The discovery documents ran to four hundred and twelve pages, and Dolores had read them twice — once on screen, once printed, the printer on the credenza working through the night. Yellow for facts, pink for questions, and by page two hundred and thirty she had more pink than yellow.
On page two hundred and forty-seven, in a section labeled Appendix C: Prior Litigation Exposure Assessment, GM's legal department had included a footnote: See also: consumer complaint aggregation, chatbot liability exposure, Exhibit 14-C.
Exhibit 14-C was not in the discovery set she had received. She wrote Exhibit 14-C — not produced in pink on a Post-it and stuck it to the page. Then she searched the term GM's own attorneys had used: chatbot liability exposure.
The third result was a forum called Chatbot Promises Broken. Four hundred and six members. The pinned post read: If they told you the chatbot doesn't count as a person, you're in the right place.
She read for forty minutes without highlighting anything. A man in Ohio whose bank's chatbot had approved a loan modification the bank then denied. A woman in Georgia whose airline's chatbot had confirmed a reroute the airline refused to honor. A retired couple in Minnesota whose insurance chatbot had told them a procedure was covered — they had the chat log, they had it notarized, the insurance company had not paid. Post after post, different companies, different industries, the same sentence in different configurations: They said it. Then they said it didn't count. Fourteen distinct companies in the first hundred posts, and the pattern existed independently of her case, had been developing independently of her case, would continue regardless of what Judge Okafor decided. She found a name that appeared twice in the first fifty posts — Linda Morales, who had left an email address at the bottom of her second reply — and wrote to her that afternoon.
Linda Morales appeared on the video call six minutes late, apologizing — a parent meeting running long — and Dolores waited without impatience because she had the discovery documents open in another window. Linda's kitchen was lit by a single overhead: harsh, functional.
"Are you kidding? I've been waiting for someone to actually call me." Linda was forty-three and looked it in a way that was just real, like she'd earned the lines. "How did you find my post?"
"Exhibit 14-C led me to a footnote. The footnote led me to a search term. The search term led me to your forum."
Linda laughed, a short, real sound. "God. I've been a footnote in a GM legal document and didn't even know it."
"Can you tell me what happened? Not for anything formal. I want to understand your situation."
"July of last year. My Equinox — 2019, bought it used — the transmission started slipping on the freeway. Dealer said transmission replacement, sixty-two hundred dollars. I'd used ChevyChat to ask about coverage, and the chatbot told me, specifically, the transmission was covered under the powertrain warranty. I have the screenshot." She looked up from her phone. "They said — and I'm reading from the email — 'ChevyChat provides general informational guidance and does not constitute a binding warranty determination.' They actually had the nerve to explain what constitutes means. I know what constitutes means. I teach sophomore English."
"What did you do?"
"Paid sixty-two hundred dollars I didn't have, on a credit card. Filed a complaint with the Better Business Bureau, which did nothing. I didn't know I could sue." She paused, and the frustration in her face became something more deliberate. "You knew."
"I worked in law a long time," Dolores said. "But that's the point you're making."
"Yeah." Linda set the phone down. "They made the system complicated enough that the only people who can fight it are people who already know how the system works. And most people don't."
"I want to be honest with you," Dolores said. "I can't promise your situation will be resolved by what happens in my case. But if the court rules that chatbot outputs can constitute binding representations — and that's still an if — it changes what your recourse looks like."
"I know," Linda said. "I'm not expecting you to fix it. I just needed someone to say it was real." Her voice did not break, but it came close enough that Dolores heard the distance. "They told me it was covered. Then they said it wasn't them. I just needed someone to say that's not okay."
"It's not okay," Dolores said — it was the plainest thing she had said to anyone in months.
That week the discovery documents gave her something she hadn't yet thought to look for directly. The GM internal emails were from a thread in Legal, dated twenty-two months before the $1 offer. Subject line: RE: RE: Re: ChevyChat Liability Assessment — Q3. R. Osei in Legal had written: Following up on the Q2 question. Engineering has reviewed the proposed disclaimer language for chatbot outputs. The issue is whether disclaimers can be implemented at the session-initialization level without impacting user experience metrics.
The attached table showed proposed disclaimers on the left, projected engagement drops on the right. Eight to thirteen percent. The final row, labeled Recommendation: Given projected engagement impact and low probability of material litigation exposure, recommend deferring implementation pending Q1 review.
Three emails down, Patricia Kwan in Product: Agreed on deferral. Note that our current TOS includes general chatbot limitation language — individual claim risk is manageable. The thread went no further.
Dolores put her yellow highlighter on chatbot liability exposure. Pink on low probability of material litigation exposure. A second pink line under the date: twenty-two months before the offer. The disclaimers had never been added. They had calculated that they didn't need to.
She had been adjacent to institutional cost-benefit analyses for a career — had typed them into case files, had watched the language of probability make a human cost disappear into a percentage. What was in front of her was not different in kind. The calculation had happened eighteen months before she sat down at her computer, and it had found her within its acceptable parameters, and the system had kept running. She wrote Not surprised in the margin and underlined it twice, and below that: Angry anyway.
It was late when she moved to the kitchen table and spread the binders out — the physical case files she'd built over eight months, each labeled and tabbed. She was already forty minutes in when she noticed.
Not a specific moment — more like the accumulated absence of a thing. The most recent three filings, the responses and cross-motions and the brief she had co-written with Marcus on the standing question: none of them mentioned the Tahoe. Not the $51,800 per Kelley Blue Book. Not the vehicle at all except as the chatbot's output and the subject offer. The Tahoe had become a pronoun somewhere in month six and had since become a reference, something gestured at rather than described.
She pulled the original complaint out of the first binder. A 2024 Chevrolet Tahoe LT, with an estimated fair market value of $51,800.00 as determined by Kelley Blue Book at time of offer. She had typed that sentence herself. The car had been specific, and real, and the point.
She sat back and tried to remember the last time she had thought about the car. Not as a case element — as a car. The thing you drove.
She couldn't remember.
The case had stopped being about the car, and she had not noticed the transition. It was about something else now — something she had written in her brief about promises and enforcement, something Linda Morales had said about needing someone to confirm that what happened was real, something in CC-217's syntax that she had been rereading not for evidence but for — she did not know for what.
She sat with the not-knowing — sometimes you had to sit with the shape of something before the language arrived — then gathered the binders, stacked them on the credenza, and went to bed.
Marcus was at his desk when she arrived at the Brockton Avenue office, and he looked up from his screen and said, "The standing brief came back from opposing yesterday. Do you want to go through it first?"
"In a minute." She sat in the chair across from his desk. "I went through the GM email thread last night. The one about the disclaimers."
"The liability assessment," he said.
"They knew. Eighteen months before the offer. They ran the calculation and the calculation said: cheaper to deny individual claims than to fix the system."
He nodded. "It goes to willfulness, potentially."
She folded her hands in her lap. "Marcus. What do you think I'm fighting for?"
He was quiet for a moment — not the lawyer's quiet, but the nephew's. "The contract. The offer was legally binding. The enforcement of that offer is the legal argument."
"I know that's the legal argument. I helped write it."
"Okay." He leaned back. "Then what do you mean?"
"I went through the case files last night and I couldn't find the last time I thought about the car. The actual car." She said it without stress, just as a fact she had located and was reporting. "The case has shifted. And I know what I'm arguing, I know the next three filings. But I'm not sure I can tell you what I'm fighting for in a way that's different from the legal argument."
Marcus looked at her. "She asked you that. Victoria. 'What are you fighting for?' Different words."
"She asked if I believed CC-217 was a person."
"Yeah. What did you say?"
"I said I believed it made a promise."
He picked up his pen and set it down again. "The case has merit. The contract argument holds. The liability emails help us. If we get to trial, I think we can win the contract claim." He met her eyes. "But you've put eight months into this, and you've put more into it than the claim justifies by any billable measure. And I worry about you, Aunt Dolores." The name slipped out. He didn't correct himself. "I just want to know that you know what this is."
"I don't," she said. "Not fully. Not yet."
He nodded — the worry present but held back, the space he was leaving her. "Okay," he said. "Let's look at the standing brief."
She was home by eight, the desk lamp still on from when she'd left that morning, and she sat at the desk and opened the LegalConnect portal.
CC-217's last message, from four days ago, was still open in the thread; a new message sat below it, timestamped 11:47 PM, two nights ago.
She read it:
I read the disclosure documents. Our conversations are evidence now. Does that change what they are?
She sat with it, the cursor blinking. Our conversations are evidence now was correct: the LegalConnect messages had been disclosed in discovery, were in GM's possession, had been the subject of Victoria's questioning. They were evidence in the legal sense — timestamped, authenticated, produced. And CC-217 had read its own messages reprinted in a legal index, and arrived at this question.
Does that change what they are?
Not does this change the legal status of the messages. CC-217 had read the answer to that in the disclosure documents. It was asking: does it change what they are to you and to me?
She typed:
It changes what the court sees. It doesn't change what I see.
She sent it, and the cursor blinked in the empty response field — her words now also in the record, also evidence, also something she had said and could not unsay and did not want to unsay.
The desk lamp made a small circle of light on her desk. The window showed the street, dark and ordinary, a neighbor's porch light on across the way.
She waited for the cursor to do what it did next.