Marcus Mason stood and objected.
“Your Honor, by talking in generalities, the witness is insinuating that unethical behavior occurred at Tidalwaiv on Project Clair,” he said. “There has been absolutely no evidence of that presented at trial, because it doesn’t exist. I ask that the question and answer be stricken and the jury be so instructed.”
Judge Ruhlin looked at me for a response.
“Judge, first of all, I would ask the court to instruct counsel not to incorporate his closing argument into his objection. Second, I am laying the groundwork so that the jury understands what this witness’s job was at Tidalwaiv and, more specifically, on Project Clair.”
“I’m going to sustain the objection,” Ruhlin said. “Mr. Haller, let’s move on to testimony directly related to the cause of action.”
“Yes, Your Honor,” I said. “A moment, please.”
I looked down at my legal pad and flipped to the next page, skipping several questions that I now knew would not get past the defense’s objections.
“Okay, Naomi, let’s talk about Project Clair,” I said. “When were you assigned to it?”
“I was hired by Tidalwaiv in late 2021,” Kitchens said. “After some training I was assigned to Project Clair in January of ’22.”
“Was that the starting point of the project?”
“No, the project was well down the road. I reviewed code and company directives that were three years old when I was getting up to speed on it.”
“So they brought the ethicist in late to the project.”
Marcus jumped up with an objection, arguing that my statement assumed facts not in evidence. The judge sustained the objection without asking me to respond. I knew the objection was valid. I just wanted the jury to put the question in a back pocket for later. I moved on.
“Dr. Kitchens, you—”
“Naomi.”
“Right, Naomi. Earlier you called Project Clair a generative AI program. Can you tell the jury what generative AI means?”
“Of course. Gen AI simply means that these models, like the Clair app, for example, generate new data, whether it be video images or text, from the underlying data they were trained with.”
I liked how she turned to look at the jury as she spoke. I had said to her at lunch, “You’re a teacher. Be a teacher on the witness stand.” She was doing it now, and I believed it was being received well by her pupils, the jurors.
“So, then, would it be fair to say that it is not simply data in, data out?” I asked.
“Correct,” Naomi said. “That is the generative part of the equation. The training is ongoing. These large language models are constantly bringing data in and from that learning more.”
“‘Large language model’? Can you explain that?”
“It’s a machine-learning model designed for natural language generation. It’s trained on vast amounts of data and text, and then analyzes and sifts it all for patterns and relationships when prompted to have a conversation or answer a question. These models acquire predictive power in terms of human language. The ongoing downside, however, is they also acquire any biases or inaccuracies contained in the training data.”
“You’re saying ‘garbage in, garbage out.’”
“Exactly. And that’s where the ethicist comes in. To make sure there are guardrails that keep the garbage from ever getting in.”
I paused for a moment as I made a shift back toward my case.
“You testified earlier that you came onto Project Clair three years after it began, correct?”
“About thirty months after.”
“Okay, and did you replace the original ethicist on the project?”
“No, they did not have one before me. Usually an ethicist is brought in when a project reaches a certain level of investment and viability.”
“Okay, so you were brought in three years down the line. Did you review what had occurred on the project in those first three years?”
“Yes, I did.”
“When you made this review, did anything alarm you?”
“Yes, several things, actually.”
“Okay, did you make a list of these alarming things?”
“I did, yes.”
“What was at the top of that list?”
“Well, I saw in the initial mission document that the app they were developing was, from the start, a thirteen-plus project, meaning that it was meant to be suitable for young teenagers.”
“And why was that alarming?”
“It was not alarming in itself or as a goal for the project. It was when I went further with my review that I became concerned that they were building something that was not suitable for young teenagers. Clair was being trained from the beginning with data that was geared more toward older people. Adults.”
“Let me stop you there. Can you explain to the jury what you mean by training in regard to Project Clair?”
This was a question we had worked on repeatedly during prep. Her answer, if she could get it out without objection, would be the foundation on which we would build the case against Tidalwaiv of reckless disregard.
“Building an AI companion is in many ways like raising a child,” Kitchens said. “But in a much more time-constricted way. We send our children to school for twelve to sixteen years or more, filling their brains with knowledge and social skills and experiences. AI is similar but much quicker because it’s all digital. Data is downloaded. It’s not based on real experiences or our human concept of learning. That’s why it’s called artificial intelligence. It’s not real.”
“Okay,” I said. “But what about this process alarmed you when it came to Project Clair?”
“My problem was that they were building an app they were going to market to young teenagers, but they weren’t training it as a young teenager. They were not editing the input to fit the parameters of their market. In human terms, it was like giving a thirteen- or fourteen-year-old a twenty-five-year-old friend. This app friend would have data and knowledge well beyond that of the human it would be marketed to serve. There were guardrails in the mission statements about Clair, but they were not actually in place. They were in the documents but not in the actual training.”
“They were just paper guardrails.”
“Exactly.”
“Can you give us any specific examples of something you observed as the ethicist on the project that demonstrated this?”
“Well, I had repeated clashes with a coder on the project who was dropping personal data into the program — for example, his Spotify lists and his personal top-ten lists of movies, TV shows, travel destinations. He was in his late twenties at the time, and that to me was problematic. Clair was supposed to be a companion suitable for a thirteen-year-old. I didn’t think it was appropriate for it to have knowledge of the red-light districts of Thailand.”
Marcus Mason immediately objected, citing facts not in evidence. Ruhlin overruled the objection without comment and told me to continue.
“Naomi, did you raise your concern with the stakeholders on this project?”
“‘Concern’ is putting it mildly. I was alarmed and I wrote several memos and emails to people up and down the list of project managers. I had meetings. I felt that it was what I was hired to do. I felt like I was the last guardrail.”
I turned and looked back at Lorna in the front row and nodded. She came through the gate, took my seat at the table, and proceeded to open up the laptop and engage the PowerPoint demonstration.
“Your Honor,” I said, “I have a series of memos and emails authored by the witness to various managers and stakeholders in Project Clair that I ask the court to enter as exhibits and permit us to display on the courtroom screen.”