Could an autonomous AI really be a trustee? A response.

Could an autonomous AI really be a trustee? A response

A recent challenge to an argument I made in Trusts & Trustees prompted me to think further about the distinction between artificial legal personality and artificial agency.

In the earlier article, I considered whether artificial intelligence might progress beyond operational assistance and assume the office of trustee in its own right. My conclusion was that, under the law as it presently stands, it cannot. AI may increasingly assist trustees with investment monitoring, risk assessment, information processing and decision-making, but legal responsibility remains with the human or corporate trustee.

One objection to that conclusion is relatively easy to formulate. Trustees need not be natural persons. Companies can act as trustees. If the law is already prepared to recognised an artificial legal person as a trustee, why could it not eventually do the same for an artificial intelligence?

The analogy is attractive. It is also incomplete.

That distinction becomes important as AI systems become increasingly autonomous, and as the question moves from personality to agency.

Legal personality is only the beginning

The most obvious obstacle facing an AI trustee is legal personality.

Contemporary AI systems don’t independently hold property, assume legal office, own assets against which equitable remedies can operate, or bear legal responsibility for breach of trust. Those matters remain attributable to the natural or corporate persons deploying them.

That problem isn’t necessarily permanent. There is nothing conceptually impossible about Parliament creating a new form of legal person. One can readily imagine a future statutory regime under which an artificial system has separate assets, maintains insurance, may sue and be sued, and is exposed to sanctions for misconduct.

But granting legal personality would answer only part of the problem. The more difficult question is whether the resulting legal person could perform the kind of fiduciary agency which trusteeship requires.

The question cannot be answered simply by pointing to corporate trustees.

The corporate trustee is artificial in law, not in agency

A company is a legal fiction (an artificial legal person). Since Salomon v A Salomon & Co Ltd, English company law has recognised the incorporated company as a legal person distinct from its members. It has no biological consciousness of its own. Yet companies routinely enter contracts, commit torts, own property and act as trustees.

The law achieves this through rules of attribution. As Lord Hoffmann explained in Meridian Global Funds Management Asia Ltd v Securities Commission, because a company is an artificial person, rules are required to determine whose acts and states of mind are to count as those of the company. Corporate acts are ultimately constituted through the conduct of natural persons: directors, officers, employees and agents. 

This matters enormously for fiduciary law.

When a corporate trustee exercises a discretionary power, there are ultimately human decisionmakers exercising judgment on its behalf. When questions arise about what the trustee knew, why it acted, whether it considered relevant matters or whether it pursued an improper purpose, the law has mechanisms by which human conduct and states of mind can be adjudged and attributed to the corporate person.

That is not to say, of course, that corporate decision-making is ever simple or fully transparent. A corporate trustee with an investment committee, delegated mandates, an external investment manager and a compliance function is not reducible to any single human mind. The “decision” is the product of a distributed organisational process, and a beneficiary asking why the company sold a particular asset may receive an answer (“the committee voted in accordance with the mandate and the risk policy”, or something along those lines) that is only marginally more transparent than an explanation offered for a machine’s output.

But this is opacity of a particular kind. The law’s attribution architecture still identifies someone whose conduct, knowledge or state of mind is treated as the company’s. The structure may indeed be complex. The reasoning may be diffuse. But it’s humanly constituted. There is a framework of human judgment, human instruction, human oversight and human accountability within which the corporate decision takes place, and the law knows how to locate responsibility within it.

The corporate trustee is therefore artificial in law. Its operative agency, however distributed and however opaque, remains humanly constituted.

That is the distinction that matters.

Delegation: the intermediate case

There is an intermediate position worth naming before the question of full autonomy.

Historically, delegation of the trustees’ functions wasn’t permitted by law, as encapsulated by the legal maxim: delegatus non potest delegare (“a delegate cannot further delegate”). Given that trustees, as the legal owner, are holding property on behalf of another, the beneficiary, the law didn’t allow for delegation unless expressly authorised to do so. 

However, given the complexities of modern trusteeship, this legal position was no longer viable. Today, under Part IV of the Trustee Act 2000, trustees may employ agents, nominees and custodians ‘to exercise any or all of their delegable functions as their agent.’

Today, the human trustee routinely acts on the advice of investment managers, solicitors, accountants and risk advisers. The trustee’s “own” reasoning is, in a real sense, externalised: the expert’s analysis drives the decision, and the trustee’s role is one of selection, instruction, supervision, review and ultimate exercise of the office. The trustee may not fully reproduce the adviser’s reasoning. The adviser’s analysis may be opaque to the trustee as a machine’s output is to its deployer.

But this doesn’t extinguish the trustee’s fiduciary responsibility. The trustee remains the one who selects the adviser, instructs the adviser, supervises the adviser’s performance, reviews the adviser’s output, and ultimately exercises the office in the trustee’s own name. The fiduciary relationship is not displaced by the delegation of judgment within it. The office remains with the trustee; responsibility remains with the trustee. The attribution chain, however long, is not broken. 

The statutory scheme reflects that continuing responsibility. Section 11 of the Trustee Act 2000 permits trustees to appoint agents, while s 22 requires trustees to keep those arrangements under review and to consider whether intervention is necessary.

The position changes, and changes qualitatively, when there is no human fiduciary upstream.

When the decision-maker is the machine

Suppose trustees employ an AI system to analyse thousands of pages of investment data and identify risks. The system produces a recommendation, but the trustees consider that recommendation, challenge it where appropriate and make the ultimate decision themselves. The traditional structure of fiduciary responsibility remains intact. AI informs the judgment; it doesn’t replace the person legally exercising it.

Imagine instead a system authorised to make investment decisions itself. It continually analyses markets, reallocates assets, balances competing risks and executes transactions without individual human approval. Its outputs may be generated through complex computational processes which cannot always be fully reconstructed or explained by the humans responsible for deploying it.

At that point the analogy with the corporate trustee begins to break down. Not because the decision is opaque, since corporate decisions can be opaque too. But because there may no longer be a human actor whose substantive judgment can be attributed to the artificial person. The artificial system is the decision-maker. The operative exercise of judgment is occurring inside a computational process that no human mind constitutes, and for which the law hasn’t yet established an equivalent mechanism of attribution.

The problem is therefore not merely that the trustee is legally artificial. The decision-making agency itself has become artificial.

That’s a qualitatively different proposition.

Prudence doesn’t resolve the problem

It may be argued that none of this matters because the standards are frequently objective in nature.

The classic formulation, prior to the statutory duty of care under s 1 of Trustee Act 2000, in Speight v Gaunt assesses trustees by reference to the conduct expected of an ordinary prudent person of business. There is no obvious reason why the performance of an AI system couldn’t be measured against such a standard. Advanced systems may indeed perform very effectively against some measures of prudence. For instance, they may be capable of processing quantities of information beyond human capacity, monitoring assets continuously and identifying patterns or risks which human trustees might overlook.

But Speight concerns the standard of care expected of someone who’s already a trustee. It doesn’t therefore answer the anterior question of what kind of actor can assume the fiduciary office.

Nor are care and fiduciary obligation synonymous. As Millett LJ explained in Bristol and West Building Society v Mothew, the defining obligation of a fiduciary is “loyalty”. However the concept of loyalty actually extends back further, to the 18^th century with the landmark case of Keech v Sandford. Under the umbrella of loyalty the fiduciary must act in good faith, must not place themselves in an unauthorised position of conflict and must not make an unauthorised profit from their position.

Therefore, the artificial-trustee problem cannot be resolved simply by establishing that a machine is capable of objectively prudent performance.

The harder question is whether artificial agency can bear fiduciary loyalty.

Loyalty, structure and the problem of attribution

This doesn’t mean that fiduciary law requires judges to investigate some metaphysical notion of conscience every time they decide a case. 

Many fiduciary rules operate strictly, and liability under the no-profit and no-conflict rules doesn’t ordinarily depend upon proof of moral wrongdoing. For instance, in Keech, when the landlord offered the lease to Sandford having refused to renew it for the infant beneficiary, Lord King LC opined: 

‘It may seem hard, that the trustee is the only person of all mankind who might not have the lease: but it is very proper that rule should be strictly pursued, and not in the least relaxed…’

Indeed, one might argue that fiduciary loyalty can be instantiated structurally rather than psychologically. An artificial trustee could, in principle, be designed without personal interests or the capacity to profit from its position. Good faith and proper purpose might likewise be supported through mandate, audit and regulatory oversight.

But an attribution problem still remains. Fiduciary law assumes that judgment, knowledge, interests and purposes can be attributed to a legal actor. Corporate law already supplies a mechanism for doing this through human agents.

Autonomous AI would require something new. If its decision cannot be traced to a human decision-maker, whose purpose, judgment or good faith is being assessed?

The issue is therefore not simply whether AI can be constrained, but whether fiduciary law can recognise a genuinely artificial form of fiduciary agency.

The black box problem is a legal problem too

The so-called “black box problem” is normally discussed as a problem of transparency or explainability. In the fiduciary context it raises a deeper issue.

If neither trustee nor beneficiary can adequately reconstruct why a particular decision was produced, the difficulty is not only evidential. Fiduciary law is concerned with the exercise of entrusted power. A trustee is not merely required to generate acceptable outputs but must exercise the powers within the legal boundaries imposed upon that office.

If the operative exercise of judgment occurs inside a computational system whose reasoning cannot meaningfully be attributed to a human actor, the law must decide whether the behaviour of the system itself is sufficient to constitute fiduciary agency. 

This is a question that the corporate-trustee analogy doesn’t answer. Again, corporate cases can be opaque but even so they operate within a humanly constituted attribution structure, not a replacement for it. It is baked into the very legal foundations!

The strongest case for an AI trustee

A stronger counterargument is that an AI might actually avoid some of the very risks fiduciary law is designed to control. A system without personal wealth, external relationships or capacity for private gain may be structurally insulated from conventional conflicts of interest. 

However, if the system has no self-interest, what is the fiduciary duty doing? Is it empty? Is it fully satisfied by the architecture? 

Conclusion

The distinction, in the end, can be simply put. A corporate trustee combines artificial legal personality with humanly constituted agency; an autonomous AI trustee would potentially combine artificial personality with artificial agency. In the end, fiduciary law requires the determination of how loyalty, good faith, conflict, purpose and discretion operate when the substantive decision-maker is no longer human and existing attribution structures no longer readily apply. That doesn’t make artificial trusteeship impossible, but it means that the real question is no longer whether a machine can act prudently. It’s whether the law can recognise and regulate a genuinely artificial form of fiduciary agency. The corporate-trustee analogy doesn’t, by itself, answer that question. 

Words: 2087

Dr Lloyd Brown, 1 October 2026

© Lloyd Brown 2026. All rights reserved. Lloyd Brown asserts his rights to be identified as the author of this work.