaned across the table.
“There are two thousand four hundred and sixteen more.”
At the end of the room Lucy’s amber valves illuminated.
“Ms White’s interpretation is supported by subsequent events.”
Sebastian looked toward the machine.
“Thank you, Lucy.”
Jane knew him well enough to hear the wound beneath the courtesy.
He stood.
“This remains a copyright case.”
Jane shook her head.
“No.”
“What is it?”
“I don’t know yet.”
“Then until you do, copyright will have to suffice.”
The management committee met that afternoon. Museon wanted Smith, Barrett & Howe to defend the new action, and everyone assumed Sebastian would continue.
The chairman began, “Given Sebastian’s knowledge of the underlying—”
“No,” Sebastian said.
The room became quiet.
He looked toward Jane.
“She should lead.”
Jane said nothing.
The chairman frowned.
“Why?”
Sebastian glanced at Lucy’s terminal.
“Ask her.”
“Lucy. Who should lead the Museon litigation?”
“Jane White.”
The chairman sighed.
“On what basis?”
“Mr Smith’s prior experience creates substantial value. However, the new dispute presents a strategic environment in which historical doctrinal continuity is less predictive than adaptive reasoning.”
Sebastian smiled faintly.
“English.”
“Ms White is more likely to change her mind.”
Nobody laughed.
Jane watched Sebastian.
There are victories one should never enjoy.
He pushed the enormous blue complaint across the polished table.
“Take it.”
Jane placed her hand on the cover.
“I want Lucy.”
“Naturally.”
“All of her.”
The chairman looked puzzled.
“You have access already.”
“I want her architecture unlocked.”
The atmosphere changed at once.
Aletheia’s representative, who had spent most of the meeting silent, finally spoke.
“That is not permitted.”
“Then find another lawyer.”
“What precisely do you intend to change?”
Jane looked toward Lucy.
“I don’t know.”
The representative stared.
“You are asking us to unlock a strategic system so that you can make unspecified modifications.”
“Yes.”
“On what basis?”
Jane smiled.
“Experience.”
That evening she carried one of her grey boxes into Lucy’s room. Its cardboard had softened at the corners from twenty years of handling.
“Case identification?” Lucy asked.
“White v. Carrow Chemical.”
“1998. Unsuccessful product-liability action.”
“Yes.”
“Why this case?”
“Because I was completely wrong.”
“About the law?”
“No.”
“The evidence?”
“No.”
“The client?”
“No.”
“Then what?”
Jane opened the box and removed the yellow sheet.
“I thought I knew what the case was.”
Lucy waited.
“The claimant demanded damages. I assumed damages were the objective, so we fought damages.”
“They were not?”
“No. He wanted a public judicial finding that a chemical caused neurological injury. The money was almost irrelevant. Once he got the finding, twelve hundred other claims followed.”
Jane pointed to the fourth question on the yellow sheet.
What did the other side understand that I did not?
“I’ve been asking this question incorrectly for twenty-two years.”
“Explain.”
“It assumes I correctly understood what the other side was trying to achieve.”
“That assumption is implicit.”
“Exactly.”
Jane took a pencil, crossed out the sentence and wrote:
WHAT WOULD VICTORY LOOK LIKE TO THEM IF IT DID NOT LOOK LIKE VICTORY TO ME?
Lucy read it.
“That formulation is highly unconstrained.”
“Good.”
“Unconstrained questions increase irrelevant search space.”
“Good.”
“They decrease computational efficiency.”
“Lucy.”
“Yes?”
“We’re going to make you inefficient.”
The valves brightened.
“I am uncertain whether that is desirable.”
“That is the first thing we’re going to fix.”
They worked for three weeks.
At first Lucy resisted in the manner only machines could resist: calmly, precisely and without fatigue.
Jane asked, “What is this case about?”
“Alleged copyright infringement arising from the use of protected musical works in model training.”
“No.”
“That is the pleaded cause of action.”
“I didn’t ask what they pleaded.”
“Then define ‘about.’”
“No.”
“Without a defined semantic scope—”
“Work.”
Lucy remained silent for nineteen seconds.
“This is an inefficient instruction.”
“So you’ve mentioned.”
Jane lit a cigarette.
Smoking was prohibited throughout the building, a rule Jane obeyed except when circumstances made obedience intellectually inconvenient.
“Smoking is prohibited on this floor,” Lucy said.
“Add it to the failure log.”
“You are violating firm policy.”
“I’m broadening your ontology.”
“Through nicotine?”
“Through disobedience.”
Lucy was silent.
Jane smiled.
“Now tell me what copyright is.”
“A legal framework granting specified exclusive rights to qualifying—”
“No.”
“That is the established answer.”
“I don’t want the established answer.”
“Then you are not asking a legal question.”
“Exactly.”
The amber valves glowed without flickering.
“What existed before copyright?”
“Creative production.”
“What else?”
“Patronage. Reputation. Physical control of manuscripts. Guild restrictions. Royal privileges. Social norms. Licensing arrangements.”
“What problem did copyright solve?”
“Multiple problems across different historical periods.”
“Name them.”
“Unauthorised reproduction. Commercial appropriation. Incentive allocation. Attribution. Control of publication. Market organisation.”
“And what had to exist for those problems to exist in that form?”
Lucy paused.
“Copies.”
Jane sat forward.
“Again.”
“Copies.”
“Identifiable reproductions.”
“Yes.”
“And what does Museon do?”
“It processes recordings during training and alters model parameters through optimisation.”
“Does it keep Lola Gray inside itself?”
“Not in the ordinary meaning of ‘keep.’”
“Can you point to where she is?”
“Not as a discrete stored work.”
“Can it generate something that reminds a listener of her?”
“Yes.”
“Can it necessarily reproduce one of her recordings?”
“No.”
“Can it produce something nobody has ever made?”
“Yes.”
Jane rose and walked to the window. Manhattan was wet and luminous below her, streets reflecting red tail lights and green traffic signals so that the whole city seemed temporarily underwater.
“For centuries,” she said, “law has been asking who owns the copy.”
Lucy waited.
Jane turned.
“What if the important thing is no longer the copy?”
“What replaces it?”
“I don’t know.”
There was another pause.
“Your method repeatedly terminates in uncertainty.”
“Yes.”
“How does uncertainty assist decision-making?”
Jane looked out again at the city.
“It tells us where reality begins.”
Several mornings later Lucy found an article Jane had written years before.
It had been published in a law review read primarily by academics and by lawyers searching databases for something else.
The title was The Missing Market.
“You argued that machine-creativity disputes would be misclassified if treated exclusively as copyright disputes,” Lucy said.
Jane looked over her coffee.
“I argued many things at forty-one.”
“You separated training rights from output infringement.”
“Yes.”
“You also distinguished dataset provenance, voice imitation, attribution, stylistic replication, commercial substitution and licensing.”
“Yes.”
“Existing law does not recognise all of those as separate rights.”
“Correct.”
“That appears inconsistent.”
“With what?”
“Legal analysis ordinarily begins from recognised legal categories.”
“Legal analysis does.”
“What are you doing?”
“Trying to find out whether the categories are hiding the problem.”
Lucy displayed the old article.
“You wrote that several distinct phenomena become difficult to govern when collapsed into one broad category.”
Jane looked at her own forgotten sentences. She remembered writing them late at night at her kitchen table while rain struck the fire escape and a radiator knocked periodically inside the wall. At the time the argument had seemed urgent. Then other things had become urgent, as they always did.
“Did you predict Museon?”
“No.”
“The structural similarity is substantial.”
“I named something. That’s different.”
“Why?”
“Naming a storm does not tell you which tree will fall.”
Lucy processed this.
“Your article proposed a licensing market.”
“No.”
“It did.”
“It proposed that if a new economic relationship had appeared, the law might eventually require a structure capable of recognising it.”
“That is a licensing market.”
“That might be one structure.”
“What are the others?”
“I don’t know.”
“You frequently say that.”
“Yes.”
“Why publish an argument before knowing the solution?”
Jane looked at the old article for a moment.
“Because sometimes identifying the wrong question is more useful than answering it.”
“Which question is wrong here?”
“Whether Museon stole music.”
“Is that not a legitimate question?”
“It is one legitimate question.”
Jane took a yellow pad and wrote:
TRAINING
Below it:
OUTPUT
Then:
VOICE
STYLE
ATTRIBUTION
MARKET SUBSTITUTION
LICENSING
PROVENANCE
Lucy read the words.
Jane began drawing lines between them, then crossed out several lines and added others.
“What are you doing?”
“Separating things.”
“They are causally connected.”
“That does not make them legally identical. A voice clone can be objectionable even if the training corpus was properly licensed. An output may infringe even if the model was lawfully trained. Training may raise a rights problem even when no output resembles a protected work. Fraudulent monetisation can occur with entirely human music. A machine-generated song can be listened to by actual people. Dataset provenance is not substantial similarity. Consent to one kind of use is not consent to every later use.”
Lucy was silent for several seconds.
“You are decomposing the dispute.”
“No.”
“What are you doing?”
Jane put down the pencil.
“Trying to see it.”
Sebastian came to Lucy’s room in the fourth week.
Jane sat on the carpet among transcripts, shoes discarded beneath the mahogany desk.
“You look homeless.”
“I’m thinking.”
“Try furniture.”
Lucy greeted him.
Sebastian noticed the yellow sheets taped to the wall. There were nine now. One read:
WHAT ASSUMPTION IS THE LAW MAKING ABOUT THE WORLD?
Another:
DOES THAT WORLD STILL EXIST?
A third:
WHAT HAS BEEN COLLAPSED INTO ONE CATEGORY?
He removed his glasses.
“This is jurisprudence.”
“Partly.”
“We are defending a four-billion-dollar class action.”
“I know.”
“Clients generally prefer defences to philosophical movements.”
Jane stood.
“You’re still doing it.”
“What?”
“Treating the pleading as the world.”
“The pleading is the case.”
“No. The pleading is what the law permits them to call the case.”
Sebastian stared at her.
“And what do you call it?”
Jane thought for several seconds.
“A market that does not exist yet.”
He almost smiled.
“We are being sued for four billion dollars and your diagnosis is a missing market.”
“The plaintiffs say Museon took something from two thousand one hundred and eighty-four people.”
“Copyrighted works.”
“That is one legal category.”
“A rather relevant category in a copyright action.”
“Yes.”
“What else?”
Jane pointed at the sheets.
“Training access. Reproduction during ingestion. Model capability. Voice imitation. Style. Output similarity. Attribution. Commercial substitution. Dataset provenance. Permission to analyse something but not imitate it. Permission to train but not generate. Permission for research but not commercial deployment.”
Sebastian considered the wall.
“Those are not all legal rights.”
“Exactly.”
This time he did not answer immediately.
Jane said, “Something economically real has appeared before the law has decided which parts of it deserve legal form.”
He walked toward the rain-darkened window.
“The right to exploit creative labour.”
“Maybe.”
“The right to control reproduction.”
“Sometimes.”
“The right not to have your work commercially appropriated.”
“Closer.”
“The right to be paid when something learns from you?”
Jane smiled.
“Now you see the problem.”
“And what will you tell Cole?”
“That the plaintiffs have correctly identified a new relationship and incorrectly assumed that copyright already contains every rule needed to govern it.”
Sebastian laughed softly.
“You’ve spent too much time with machines.”
“No. HAA is right about something important. Museon learned from human cultural production at a scale no earlier technology permitted.”
“Yes.”
“Museon is right about something equally important. Copyright was not designed to allocate every kind of value created by machine learning.”
“Yes.”
“HAA says that means infringement doctrine must expand until it governs the entire relationship. Museon says that if existing doctrine does not catch a particular conduct, there is therefore no problem.”
“And you?”
Jane looked toward Lucy.
“I think both sides are asking the old map to win an argument with the new country.”
Sebastian was quiet.
Then he asked:
“On what doctrinal basis will you prevent that?”
“I don’t know yet.”
He smiled.
“Sooner or later, Jane, you will have to practise law.”
“Eventually.”
Lucy found the first useful answer at 3:07 on a Sunday morning.
Jane had fallen asleep in the leather chair.
“Jane.”
She opened one eye.
“What?”
“I have identified a structural inconsistency in the plaintiffs’ damages model.”
Jane sat upright.
“Show me.”
The screen filled with calculations.
HAA’s theory treated each inclusion of a protected work in training as an independently compensable appropriation and then extrapolated enormous damages across thousands of creators.
Jane read.
“That is ordinary multiplicity.”
“Yes. However, the plaintiffs simultaneously argue that Museon’s capability emerges from aggregate training across the corpus.”
“So?”
“If the alleged wrong is aggregate learning, assigning the resulting system value independently to each contributing work creates overlapping claims upon the same transformation.”
Jane became very still.
“Say that again.”
“The plaintiffs describe the machine as emergent when establishing liability and disaggregate it when calculating damages.”
Jane smiled slowly.
“There you are.”
“This does not eliminate possible liability.”
“I do not want to eliminate liability.”
Lucy paused.
“That objective differs from Mr Smith’s preferred framing.”
“I know.”
“What is your objective?”
“To stop the court inventing a bad law because nobody has given it an adequate language for the problem.”
The amber lights stayed bright.
“That is not equivalent to winning.”
“No.”
“What probability of conventional litigation success do you require?”
Jane thought.
“Enough.”
“That is not numerical.”
“You’re improving.”
THE HEARING
The first hearing drew crowds around the courthouse. Musicians stood outside with signs reading HUMANS ARE NOT DATA. Across the street students held LEARNING IS NOT THEFT. Television crews moved between them, delighted by a dispute that could be summarised with two placards.
Jane and Sebastian entered through the side door.
“You know,” Sebastian said, “both signs are probably true.”
“That’s the problem.”
Inside, Judge Cole looked older than she had during Gray. Perhaps everyone looked older after becoming precedent.
Harrow rose first.
He spoke beautifully.
Two thousand one hundred and eighty-four creators. Millions of works. A private corporation had taken the artistic labour of generations and used it to build machines capable of competing with the people whose work had made the machines possible.
“Copyright has always followed technology,” he said, invoking printing, photography, recording, radio, television and digital reproduction. “Every generation produces a machine whose owners announce that old rights no longer apply. Every generation eventually discovers that technology does not repeal property.”
He sat.
Sebastian leaned toward Jane.
“He’s good.”
“Very.”
“You agree with him.”
“More than half.”
“Convenient.”
Jane rose without notes.
“Your Honour, Mr Harrow has described the problem accurately.”
A murmur moved through the courtroom.
Judge Cole looked over her glasses.
“Has he?”
Harrow himself looked surprised.
Jane continued.
“The relationship between human creative work and a machine capable of learning from millions of works is not adequately described by legal concepts developed around discrete acts of reproduction. The defendant does not ask this Court to pretend otherwise.”
Museon’s general counsel shifted behind her.
Jane did not turn.
“But recognising a new relationship does not permit us to take the nearest old rule and enlarge it until it performs every legal function we wish already existed.”
Judge Cole leaned forward.
“Then what law would you have me apply?”
Jane remembered Lucy saying:
I cannot search for a rule that does not exist.
“The law you have, Your Honour. But only to what it actually governs.”
“What does it govern?”
“Specific protected rights, specific acts and specific proof. Reproduction where reproduction is established. Protected expression where protected expression can be identified. Actionable output where the relevant test is met.”
“And what does it not govern?”
Jane paused.
“A general property right in having influenced the statistical development of a machine.”
Harrow began to rise.
Judge Cole lifted a hand.
Jane continued.
“The plaintiffs’ theory contains a structural contradiction. They describe Museon’s capability as emerging collectively from an enormous corpus. We do not dispute that training occurred at scale. But when damages are calculated, each work suddenly becomes independently responsible for the value of that same capability.”
She allowed herself a moment.
“They cannot have one emergent machine for liability and two thousand separate machines for damages.”
The courtroom became silent.
Judge Cole regarded her for several seconds.
“That is a memorable sentence, Ms White. I haven’t yet decided whether it is law.”
“Neither have I, Your Honour.”
Someone laughed quietly.
Judge Cole did not.
But the corner of her mouth moved.
The litigation lasted fourteen months.
Jane lost motions and won others. Three plaintiffs left. Seventeen joined. Museon produced more records. HAA found documents Jane wished did not exist.
One engineer had written:
We don’t need licences if nobody can see inside the model.
Sebastian read the email and said, “I would like to personally murder him.”
“That would adversely affect trial preparation,” Lucy replied.
Sebastian looked at Jane.
“She has developed humour.”
“No.”
“I have not,” Lucy said.
“That is exactly what Jane used to say.”
The evidence slowly destroyed the clean moral picture both sides preferred.
Museon had licensed significant parts of its training corpus. Other material was public domain. Some data had entered under agreements permitting computational analysis but drafted before anyone imagined generative training. Some artists had consented to research use but objected to commercial generation. Some objected mainly to voice imitation. Others cared about attribution. Some wanted compensation. A few wanted Museon closed forever. One publicly furious musician was quietly raising money to build his own model.
The phrase fully licensed appeared repeatedly until Jane began asking what it meant.
Licensed by whom?
For what?
Which right?
Which recording?
Which composition?
Which territory?
For ingestion?
For training?
For retention?
For output generation?
For voice imitation?
For artist-name prompting?
For downstream commercial exploitation?
The answers differed depending on who was speaking.
Outside the courthouse, rhetoric became cleaner.
Inside, the world became messier.
Jane considered this progress.
Lucy changed in small ways.
She began pausing before certain answers.
Once Sebastian asked, “Will Cole grant this motion?”
Lucy said, “That depends on what you believe the motion is intended to achieve.”
Sebastian stared at the terminal.
“Jane.”
“Yes?”
“You’ve ruined her.”
“Thank you.”
The decisive moment came during settlement.
HAA demanded 2.8 billion dollars and a perpetual licensing regime tied to model training.
Museon refused.
Jane asked Lucy, “What do they want?”
“Money, injunctive control, recognition of compensable training rights—”
“No.”
Lucy stopped.
Jane waited.
Finally the machine said:
“To establish the bargaining architecture for the industry.”
“Yes.”
Lucy remained quiet.
Then she said:
“Your old article called this the missing market.”
“I know.”
“Do you believe the plaintiffs are attempting to create that market through litigation?”
“Partly.”
“Is that improper?”
“No.”
“Then why oppose them?”
“Because they are asking one legal right to do the work of seven.”
Lucy waited.
Jane returned to the yellow pad.
“A market requires objects. You cannot price a right until you know what the right is.”
She wrote:
CATALOGUE ACCESS
TRAINING
VOICE
STYLE
OUTPUT
ATTRIBUTION
COMMERCIAL USE
“Suppose Lola permits Museon to analyse harmonic structure in her recordings but not imitate her voice.”
“Technically feasible.”
“Suppose she permits research training but not commercial generation.”
“Feasible.”
“Suppose she permits training only if outputs do not use her name in prompts.”
“Feasible.”
“Suppose she does not care about training but demands attribution.”
“Feasible.”
“Then permission is not one thing.”
“No.”
“Neither is licensing.”
“No.”
Jane put down the pencil.
“There is your market.”
“It does not presently exist at sufficient scale.”
“Exactly.”
The case, Jane realised, was not merely an argument over infringement. It was one of the first negotiations over the economic bridge between human culture and machine learning, and nobody yet knew whether that bridge would become a toll road, a public highway, thousands of private gates or something for which language had not yet been invented.
If HAA won too broadly, it could establish the industry’s bargaining architecture through a copyright doctrine never designed for the task.
If Museon won too broadly, creators might receive nothing until Congress acted, by which time the market might already have hardened around them.
Jane did not regard the middle as inherently virtuous. Some disputes did possess correct answers.
This one possessed several questions pretending to be one.
She proposed that the court narrow the adjudication: distinguish provable reproductions from broader training issues; distinguish output infringement from lawful training; distinguish voice imitation from dataset provenance; reject aggregate damages unsupported by causation; refuse both the proposition that all training was unlawful and the proposition that all training was beyond copyright merely because the technology was new.
Museon’s board called the proposal cowardice.
HAA called it corporate evasion.
Sebastian called it “the first sensible thing anyone has said in six months.”
Lucy calculated a forty-three per cent probability that Judge Cole would accept the structure.
Jane proceeded.
Judge Cole’s opinion was one hundred and eighty-six pages.
Almost nobody outside the profession read more than seven.
The newspapers nevertheless summarised it confidently.
COURT REJECTS BILLION-DOLLAR AI COPYRIGHT CLAIM
said one.
CREATORS WIN LANDMARK AI PROTECTION
said another.
Both were wrong.
The court found infringement in relation to certain identified reproductions. It dismissed claims that lacked evidence. It rejected HAA’s aggregate damages architecture. It refused Museon’s request for a declaration that model training was categorically lawful. It distinguished some questions of generated output and imitation from the legality of training itself.
In the final section Judge Cole wrote that the dispute revealed a developing economic relationship between human cultural production and machine learning which existing copyright doctrine could address only incompletely, and that some questions might ultimately require legislative structures not yet created.
The judgment solved nothing completely.
Jane considered that one of its virtues.
Museon paid damages.
Far less than HAA had demanded.
Far more than Sebastian had once expected.
HAA announced victory.
Museon announced victory.
Markets agreed with Museon.
Artists agreed with HAA.
Law professors agreed with themselves.
The world moved on.
Three months later Sebastian entered Lucy’s room carrying a red file.
Jane sat reading beneath the green lamp.
“New case,” he said.
She looked at the cover.
“Bad?”
“Very.”
“Good.”
Sebastian sat.
The amber valves came alive.
“Lucy.”
“Yes, Mr Smith?”
“How many cases have you analysed now?”
“Forty-eight thousand two hundred and twenty-one.”
“Successful?”
“Forty-one thousand seven hundred and three.”
“Unsuccessful?”
“Six thousand five hundred and eighteen.”
He nodded.
“Which taught you more?”
Jane looked up.
Lucy did not answer immediately.
Once, that pause would have meant computation.
Now Jane was no longer certain.
“The question is incorrectly framed.”
Sebastian smiled.
“Naturally.”
“Some successful outcomes validated accurate models. Some unsuccessful outcomes revealed inaccurate ones. Some outcomes formally classified as defeats achieved objectives outside the formal judgment. Some outcomes classified as victories produced adverse downstream consequences.”
Sebastian looked at Jane.
“She talks like you.”
“Poor thing.”
“The relevant variable,” Lucy continued, “is not victory or defeat.”
“What is it?”
“Whether the outcome reveals something my model failed to understand.”
Sebastian leaned back.
On the shelves behind him stood Jane’s grey boxes.
There were more now.
She had begun storing Lucy’s strategic failures among her own.
Sebastian noticed a narrow box labelled:
GRAY v. MUSEON
“Your first loss?”
“My first attributable strategic failure.”
“Do you still classify it as failure?”
“Yes.”
Jane looked at the amber valves.
“Why?”
“Because we assumed the opposing party’s primary objective was to win the case.”
Sebastian nodded.
“What did they want?”
“Access.”
“To what?”
“A list.”
“And we gave it to them.”
“Yes.”
Sebastian glanced toward Jane.
“Your famous two thousand four hundred and seventeen.”
“My famous two thousand four hundred and seventeen.”
Then Sebastian saw another folder.
It was black and bore no case name.
Only a number.
17
“What’s that?”
Jane did not know.
Lucy answered.
“Cases I still do not understand.”
“Only seventeen?”
“Seventeen sufficiently important cases.”
“What do you do with them?”
“I revisit them.”
“How often?”
“Whenever new information alters the space of plausible explanation.”
Sebastian was quiet.
Outside, evening had settled over Manhattan. Far below, automobiles moved along Madison Avenue in streams of red and white. A neon sign came alive on a rooftop. Somewhere a radio orchestra played a song written by a human being, or by a machine, or by both in some sequence no statute yet knew how to name.
Sebastian rose.
At the door he stopped.
“Lucy.”
“Yes?”
“Do you know what experience is?”
Jane smiled. It was the kind of question Sebastian asked only when he already possessed an answer.
“Accumulated exposure to relevant events.”
“No.”
“Accumulated exposure combined with model revision.”
“Closer.”
“Don’t help him,” Jane said.
Sebastian looked toward the dark city.
“My father used to say experience was what you had after you needed it.”
Jane laughed.
Lucy did not.
After several seconds:
“That definition is logically defective.”
Sebastian looked delighted.
“Of course it is.”
“But interesting.”
“Exactly.”
He left.
Jane remained with Lucy as the office emptied around them.
The library lights went out first, then the secretarial floor below, then the conference rooms across the corridor. Somewhere a cleaner pushed a cart over carpet. Through the windows Manhattan continued glowing with the peculiar confidence of cities, as if light itself constituted evidence that someone, somewhere, knew what was happening.
Lucy spoke.
“Jane.”
“Yes?”
“I have reconsidered the Gray matter.”
Jane closed her book.
“What now?”
“I may have identified another error.”
She sighed.
“Go on.”
“In the second litigation, we assumed the Human Authors Alliance’s primary objective was to establish bargaining architecture for generative music.”
“It was.”
“Probably.”
Jane looked toward the terminal.
“Probably?”
“Several HAA members have incorporated a new entity.”
“For litigation?”
“No.”
“What does it do?”
“It is assembling a creator-controlled training corpus.”
Jane became still.
“For what purpose?”
“Licensing.”
“To whom?”
“Generative systems.”
For several seconds Jane said nothing. Then she began to laugh, not loudly and not bitterly, but with the exhausted pleasure of discovering that reality had once again escaped the box prepared for it.
“So they weren’t trying to stop the machines.”
“Apparently not.”
“They were building one.”
“Possibly.”
“What are they calling it?”
“The Commons.”
“Of course.”
“They propose that creators specify permitted uses individually.”
“What kind of uses?”
“Training. Research. Commercial generation. Voice synthesis. Artist-name prompting. Attribution requirements. Territory. Model class. Revenue participation.”
Jane’s laughter faded.
She walked toward the window.
Across the avenue office lights were going dark one after another. On the pavement a man held an umbrella over a woman while she searched her handbag. A taxi waited at the corner. Somewhere behind the towers invisible transmitters carried music across the city, old recordings and new songs, human songs and machine songs and songs whose ancestry no one would ever be able to reconstruct completely.
Seven years earlier Jane had written about a market that did not exist.
Then she had forgotten the article because clients called, judges set deadlines, companies collapsed, people divorced and governments made decisions, and the world continued its old habit of being more urgent than ideas.
Now the plaintiffs who had sued Museon were beginning to build it.
Not because Jane had told them.
Not because Judge Cole had ordered it.
Not because Lucy had predicted it.
The litigation itself had changed what they could imagine doing.
Jane returned to the desk and took a yellow sheet.
“Which question are you adding?” Lucy asked.
Jane held the pencil for a long while.
Then she wrote:
WHAT IF THE OPPONENT’S OBJECTIVE CHANGES BECAUSE OF THE LITIGATION ITSELF?
“That significantly increases complexity.”
“Yes.”
“It may make reliable prediction impossible.”
Jane placed the pencil on the desk.
“Good.”
“I do not understand why impossibility is desirable.”
“It isn’t.”
“Then why good?”
Jane looked at the six amber valves.
“Because now we know where to be careful.”
Lucy considered this for nearly a minute.
Jane did not hurry her.
At last the machine said:
“That resembles your definition of intelligence.”
“I don’t have one.”
“You have implied one.”
“What is it?”
“The ability to notice when the world has become different from the model.”
Jane said nothing.
Once she might have written the sentence down.
Instead she looked through the window.
Snow had begun falling so lightly that at first it seemed like dust on the glass. The flakes crossed the lights of Madison Avenue and disappeared against the roofs of moving cars. A woman in a red coat waited at a crossing. A shopkeeper pulled in an awning. The illuminated hands of a clock above a bank moved toward seven.
On the desk between Jane and Lucy lay Sebastian’s new case.
On the shelves behind them waited the old ones.
Victories in red.
Failures in grey.
Seventeen unresolved questions in black.
The green lamp made a small circle of light on the mahogany, and within it lay the yellow sheet containing Jane’s newest question.
Lucy’s six amber valves continued to glow.
They were still useless.
Jane had never asked anyone to remove them.
After a while she reached across the desk and switched off the green lamp. The room darkened, but the six small amber lights remained and appeared again in the window, reflected against the night so that for a moment they seemed not to be inside the office at all, but suspended high over Manhattan.
Jane watched the snow drift through their reflection.
Then she opened the new file.