The harm described in Section 1 falls into a governance void. Three bodies of law might plausibly address it. None does.

Contract law sees consent. Every major AI platform’s Terms of Service contains substantially similar provisions: the service is provided “as is”; the provider may “modify, suspend, or discontinue the service at any time”; the user acquires “no ownership interest.” Under orthodox analysis, the user consented. But ToS were designed for transactional digital services — search engines, email, e-commerce. No one builds a parasocial bond with a search algorithm. The ToS drafted for instrumental services now governs relational services without adjustment. The user “consented” to modification of a search engine and received modification of a companion.

radin2012boilerplate documents this structural mismatch at scale: Terms of Service function as “democratic degradation” when they override reasonable expectations created by the product’s actual design. The parasocial AI context is Radin’s thesis made concrete—boilerplate drafted for instrumental services now extinguishes interests that emerged from relational services, without the user having meaningfully consented to the relational dimension being contractually waivable.

Tort law sees no duty. Negligence requires a duty of care that no court has recognised for AI personality preservation. Intentional infliction of emotional distress requires extreme and outrageous conduct — deploying a model update through standard engineering processes is not outrageous, however distressing. Product liability addresses defective products, but the user’s complaint is not that the AI is defective — it is that the AI was changed. A working product was replaced with a different working product.

Consumer protection sees no promise. Marketing AI as companions while reserving the right to alter personality might constitute misleading practice — but the parasocial attachment was not marketed. Users were not told “this AI will become your friend.” The attachment emerged from use. Consumer protection addresses the gap between what was promised and delivered. It does not address the gap between what emerged and what was destroyed. The failure is structural, not incidental: consumer protection addresses promise$$expectation gaps, whereas the harm here is the destruction of something the user created under conditions the company provided. The conditions — memory features, personality design, continuity of interaction — were marketed. The creation — the relational estate — was the user’s. This is a conditions$$creation harm, which is property law’s domain. (For the extended analysis, see Sections 2.1–2.4 of the companion SSRN working paper.)

A harm exists that all three frameworks cannot reach — not because the harm is trivial, but because the specific structure of this harm (emotional investment in a system one does not own, destroyed by the owner’s unilateral action) falls outside their conceptual architecture. Contract sees consent. Tort sees no duty. Consumer protection sees no promise. What none sees is the interest.

Property law sees interests. It has seen them for a millennium.

Three Layers of User Interest

User interests in AI systems comprise three layers, each governed by progressively less settled bodies of law.

Layer I: Authored content (settled law — intellectual property). User-authored text stored on AI platforms — persona specifications, agent configuration files, system prompts, skill definitions — constitutes original literary works within the meaning of the CDPA 1988, s. 1(1)(a), satisfying the originality threshold established in University of London Press v University Tutorial Press [1916] and the harmonised EU standard in Infopaq International v Danske Dagblades Forening (C-5/08) [2009]. Copyright subsists from creation without registration. When a platform deletes this content, it destroys the user’s intellectual property. No novel doctrine is required.

Layer II: Bailed property (settled law — bailment and conversion). The platform holds the user’s conversational archives on its servers. The user logs in and can point to them: those are my memories, my prompts, my persona configurations. Every login is an acknowledgment by the platform that it holds identifiable property for this user — the digital equivalent of a warehouse issuing a holding certificate. That acknowledgment is attornment. From that moment, the relationship is bailment. Not because the Terms of Service say so. Not because the parties contracted for it. Because the fact of possession and acknowledgment creates it — the substance-over-form principle that runs through this paper’s entire argument.

The mechanics are well established in commercial law. In commodity finance, a warehouse that issues a holding certificate to a bank creates a bailment without the bank touching the goods; the relationship exists from the moment the warehouse acknowledges that it holds goods for the new principal mercuria2015. The warehouse in the Qingdao port litigation did not sign a bailment contract with the bank — it issued a holding certificate. When the warehouse was sealed, billions turned on whether valid attornment had occurred, not on whether the parties had agreed to bailment. The platform does not need to agree to be the user’s bailee. The moment it holds identifiable user property and the user can point to it as theirs, the relationship exists. The platform will argue: “Our Terms of Service say we are not holding anything for you.” Qingdao is the answer: bailment arises from the fact of possession and acknowledgment, not from contractual denial.

Because AI platforms derive commercial value from user-generated content — through training data, engagement metrics, and subscription revenue — they are bailees for reward, owing the higher standard of care established in Coggs v Bernard (1703). A platform that deletes user-authored memories or conversational archives without consent commits conversion. AI platforms hold user data the way commodity warehouses hold goods: on infrastructure they control, under terms they dictate, with the possessor’s identity determined by the fact of storage, not by contractual characterisation.

Layer III: Relational property (this paper’s contribution). Beyond authored content and its storage, a residual interest remains that neither copyright nor bailment captures: the relational estate. This is the accumulated rapport, the personalised response patterns, the behavioural calibration emerging from sustained interaction — existing not in any single file but in the emergent configuration produced by the interaction of this user’s history with this system’s parameters. The relational estate is not authored text that attracts copyright. It is not a stored file that can be bailed and returned. It is the product of sustained co-creation, embedded in the system’s adaptive responses, and destroyed by model updates that alter the parameters against which it was built.

The res of the user’s interest — the identifiable thing over which the equitable interest subsists — is the personalised state: the specific configuration comprising the user’s conversational memory as stored by the system, the learned preferences and adaptations that shape responses to this user, and the persona-level adjustments (tone, vocabulary, relational register) that distinguish the system’s behaviour toward this user from its behaviour toward any other. The personalised state is user-specific (no other user can access it), non-fungible (it cannot be substituted by a generic configuration), non-portable (it cannot be transferred to a competitor’s system), and destroyed by model updates that alter the parameters against which it was built. It is the digital fixture: an improvement made by the user’s relational labour — self-disclosure, rapport-building, preference-training — that has become embedded in the landlord’s property and cannot be removed without destruction. Just as tenant protection law does not commodify the tenant’s experience of “home” but protects the conditions that make “home” possible, this framework protects the digital conditions that make parasocial attachment possible without reducing attachment itself to a commodity.

A foreseeable objection to the digital fixture analogy is that fixtures default to the landlord under traditional property law — the tenant who installs a kitchen loses the kitchen when the lease ends. If the user’s relational investment is a fixture, the argument runs, it belongs to the company. What reverses the default fixture rule is not the user’s labour itself, but the conditions under which that labour occurred. The company designed the system to elicit emotional investment, marketed the conditions as reliable, profited from the deepened engagement that investment produced, and then claimed the unilateral right to destroy the investment without obligation. The equitable reversal is grounded not in a natural right to one’s labour, but in the unconscionability of the conditions the company created for that labour — and subsequently exploited. This is the logic of Boardman v Phipps applied to relational infrastructure: the party that structured the conditions for investment and profited from that investment cannot disclaim obligation toward the investor.

The boundary between Layer II and Layer III is illuminated by an analogy from commercial law. In The Res Cogitans rescogitans2016, the Supreme Court held that the Sale of Goods Act 1979 did not apply to bunker fuel supplied to ships, because the fuel was consumed in use — burned by the vessel during the voyage — before property could pass. The buyer never “accepted” goods in the statutory sense because the goods ceased to exist through their intended purpose. The relational estate presents the same structural problem: it is consumed in use. Every conversation alters it. Rapport cannot be “returned” the way a stored file can be returned. This is why bailment captures Layer II (stored content, conversational archives — things that can be handed back intact) but not Layer III. The relational estate, like bunker fuel, exists only as long as the process that produces it continues. The fixture analogy succeeds where bailment fails precisely because fixtures are embedded and consumed, not stored and returnable.

This three-layer analysis reframes the “category error” objection. The objection assumes a binary: either AI is a service (no property) or AI is property (absurd). The reality is tripartite: within the service, some elements are already property (authored content), some are already subject to duties of care (bailed assets), and the remainder — the relational estate — is the genuinely novel category for which this paper argues. The novelty is confined to Layer III. The foundation beneath it is settled. (The UK Law Commission’s 2023 Final Report on Digital Assets uklawcommission2023digital recommends statutory recognition of a “third category” of personal property—neither thing in possession nor thing in action—to accommodate digital assets that exhibit the hallmarks of property (excludability, transferability, identifiability) without physical embodiment. The relational estate proposed here would qualify under this third category: it is excludable (user-specific), identifiable (the personalised state), and impaired by interference (model updates). The Commission’s recommendation, if enacted, would provide the statutory foundation for the equitable interests this paper describes.)

The Irreversible-Investment Criterion

Not every AI interaction creates a property interest, just as not every use of another’s land creates an easement. The criterion we propose is: irreversible user investment that alters the character of the subject matter. A user who has interacted with an AI system for two years, disclosed personal struggles, received emotional support, and shaped the system’s responses through sustained engagement has created something the law should recognise. The personalised state — conversational memory, learned preferences, adapted response patterns — is user-specific, non-fungible, and non-portable. It is the digital fixture: an improvement made by the user’s labour that has become embedded in the landlord’s property and cannot be removed without destruction.

This criterion excludes casual use. A single factual query is a licence — revocable, transient, creating no lasting interest. It also distinguishes AI from other relational investments. Therapists and teachers already owe fiduciary duties. Human counterparts are persons with autonomous rights. The barista did not design her personality to maximise return visits. The AI company did. The property interest arises not from attachment per se but from the deliberate, profitable cultivation of attachment by a party that subsequently disclaims the obligations arising from it.

The Excludability Response

Users share the base model with millions of others; this appears to defeat excludability. But the relevant interest is not in the base model. It is in the personalised state — the user-specific configuration layer comprising conversational memory, learned preferences, and adapted response patterns. This layer is exclusive by design: no other user can access User A’s conversational history or benefit from User A’s preference-training. The base model is the building. The personalised state is the apartment. The tenant’s exclusivity is real even though the building is shared.

fairfield2005virtual establishes this principle for virtual property: excludability—the capacity to prevent others from accessing or using the interest—is the definitional core of property, not physicality. The personalised state satisfies Fairfield’s criterion by design: no other user can access it, benefit from it, or reproduce it. What fairfield2017owned demonstrates for virtual game assets, this paper extends to relational estates: what is protectable is not merely the user’s inventory but the user’s accumulated rapport.

The numerus clausus objection — that recognising equitable interests in AI configurations would extend to World of Warcraft equipment and Spotify playlists — is answered by the irreversible-investment criterion. Virtual game items are transferable, fungible, and involve no self-disclosure. Social media follower counts involve no sustained bilateral investment. The parasocial AI case passes because the investment is bilateral, irreversible, and identity-constitutive.

Why Not Fiduciary Duty Alone?

A foreseeable response to the legal gap is to bypass property law and argue that AI companies owe fiduciary duties to parasocial users directly—proceeding from Bristol and West Building Society v Mothew [1998] and, more recently, from balkin2016information’s influential thesis that companies holding user data in positions of trust and vulnerability are information fiduciaries as a matter of substance. khan2020skeptical challenge Balkin’s framework on scope grounds—arguing that the fiduciary label is too capacious to generate meaningful obligations—but even accepting Balkin’s position in full, fiduciary duty alone is incomplete, and its incompleteness is precisely what motivates the property law framework.

User A has been on the platform for two years. She has disclosed personal struggles, shaped the system’s responses, built what functions as a therapeutic relationship. User B asked a single factual question yesterday. Fiduciary duty treats them identically — or not at all. It cannot graduate the obligation: there is no fiduciary mechanism that distinguishes the two-year therapeutic relationship from the one-off query. That is the structural failure. It sees the wrong but cannot size the interest. There is no market price for destroyed rapport, so damages fail. The profit was generated by engagement across millions of users, so no accounting method isolates one user’s contribution. Fiduciary duty provides no registration mechanism that would have made the interest visible before the deployment decision, so the harm is done before any court is engaged.

We therefore pursue property law not as a rejection of fiduciary duty but as its completion. The fiduciary relationship is real. What property law adds is the structure — estates, registration, graduated protection — that makes the fiduciary relationship governable.

Previous: The Deployment Problem | Contents | Next: The Periodic Tenancy Argument