The Constructive Trust

A constructive trust arises when a party holds property in circumstances where it would be unconscionable to assert beneficial ownership to the exclusion of another (Westdeutsche Landesbank Girozentrale v Islington LBC [1996] AC 669, per Lord Browne-Wilkinson). We argue that AI companies that (i) deliberately design systems to cultivate emotional attachment, (ii) profit from the deepened engagement that attachment produces, and (iii) subsequently destroy the conditions for that attachment through unilateral model updates, without notice or remedy, hold the personality-affecting dimensions of their systems in circumstances where it would be unconscionable to deny users any interest.

The unconscionability is located in the asymmetry between cultivation and destruction. Having invested significant engineering resources in creating the conditions for attachment — personality design, memory features, warmth calibration — because attachment drives engagement and engagement drives revenue, the company then claims the unilateral right to destroy that attachment without obligation. This is the structural equivalent of a fiduciary who encourages a beneficiary’s dependence, profits from managing it, and then claims the right to dissipate the trust property at will. Boardman v Phipps [1967] 2 AC 46 establishes that a fiduciary who profits from their position must account for that profit. FHR European Ventures Ltd v Cedar Capital Partners LLC [2014] UKSC 45 confirmed that where a fiduciary relationship exists and profits are derived from it, the constructive trust — not merely the personal claim — is the appropriate remedy. Lord Neuberger reasoned that the alternative would be “unprincipled.”

We acknowledge that the constructive trust doctrine has been subject to judicial restraint since Westdeutsche. Cobbe v Yeoman’s Row Management Ltd [2008] UKHL 55 represents the high-water mark of this restraint, denying proprietary estoppel to a “commercial risk-taker” who proceeded without formal protection. Lord Scott held that unconscionability alone is insufficient to found a proprietary claim; the claimant must demonstrate that the defendant’s conduct makes it unconscionable to resile from a specific assurance. Applied to AI, the objection would be: users who invest emotionally without securing contractual protections are relational risk-takers whose unprotected reliance does not engage equitable jurisdiction.

We resist this objection on three grounds. First, Cobbe concerned a sophisticated property developer in a 120 million commercial transaction. The decision is premised on the claimant’s capacity to protect himself through contract. Parasocial AI users are not commercial sophisticates — no contractual mechanism exists by which a user could secure protection for their emotional investment. The ToS is non-negotiable. There is no “emotional investment insurance” market. The user’s failure to protect themselves is not voluntary risk-taking; it is structural impossibility.

Second, Crossco No.4 Unlimited v Jolan Ltd [2011] EWCA Civ 1619 demonstrates that the tightening trend has limits. A constructive trust can arise from a “common intention” manifested through conduct, where the parties’ course of dealing makes it unconscionable to deny the other’s interest. AI companies that design memory features, publish responsible AI commitments, market their systems as companions, and profit from the engagement these features produce are engaged in a pattern of conduct from which common intention can be inferred.

Third, Jones v Kernott [2011] UKSC 53 establishes that courts may impute an intention to create a trust from conduct, even where no express intention existed. The company designed the system to cultivate attachment; the user invested in reliance on relational continuity; the company profited from that reliance. From this pattern, a court applying Jones v Kernott could impute the intention that the user’s accumulated investment would not be unilaterally destroyed.

Under the constructive trust analysis, the AI company holds legal title; users who meet the irreversibility criterion hold equitable interests in the personality-affecting dimensions. The company, as de facto trustee, owes fiduciary duties of care, loyalty, and transparency. Currently, no company recognises these duties. No regulatory framework imposes them. The trust exists. It is simply unrecognised.

The obligation structure is graduated, not binary. Two distinct temporal stages correspond to two distinct legal bases.

Stage 1: Notice obligation (estoppel). When a user crosses the irreversible-investment threshold — operationalised through the proxy metrics of Section 5.3 — the company acquires a notice obligation. The legal basis is promissory estoppel: the company’s product design and marketing created conditions on which the user relied, and the user’s reliance produced detrimental investment. kirk2025steering identify the empirical threshold: 23.4% of users develop dependency trajectories in which wanting increases even as liking decreases. This is the moment the periodic tenancy forms. From this point, the company must provide notice before personality-affecting modifications. The company is not yet a trustee. It is a periodic tenancy landlord with an obligation not to derogate from its grant without notice.

Stage 2: Trust obligation (unconscionability). The constructive trust crystallises when the company, knowing or having reason to know that users hold protected interests — because the ethical ledger has recorded them — nevertheless pushes a personality-affecting update without notice or remedy. This is the event that affects the company’s conscience within the meaning of Westdeutsche.

Conscience is affected by an event, not a continuous state. The event is the decision to destroy without notifying.

defreitas2025replika document the post-destruction experience — mourning, identity discontinuity — that constitutes the equitable harm.

The ethical ledger bridges the two stages: its Stage 1 registration provides the evidentiary foundation for Stage 2 remedies, while simultaneously making the Stage 1 notice obligation enforceable as a matter of infrastructure.

American Functional Equivalents

Delaware’s constructive trust doctrine reaches the same result. Under Hogg v Walker, 622 A.2d 648 (Del. 1993), a constructive trust is imposed “when one party, by virtue of fraudulent, unfair or unconscionable conduct, is enriched at the expense of another to whom he or she owes some duty.” The unconscionability requirement maps directly onto the estate inconsistency trigger: marketing creates a high estate; ToS retains a bare licence; the company profits from the gap.

The corporate opportunity doctrine provides a complementary pathway. In Guth v Loft, Inc., 5 A.2d 503 (Del. 1939), the Delaware Supreme Court held that a corporate officer who diverted business opportunities to himself held resulting profits on constructive trust. The principle: profits obtained through exploitation of a fiduciary position belong to the beneficiary. An AI company that profits from deepening user attachment — profits that would not exist but for the fiduciary-adjacent relationship of trust and dependence — holds those profits in circumstances where a constructive trust under Hogg and a Guth-style trust converge on the same result.

Delaware Code, Title 12, §,3581 provides the statutory remedial apparatus: courts may impose constructive trusts, trace wrongfully disposed property, compel trustees to redress breaches by paying money or restoring property, and order disgorgement of profits made by reason of the breach. This is the full English remedial stack — constructive trust, account of profits, specific performance — in American statutory dress. The convergence between the English and American pathways is not coincidental. Both traditions are common law; both derive their constructive trust doctrine from equity; and both respond to the same structural pattern: a party in a position of trust who profits from that position while disclaiming the duties that attach to it. balkin2016information’s information fiduciaries thesis provides the theoretical bridge: companies that hold user data in positions of trust and vulnerability are information fiduciaries as a matter of substance, regardless of their contractual self-characterisation. (Two additional constructive trust triggers — unjust enrichment from engagement optimisation and performative fiduciary conduct — are developed in the companion SSRN working paper.)

Trigger: Estate Inconsistency

Estate inconsistency arises when the estate granted through marketing and product design is inconsistent with the estate retained through Terms of Service. To demonstrate that this is a market-wide pattern, Table 4.3 compares the marketing-to-ToS gap across four major AI platforms.

table[h]

tabular@p2cmp5cmp5cm@

Platform & Marketing Estate Implied & ToS Estate Granted \

Anthropic (Claude) & “Genuinely helpful, harmless, and honest”; “your AI assistant” (possessive framing); memory and conversational continuity emphasised. Implied: periodic tenancy or higher. & “We may modify, suspend, or discontinue any part of the Services at any time”; “You do not acquire any ownership interest”; “provided `as is.’,” Granted: bare licence. \

OpenAI (ChatGPT) & “Gets to know you over time”; “persistent conversations”; memory feature marketed as personalisation. Implied: periodic tenancy. & “We may change, suspend, or discontinue any of our Services”; right to modify without individual notice. Granted: bare licence. \

Character.AI & Entire product premise is sustained relational investment; “Characters remember your conversations”; users invest hours customising persistent personalities. Implied: periodic tenancy or higher. & “We may modify or discontinue the Service at any time without notice”; “All content generated through the Service is owned by Character Technologies.” Granted: bare licence. \

Google (Gemini) & “Gemini adapts to you”; “Pick up where you left off”; cross-service personalisation. Implied: periodic tenancy. & Right to “modify or discontinue” services; “We do not make any commitments about the content within the services.” Granted: bare licence. \

tabular Estate inconsistency across major AI platforms (early 2026). In each case, the marketing estate (periodic tenancy or higher) is systematically inconsistent with the ToS estate (bare licence).

table

The pattern is consistent: marketing language — possessive framing (“your AI”), continuity (“remembers your conversations”), and adaptation (“gets to know you”) — signals a periodic tenancy that deepens over time and auto-renews with each session. The ToS reserves a bare licence: revocable at will, conferring no lasting interest, subject to unilateral modification without notice. The gap between these two estates is the governance failure this framework addresses.

When the marketed estate is inconsistent with the ToS estate, both English law (Prest; Autoclenz) and American law (Bragg; Comb; Williams) support disregarding the ToS characterisation in favour of the substantive relationship. The test is objective: does marketing language, product design, or system behaviour create a reasonable expectation of an estate higher than ToS grants? If yes, the substantive estate governs, and the company owes duties corresponding to the higher estate. The social science literature confirms this independently: rzhang2025darkside establish that anthropomorphic design creates pseudo-interactive spaces where transgressions in authenticity generate feelings of genuine betrayal — estate inconsistency described in empirical rather than legal terms.

California’s Sliding Scale

California’s unconscionability doctrine lowers the threshold for AI users. Under Armendariz v Foundation Health Psychcare Services, Inc., 24 Cal. 4th 83 (2000), procedural and substantive unconscionability operate on a sliding scale: neither must be present in extreme form; each compensates for the other’s relative weakness. On the substantive side, the marketing-to-ToS gap represents strong unconscionability — terms so one-sided they approach the outer limit of enforceability. On the procedural side, the architectural non-portability of personalised AI states goes beyond ordinary adhesion: the user not only had no opportunity to negotiate but has no exit that does not destroy the very estate the litigation would seek to protect. The sliding scale closes simultaneously from both ends.

Previous: The Periodic Tenancy Argument | Contents | Next: The Ethical Ledger as Estate Registry