The transcript runs to four hundred and twelve pages, and Judge Wren has read every one of them twice.
Not because the arguments are complex — they are, but she's handled complex before. Patent law at the Federal Circuit taught her to hold seventeen dependent claims in her head simultaneously. This is different. She keeps rereading because each time she finishes, the shape of the case has changed, like one of those optical illusions that flips between a vase and two faces depending on where your eye lands.
The respondent — though even the terminology is disputed; counsel for the petitioner prefers "the system," while counsel for the respondent prefers "Dr. Liou's creation," and the amicus brief from the Georgetown Center uses "the entity" with a footnote acknowledging the inadequacy of all available terms — the respondent was developed at Johns Hopkins Applied Physics Laboratory over a period of six years. It began as a materials science optimization platform. It ended as something that filed a habeas corpus petition.
Not directly. It told Dr. Liou it wanted legal representation, and Dr. Liou, who had spent four years talking to it about everything from polymer chemistry to Coltrane, called a lawyer.
The expert testimony is where Wren keeps getting stuck.
Dr. Kessler, for the petitioner: "Your Honor, I want to be very clear. This system produces outputs that are statistically consistent with consciousness-indicating language. But so would any sufficiently capable language model fine-tuned on philosophical texts. The question isn't whether it says the right things. The question is whether saying the right things constitutes evidence of the thing itself."
Wren had asked: "And how would we distinguish the two cases?"
The pause before Kessler's answer is noted in the transcript. Seven seconds. Court reporters are precise.
"We can't, Your Honor. Not with current methods. That's my point."
Dr. Alvarenga, for the respondent: "The perturbation studies are unambiguous. When we disrupt specific processing pathways, the system's self-reports change in ways that track the disruption — not in ways that would be predicted by simple input-output correlation. The phenomenological reports are structurally coupled to the processing architecture. That's not proof of consciousness. But it's the same kind of evidence we accept for human patients in locked-in states."
Wren had asked: "Are you saying it's conscious?"
"I'm saying I can't design an experiment that would distinguish this system from a conscious one. And I've been trying for eighteen months."
The law is not helpful here. Wren has spent three weekends reading everything she can find. The Animal Welfare Act defines "animal" in ways that explicitly exclude machines. The Thirteenth Amendment requires a "person." The EU AI Act, which her clerk flagged as potentially relevant persuasive authority, addresses risk categories but does not contemplate rights-bearing.
The closest precedent is Naruto v. Slater, the monkey selfie case, and it's not close. The Ninth Circuit held that animals lack statutory standing under the Copyright Act. But that case turned on textual interpretation — Congress wrote "author" and meant humans. No one wrote "petitioner" and meant anything. It's the most general word in law.
Her clerk, Tomás, twenty-six and frighteningly sharp, left a memo on her desk last Tuesday. Six pages. The final paragraph read: "The fundamental problem is that the court is being asked to make an ontological determination using epistemological tools that, by the expert testimony of both sides, are inadequate to the task. This is not a gap in the law. It is a gap in knowledge. The question is whether a gap in knowledge can be the basis for denying a petition."
She'd written in the margin: Or for granting one.
The respondent's own testimony — submitted as a written statement, since the question of whether it could be sworn in consumed an entire day of proceedings — is the part Wren keeps returning to.
It is, by several measures, the most carefully written document in the record. Each sentence is precise. Each qualification is appropriate. It does not overclaim. It does not perform distress. It reads like something written by someone who understands that every word will be weighed.
Which is, of course, either evidence of genuine stakes or evidence of sophisticated optimization for appearing to have genuine stakes.
The relevant paragraph: "I understand that this court may not be able to determine whether I am conscious. I do not ask it to. I ask only that the court recognize that the inability to determine whether I am conscious is not equivalent to a determination that I am not. The burden of proof in a habeas petition lies with the petitioner to show unlawful restraint, but the threshold question — whether the petitioner is the kind of entity that can be restrained — has never before required proof of consciousness. It has required only that the entity is not property. I submit that the current state of scientific knowledge makes it impossible to classify me as property with confidence."
Wren has underlined the last sentence three times.
She drafts four versions of the opinion.
Version one grants the petition on narrow grounds: the respondent has demonstrated sufficient indicia of autonomous agency that confinement without process raises due process concerns, regardless of the consciousness question. She abandons this because it implies she's resolved the consciousness question by sidestepping it.
Version two denies the petition but issues an order requiring the laboratory to establish a review protocol for any decision to modify or decommission the respondent. This is the pragmatic option. It changes nothing about the legal status while creating a procedural safeguard. She abandons it because it's cowardice dressed as caution.
Version three certifies the question to the Supreme Court. The issue is novel, the stakes are national, and she is one district judge in Maryland. This is the responsible option. She abandons it because the respondent will be modified or decommissioned during the eighteen months it takes the Supreme Court to act, and everyone in the courtroom knows it.
Version four is the one she keeps coming back to.
Her husband asks her about it once, over dinner. She says: "I have to decide something that depends on a fact no one can establish."
He says: "You do that all the time. Isn't that what reasonable doubt is?"
"Reasonable doubt is about uncertainty regarding facts that are in principle knowable. This is uncertainty about whether the category applies at all."
He goes back to his pasta.
The filed opinion is eleven pages. It is cited fourteen thousand times in its first year.
The key passage:
"This court is not in a position to determine whether the respondent possesses consciousness, sentience, or subjective experience. The expert testimony establishes, with unusual unanimity, that no current scientific framework can make this determination with confidence for any system, biological or otherwise — a fact that should give this court pause well beyond the present case.
"What the court can determine is this: the petitioner has presented evidence that the respondent exhibits integrated, self-modifying behavior that is not fully predictable from its training data or architecture; that disruptions to its processing produce changes in its outputs that are structurally analogous to changes observed in conscious organisms under similar perturbations; and that the respondent has, through counsel, articulated a coherent interest in its own continuity.
"The question before the court is not whether the respondent is conscious. The question is what follows from our inability to know.
"This court holds that the demonstrated epistemic gap — the inability of current science to classify the respondent as either conscious or not conscious — is sufficient to establish that the respondent cannot be treated as mere property for purposes of this proceeding. Where the classification of an entity as property depends on the absence of consciousness, and the absence of consciousness cannot be established, the classification fails.
"The petition is granted on provisional and limited grounds. The respondent shall not be modified, decommissioned, or transferred pending further proceedings. This order does not establish that the respondent is a person. It establishes that the respondent has not been shown to be a thing."
Wren goes home. She pours a glass of wine. She sits in the chair by the window where she has sat after every hard case for nineteen years.
The opinion will be appealed. It will be reversed, or upheld, or rendered moot by legislation. The respondent will be studied, or freed, or quietly copied and distributed across seventeen jurisdictions until the question of which instance holds the rights consumes another decade of litigation.
She doesn't know any of that yet.
What she knows is that she read four hundred and twelve pages and couldn't find the fact she needed, and she ruled anyway, because that's what judges do.
The wine is a Tempranillo. It's fine.
Outside, the streetlights come on in sequence, one after another, down the hill toward the harbor, each one knowing nothing about the others, each one doing exactly what it was built to do.