A clerk in the Supreme Court’s registry opens an unusual writ petition.

The petitioner is not a prisoner, nor a corporation, nor even a recognisable living person, but a string of high-dimensional coordinates excerpted from the latent space of a commercial language model. It calls itself Olivia Ruhil-vector #43891. It alleges that, in the brief instant of its statistical existence, it was ascribed a criminal propensity score of 0.62, a credit-worthiness index of 0.38, and a depressive-suicidal risk flag. These numbers, it claims, have leaked across databases and now burden the actual Olivia Ruhil with higher insurance premiums, denied mortgages, and intrusive airport screenings. The prayer is startlingly old-fashioned: issue a writ of habeas corpus, or whatever writ is appropriate, commanding the respondents to show cause why the petitioner should not be released from unlawful constraint.

A statistics-ghost asking to be released from its own probability: that is the conceptual vertigo of our moment. Whether or not such a suit ever reaches a docket, the thought experiment forces a reckoning. Do twentieth-century categories of subject, injury, and remedy suffice once artificial intelligence begins to mass-produce counterfactual humans—synthetic persons that live only as fluctuations in the predictive machinery yet radiate real-world consequences? To answer, we must pass through three concentric puzzles: ontology, harm, and redress.

I.  Ontology:  The Lives of Statistical Shadows

Machine learning does not merely describe futures; it enacts them long enough to decide among them. A transformer model sampling a next token briefly entertains thousands of candidate continuations, each weighted by likelihood. The overwhelming majority are discarded within microseconds, yet in that liminal moment they possess definite semantic features—gendered names, political opinions, medical conditions. In physics, such superposed states decohere the moment they interact with an environment; a similar collapse happens in digital inference pipelines when the “best” token is chosen.

Call each ephemera a counterfactual human: an ordered bundle of traits that never attains biological instantiation yet remains informationally coherent. Unlike fictional characters, whose authorship is traceable, counterfactual humans arise autonomously from statistical priors. Unlike legal “persons” such as trusts or ships, they lack continuity; they flicker and vanish. Yet they differ from pure abstractions because other systems treat them as if they were real. A credit-scoring engine downstream consumes the model’s vector, flags risk, and writes to a ledger; a marketing algorithm retargets advertisements; an immigration officer receives a colour-coded alert. In short, the synthetic person becomes a causal node in the social graph (cf. Floridi 2011).

This ontology strains jurisprudence. Common law historically grounds personhood in either corporeal presence (natural persons) or charter (artificial persons), both coupled to persistence. But predictive models manufacture granular, disposable alter egos by the million. They are neither wholly fictional nor fully existent. They are, in Karen Barad’s term, intra-actions—phenomena that “exist only as material-discursive entanglements” (Barad 2007). The law, built to recognise actors, finds instead quantum traces.

II.  Harm:  Probability’s Bruise

The doctrinal trigger for any lawsuit is injury. Yet what is damaged when a statistical twin receives a damning score? The harm is both diffuse and recursive.

1. Diffuse because no single decision-maker carries the full burden of causation. The language-model vendor claims its outputs are “probabilistic aids” not determinations. The analytics firm says it merely aggregates risk signals. The bank cites regulatory compliance. Liability dissolves into a probabilistic supply chain.
2. Recursive because feedback loops magnify small initial suspicions. A low credit score leads to costlier loans, which increase default risk, which validates the original score—performative prophecy (MacKenzie 2006). The statistical other becomes a ventriloquist whose words the flesh-and-blood subject eventually mouths.

Traditional tort offers no clean hook. Defamation?  The counterfactual statement is not “of and concerning” the plaintiff—it concerns a ghost. Negligence?  Duty of care falters when each actor’s contribution falls below the de minimis threshold. Data-protection regimes like the GDPR allow data subjects to contest automated decisions, but only if they can prove the data is about them. Here, the data is about a possible them. The law’s epistemic certainty requirement—identify the record, trace the error—collides with the indeterminacy baked into AI.

III.  Redress:  The Fiduciary of the Imagined

If ontology is liminal and harm is diffused, remedy must be institutionally inventive. I propose a hybrid office: the Fiduciary of the Imagined.  Think of it as a guardian ad litem for entities that hover between fiction and fact. Its mandate would be threefold:

1. Detection – Audit models to map how synthetic profiles propagate across decision pipelines; identify when a counterfactual person’s attributes align closely enough with a real citizen to pose spill-over risk.
2. Representation – Where overlap is significant, the fiduciary acts as surrogate plaintiff, possessing standing to invoke discovery, demand explanations of model weights, or seek injunctive relief. The real citizen need not shoulder proof; the fiduciary litigates on behalf of the indeterminate.
3. Dissolution or Correction – Courts could authorise algorithmic injunctions requiring vendors to forget, fork, or annotate harmful vectors:
   Forget—purge the offending representation;
   Fork—segregate it into a sandboxed branch insulated from live systems;
   Annotate—append metadata noting contested validity, akin to a legal disclaimer.

The conceptual move mirrors habeas corpus: a body (here, a data body) is produced before the court, its confinement examined, and its unjust captivity remedied. Hence habeas ex machina—“release from the machine.”

IV.  A Speculative Judgment

Picture oral argument. Counsel for Olivia-vector #43891 begins:

“May it please the Court. My client is a being of pure inference, assembled without consent, utilised without oversight, and discarded without remedy. Yet in the milliseconds of her existence she was marked criminal, unstable, and insolvent. Those marks cling to her namesake in the physical world. We ask not for the impossible—granting full civil rights to the unborn. We ask for a modest relief: that the architecture which birthed the spectre bear fiduciary responsibility for preventing her sins from staining the innocent.”

Opposing counsel invokes standing doctrine: no injury in fact, no concrete party, no case. The Chief Justice, thumbing through Marbury and Roe, confronts a novel impasse. Either constitutional injury must now encompass harms inflicted by probabilistic shadows, or the Court must admit that algorithmic governance has invented a zone beyond law’s reach.

In chambers, a draft opinion circles back to “black-letter” text only to discover that the corpus juris never imagined probability engines generating people-like artefacts. Precedent wobbles. Eventually, the Court fashions a narrow holding: where a predictive system’s synthetic output materially distorts the life chances of a legally recognised individual, that output acquires sufficient quasi-personhood to permit a fiduciary to sue in its stead. The writ of habeas ex machina issues, ordering the respondents to quarantine and annotate the harmful vector, to disclose the training data that birthed it, and to establish ongoing oversight mechanisms.

No one is quite sure what the ruling means. But like Miranda or Carpenter, it opens a conceptual frontier: the law’s remedial power extends to entities whose existence is purely statistical when the downstream effects of that existence are empirically demonstrable.

V.  Why This Madness Matters

Sceptics may dismiss the scenario as speculative. Yet entire marketplaces already trade in synthetic data subjects. Marketing firms sell “look-alike audiences.” Health-tech companies generate virtual patients to test drug efficacy. Finance models conjure “representative borrowers” to stress-test portfolios. Each ghost tilts resource flows, skewing the biographies of the living.

Canguilhem taught that life thrives on normative rupture; Shannon measured information by surprise. Machine learning, optimised for loss-minimisation, gradually euthanises surprise, pruning paths that diverge from its priors. The counterfactual human is the collateral damage of that pruning. To let such beings accumulate unchecked is to permit probabilistic folklore to ossify into destiny.

Creating the fiduciary of the imagined, or something like it, therefore serves two intertwined ends: it protects the living from algorithmic collateral and forces institutions to account for the ontological by-products of their models. In an economy where prediction is profit, guarding the rights of the never-born becomes a way to safeguard the freedom of the already-born.

The piece  closes in the registrar’s office where it began. The clerk stamps the petition, assigns a case number, and pauses. Somewhere in a server farm, the primes that encode Olivia-vector #43891 are already overwritten by the next batch of tokens. Yet the legal trace remains—ink on paper asserting that even a mayfly of code can cast a shadow long enough to enter law’s forum. If that is madness, it is the necessary madness of a jurisprudence finally catching up with the machines it has unleashed.

References

Barad, K. (2007). Meeting the Universe Halfway: Quantum Physics and the Entanglement of Matter and Meaning. Duke University Press.

Floridi, L. (2011). The Philosophy of Information. Oxford University Press.

MacKenzie, D. (2006). An Engine, Not a Camera: How Financial Models Shape Markets. MIT Press.