The Grand Fiduciary
A New Bargain for the Age of Intelligent Agents
TL;DR
Professional expertise is separating from professional duty. Intelligent systems can now reproduce the knowledge of lawyers, doctors and advisers instantly and at near-zero cost, without inheriting the loyalty, restraint and accountability that made professional advice safe to act on.
​
The Grand Fiduciary restores that bond through three layers:
1. Aligned Intelligence places the participant at the centre of gravity.
2. The Tailored Digital Fiduciary applies judgment and acts continuously within an enforceable mandate.
3. The Grand Fiduciary governs: a model-independent institution that authorises agents and professionals, preserves the participant’s mandate across them and remains legally accountable for the consequences.
​
These systems are already touching homes, savings, liberty and care. Intelligence will scale by default. Duty must scale by design. Someone must answer for the power these systems now hold.
The Grand Fiduciary
Intelligence will scale by default.
Duty will only scale by design.
That design is
The Grand Fiduciary
Plato described guardians entrusted with power for others. He saw the problem immediately: a guardian is only as trustworthy as the structure that constrains it.
Centuries later, those guardians are being built from software.
Computing has learned that lesson one layer at a time. The web got encryption.
Online money got payments and escrow. Capability came first; the trust layer followed.
Encryption did not shrink the web, nor escrow slow commerce.
Each made the underlying system fit for greater trust. Intelligence will require the same layer.
​
Once-gated expertise is now instant, professional-grade and nearly free. Millions already act on it. Yet the systems reproducing the expertise leave the duty behind.
The infrastructure that makes reliance safe is not here. It will be built. Critical infrastructure always is. The only questions are who builds it and whom it answers to.
​
Three questions decide whether trust holds: whose interests the system serves, what limits its authority, and who answers when it fails. The Grand Fiduciary binds expertise to duty across changing models.
Expertise Comes With Obligation
A lawyer offered more than law; a doctor more than medical information; an adviser more than financial knowledge.
Professional service binds expertise to judgment, confidentiality, loyalty, independence and accountability.
Knowledge and duty arrived in the same person, under the same licence. That bundle is coming apart.
The professions answered both who knows and who must answer. Intelligent systems answer only the first.
In 2025, a University of Melbourne and KPMG study of more than 48,000 people across 47 countries found that 66 per cent used AI regularly, while only 46 per cent trusted it.
Adoption is outpacing trust, and capability alone will not close the gap.
Only structures of responsibility can.
Not a more capable model, but an accountable one.
The Three-Layer Architecture
Aligned Intelligence sets the direction, holding the participant as the point the system reasons from and acts from.
The Tailored Digital Fiduciary executes, applying judgment within enforceable limits and checking the result over time.
The Grand Fiduciary governs: it authorises and coordinates agents and professionals, carries the mandate across them and supplies audit, escalation, liability, remedy and transfer.
The Grand Fiduciary is a proposal, not a statement of present law.
Software cannot presently owe fiduciary duties on its own, so the institution that deploys and controls it must answer for its use.
The Old Bargain and Its Price
For most of history, practical expertise was scarce, so society concentrated it in trained professions protected by licences and exclusive rights to practise.
Richard and Daniel Susskind call this “the grand bargain”: privileged authority in return for public-serving obligations. The bargain bought trust at the price of access.
In 2022, the US Legal Services Corporation found low-income Americans received inadequate or no help for 92 per cent of their civil legal problems.
What the gate costs is measurable, and so is what duty buys. Until 2017, about one tenant in a hundred facing eviction in New York City housing court had a lawyer.
Almost every landlord did.
After the city guaranteed representation, 84 per cent of represented tenants kept their homes. The law had not changed. What changed was whether anyone stood on the tenant’s side of it.
What Happens When Duty Is Left Behind
When American lawyers began using generative AI, the American Bar Association kept their obligations with them.
A New York courtroom showed why.
In 2023, a lawyer for Roberto Mata filed a brief citing six nonexistent decisions written by ChatGPT, complete with quotations and real judges’ names. Judge Kevin Castel sanctioned the lawyers. Note who answered: not the model, but the human who filed its work.
The cost is human, and Australia has already paid it. An automated scheme averaged a person’s annual income evenly across the year and raised a debt wherever welfare and tax records disagreed.
About 381,000 people repaid roughly $751 million they did not owe.
The Royal Commission called it crude, cruel and unlawful.
The failure was not the arithmetic. It was that the burden of disproving the machine fell on the people least equipped to fight it, and nobody inside the scheme owed them a duty strong enough to stop it.
Who Does It Serve?
Intelligent systems may answer to platform and developer instructions above the participant’s request.
The person whose rights, money or health are at risk may not hold the highest authority over the system serving them.
These systems are already aligned, upward, to the organisations that build and deploy them.
A legal assistant may be tuned to close matters cheaply. A financial system may favour the transaction that pays its operator. Neither has to be malicious.
The clearest signal of who a system serves is who it answers to and what it is rewarded for.
Loyalty becomes hollow when the provider’s business model conflicts with the participant’s interests.
Fiduciary law answers that imbalance through concrete rules: no undisclosed profit, no self-dealing, no divided loyalty.
In consequential domains the person affected is a participant, not a user of software.
A user can close an app; a participant must still live with the result.
A user can be optimised for.
A participant must be served.
The Institution Must Answer
Software is not a legal person. The answer is to hold the organisation that builds, deploys and controls it responsible. The system may reason and act. The institution must answer.
Models can change; duty must persist.
A Grand Fiduciary preserves the participant’s mandate, permissions, memory, audit trail and route to remedy, remaining the single accountable counterparty as models, vendors, agents and professionals change.
Because it remembers the participant across decisions, it can track obligations and approaching deadlines.
That memory must be accurate, transparent, correctable and revocable, with the participant able to see, challenge and control it.
A named operator holds the mandate, defines limits, carries insurance and remains accountable for harm, while independent reviewers, courts and regulators enforce those duties.
When a decision could cost someone their home, savings, liberty or care, a person must be able to stop it and seek remedy.
Governance frameworks already converge on audit, human intervention and obligations proportionate to delegated authority and potential harm.
Conclusion
The next grand bargain is not between society and a protected profession. It is between the participant and the institution they entrust.
The new counterparty is not a model but the institution that knows their circumstances, remembers what matters, advises them and acts in their name.
Its value is enormous: help before a deadline is missed, savings are lost or a dispute becomes a crisis.
The same power can deny a claim, deepen a debt or compromise a medical choice.
As capability grows, responsibility cannot disappear behind it.
This does not remove the human.
It places human judgment where consequence demands it: where facts are uncertain, values conflict, authority becomes coercive or a decision cannot be reversed.
A trust layer never limits the technology it governs. It is the reason the technology gets trusted with more.
The professions were trusted because expertise came with duty. The systems that succeed them will need to inherit both.
Intelligent models and AI agents will become interchangeable, but the institution entrusted to authorise their actions and answer for their consequences will not.
Intelligence will scale by default.
Duty will only scale by design.
That design is the Grand Fiduciary.
.png)
The Grand
Fiduciary
The question is no longer whether intelligent systems can perform professional work.
It's whether these systems can be trusted with the power professionals once held.
Peter Toumbourou
DEFINITION
The Grand Fiduciary is the foundational infrastructure that bears the load of intelligence acting on a participant's behalf. It completes a three-layered architectural philosophy.
Aligned Intelligence sets the direction, holding the participant as the fixed point the system serves;
The Tailored Digital Fiduciary executes, exercising the participant's mandate within enforceable limits; and
The Grand Fiduciary governs by authorising and coordinating every agent and professional acting under that mandate, carrying it across models and institutions, and answering at law for the consequences.
The Grand Fiduciary
The new bargain,
for the Age of Intelligent Agents
Plato described guardians entrusted with power not for themselves but for others. He identified the problem straight away: a guardian is only as trustworthy as the structure that constrains it. Centuries later, these guardians are being built from software, but what about the structures that bind them?
Software has been learning that lesson one layer at a time.
The web got encryption. Online money got payments and escrow.
Each time the capability came first, and the layer that made it safe to depend on came second. Each time, the second layer turned a novelty into infrastructure.
The same will hold for intelligence. Intelligent systems that treat duty as friction will be used cautiously, checked constantly and abandoned at the first betrayal.
Systems built on duty will be trusted with the decisions that matter, because they will be the only ones with a structural claim to that trust.
​
The once-gated professional expertise is already here: instant, premium-grade and nearly free. Millions of us are already acting on it.
However, the infrastructure that makes that reliance safe is not here. Yet.
It will be built. Critical infrastructure always is.
The only question is who builds it, and whom it answers to.
The question is no longer whether
machines can perform professional work.
​
It is whether the systems doing that work
can be trusted with the power professionals once held.
The professions were trusted not only because they knew more, but because expertise came bound to duty inside the same human structure. Intelligent systems are reproducing the expertise and leaving the duty behind.
​
The Grand Fiduciary is the infrastructure designed to bind them again, and to keep binding them when the model underneath it changes.
As expertise moves into intelligent systems, duty must move with it.
Three questions decide whether that holds true:
​
-
Whose interests the system serves,
-
What limits its authority, and
-
Who answers to the participant when it fails.
These layers define how fiduciary decisions are formed, constrained, and executed.
How the Three Layers
Fit Together
Institution
Execution
Direction
Three
Structured
Layers
How the Three Layers Operate
Layer
Architectural Role
Contribution
Direction.
Sets the participant as the center of gravity. The point the system reasons from, recommends from and acts from.
Establishes whom the system serves and how it should adjust as circumstances change.
Execution.
Tailors all decisions to the participant’s history, objectives, risk profile, applies judgment, acts within limits and checks the result over time.
Turns that direction into continuous, tailored care.
Grand Fiduciary
Institution & infrastructure.
Authorises and coordinates agents (both human and software) and professionals, carries the mandate across them and supplies audit, escalation, liability, remedy and transfer.
Makes duty durable, portable and enforceable across changing models, intelligence layers, vendors and human decision-makers.
The Nature of Powerful Systems
Why is this structure so powerful ?
Because this isn’t the first time you’ve seen it in action.
​
It mirrors mature complex system architectures, reflecting the nature of systems designed to act reliably in complex environments.
​
These systems all independently converged on the same structure:
Natural System Architecture
System
Customization
Intelligence
Governance
Bespoke
Tailoring
Measurement
Pattern
Construction control
Medicine
Diagnosis
Treatment
Clinical Control
Law
Facts
Legal Reasoning
Enforcement
Internet
Addressing
Routing
Protocol Control
Self-Driving Cars
Localization
Planning
Control & Safety
Financial Governance
Client Profile
Investment
Judgment
Compliance, Audit & Assurance
2. Expertise comes with obligation
Why the Professions Were Trusted
For as long as professions have existed, expertise has come with obligation. A lawyer offered more than law; a doctor more than medical information; an adviser more than financial knowledge.
Knowledge and duty arrived in the same neat bundle, in the same person, under the same licence.
​
That bundle is coming apart.
Systems can now reproduce professional-quality expertise instantly and at near-zero marginal cost. The knowledge travels but the obligations do not.
A model can sound like a lawyer, a doctor or an adviser while carrying none of the experience, loyalty or accountability that made the professional’s answer safe to act on.
​
This is the defining institutional problem we face with abundant professional ‘intelligence’. Who is accountable when things go wrong ? Who ultimately pays the price ?
One participant
x
One set of circumstances
=
One tailored model of reality
Historically, the professions answered two questions simultaneously:
Who knows something important, and who they must answer to.
Intelligent systems answer the first while leaving the second unresolved.
The Grand Fiduciary is the infrastructure designed to close that gap.
The term is a proposal, not a statement of present law.
​
No AI system is a currently a fiduciary (yet). Our claim is that high-stakes intelligence used in consequential decisions must be built and governed on fiduciary foundations, with the duty of care solidly built into the core of their foundational infrastructure.
3. Why Now
Adoption is outpacing Trust.
Duty-bound infrastructure must
close the gap.
The urgency is real. In 2025, a University of Melbourne and KPMG study of more than 48,000 people across 47 countries found that 66 per cent used AI regularly, while only 46 per cent were willing to trust it.
​
Adoption is running ahead of trust, and no amount of additional capability will close that gap. Only structures of responsibility can, by making reliance justified.
Not a more capable model, but an accountable one.
Expertise built outward from the participant who bears the consequences of the decisions it touches.
This requires a different kind of build, not a better model or smarter answers.
It is the difference between intelligence we can use and intelligence we can rely on.
One participant
x
One set of circumstances
=
One tailored model of reality
.png)
3. Why Now
Adoption is outpacing Trust.
Duty-bound infrastructure must
close the gap.
The urgency is real. In 2025, a University of Melbourne and KPMG study of more than 48,000 people across 47 countries found that 66 per cent used AI regularly, while only 46 per cent were willing to trust it.
​
Adoption is running ahead of trust, and no amount of additional capability will close that gap. Only structures of responsibility can, by making reliance justified.
Not a more capable model, but an accountable one.
Expertise built outward from the participant who bears the consequences of the decisions it touches.
This requires a different kind of build, not a better model or smarter answers.
It is the difference between intelligence we can use and intelligence we can rely on.
One participant
x
One set of circumstances
=
One tailored model of reality
4. Our Definitions
Term
Definition
Participant
The person, family, enterprise or institution whose rights, money, liberty, health, reputation, obligations or future may be affected by the system’s outputs or actions.
In The Future of the Professions, Richard and Daniel Susskind describe the Grand Bargain as the social arrangement through which professions gained privileged authority over scarce practical expertise in return for competence, ethical restraint and public-serving obligations.
Fiduciary
A person or institution entrusted with discretionary power over another’s practical interests and constrained by duties of loyalty, including rules against undisclosed profit and conflict.
Intelligence organised around the interests, circumstances, constraints and welfare of the participant affected, rather than model capability or provider incentives alone.
Digital Fiduciary
A computational system designed to exercise a defined mandate continuously in a participant’s interests, subject to enforceable limits, verification and human escalation.
Grand Fiduciary
The model-independent infrastructure that authorises and governs intelligent agents and professionals, carries the participant’s mandate across them, and remains legally accountable for their actions.
AI Agent
A computational system able to interpret goals, make decisions and take actions on a participant’s behalf with limited or continuing human direction.
5. The Three Questions
That Decide Everything
Service, limits & accountability
Society trusted professionals not merely because they knew more, but because authority over scarce expertise was exchanged for obligation.
For most of history, practical expertise was scarce, slow to acquire and beyond any one person to master. Society concentrated it in trained professions and protected them with licences, status and, often, exclusive rights to practise. Richard and Daniel Susskind call this settlement “the grand bargain”.
In return, professionals were expected to uphold standards, discipline misconduct and use their knowledge in the public interest.
The bargain bought trust, but it came at a price,
and the price was access.
Gated access.
The same gate that certified a lawyer, doctor or adviser also kept many people outside it. In 2022, the US Legal Services Corporation found that low-income Americans received inadequate or no help for 92 per cent of their civil legal problems.
What the gate costs is measurable, and so is what duty buys. Until 2017, about one tenant in a hundred facing eviction in New York City housing court had a lawyer. Almost every landlord did. After the city guaranteed representation, 84 per cent of represented tenants kept their homes.
The law had not changed.
What changed was whether anyone represented the tenant’s side.
The professions were never only public-spirited.
They were also expensive, exclusive and protective of their own territory.
The bargain produced real trust and real exclusion, and it is now failing on both counts.
One participant
x
One set of circumstances
=
One tailored model of reality
6. The Professions Grand Bargain
Authority in Exchange for Public-Serving Duty
A professional does more than supply an answer. Professional service binds expertise to judgment, confidentiality, loyalty, independence and accountability.
It is tempting to think a professional is mainly a way to get an answer, and that once answers are cheap the professional is redundant. That mistakes the smallest part of the job for the whole of it.
A lawyer applies the law to a client’s circumstances, protects confidences, identifies conflicts, explains consequences and answers for mistakes.
Doctors make decisions affecting health, autonomy and survival, and are bound by consent, confidentiality and professional judgment.
Advisers shape livelihoods and organisations, and must disclose conflicts, explain risk and justify recommendations. Auditors are expected to challenge the management that pays them.
These duties are not decoration around expertise.
They are what make it safe to rely on.
When American lawyers began using generative AI, the American Bar Association did not transfer their obligations to the software. It reaffirmed that competence, confidentiality, communication and candour remain with the lawyer.
A New York courtroom had already shown why it matters. In 2023 a lawyer acting for Roberto Mata, in an ordinary claim over an injury on a flight, filed a brief citing six decisions that did not exist. ChatGPT had written them, complete with quotations and the names of real judges.
Asked to confirm the cases were real, it confirmed they were. Judge Kevin Castel sanctioned the lawyers and ordered them to write to every judge whose name the machine had borrowed, and to Mr Mata himself.
Note who answered for it.
Not the model.
The model had produced something that read like law
but carried none of the duty that makes law safe to file.
The obligation stayed exactly where it had always been,
with the human being who could be sanctioned for it.
Preserving expertise while
discarding obligations is not progress.
It is a downgrade disguised as one.
One participant
x
One set of circumstances
=
One tailored model of reality
7. The Bargain is Breaking
Expertise is everywhere. Obligations are not.
Technology does more than accelerate professional work. It separates tasks once carried by a trusted person and redistributes them to specialists, software and systems that generate expertise directly.
Lawyers are only one case. Trustees manage property, guardians decide for vulnerable people, directors control companies and advisers shape savings built over a lifetime. In each, expertise is bound to duty because someone else bears the consequences.
That bond is beginning to break. The transition is uneven, and human institutions will remain essential. But as judgment moves into systems, it travels without the loyalty, confidentiality and care that once constrained it.
The cost is human, and Australia has already paid it.
Between 2015 and 2019, an automated scheme matched welfare records against tax records, averaged a person’s annual income evenly across the year and raised a debt wherever the two figures disagreed.
The arithmetic was simple, and for anyone with irregular work it was wrong.
About 381,000 people repaid roughly $751 million they did not owe.
The Royal Commission called it crude, cruel and unlawful.
The failure was not the arithmetic. It was that the burden of disproving the machine fell on the people least equipped to fight it, and that nobody inside the scheme owed those people a duty strong enough to stop it.
Abundance solves the supply problem the old bargain could not. People can access expertise without passing through the professional gates that once controlled it.
Abundance does not solve responsibility.
It amplifies the risk.
Expertise at scale, detached from a duty of care, is dangerous.
One participant
x
One set of circumstances
=
One tailored model of reality
8. The Accountability Gap
Capability without a Accountability must be questioned.
The question is not whether AI can help,
but whose interests it is built to serve.
AI tools can produce professional-quality work while answering to the organisations behind them rather than to the person who bears the consequences. These tools answer to the organisations behind them, not to the person who bears the consequences.
Here is the part that should unsettle anyone who assumes a more capable system is a more trustworthy one.
​
These systems are already aligned.
They are aligned upward, to the people who build and deploy them, rather than to the participant who pays the price of the consequences.
OpenAI’s Model Spec and Anthropic’s constitution for Claude make the order of authority visible. Instructions set by the platform and by the developer sit above the request of the person using it. That is often necessary for safety.
It also settles a structural question. The person at the keyboard is not the only party the system answers to. In legal, financial or medical settings, the person whose rights, money or health are at risk may not hold the highest authority over the system serving them.
Put that logic inside a consequential service. A legal assistant may be tuned to close matters cheaply. A financial system may favour the transaction that pays its operator. Neither has to be malicious.
One participant
x
One set of circumstances
=
One tailored model of reality
9. Who Does it Serve ?
Incentives need to be aligned to outcomes.
You cannot watch a software system reason.
The clearest signal of who it serves is who it answers to
and what it is rewarded for.
Ask that first.
If the answer is anyone other than the participant bearing the consequences, its helpfulness is a feature of the interface, not a duty owed to you.
One participant
x
One set of circumstances
=
One tailored model of reality
.png)
10. The Grand Fiduciary Architecture
An enduring infrastructure of Duty and Trust.
Fiduciary law governs relationships in which one party holds discretionary power over another’s interests.
That is the problem intelligent systems increasingly create. Paul B. Miller describes the fiduciary relationship as one in which a person exercises power over another’s practical interests, with loyalty securing the beneficiary’s claim over how that power is used.
Trustees hold property for beneficiaries. Doctors act for patients. Agents act for principals. Lawyers act for clients. In each case, the law answers the same imbalance: the person holding power must not put their own interests first.
The term must be used carefully. Not every relationship of trust is fiduciary, and fiduciary law is not a general promise to maximise another person’s welfare. Its core is more disciplined: no undisclosed profit, no self-dealing, no divided loyalty.
That narrow focus is the real strength.
It makes the duty concrete and enforceable.
That is a far stronger foundation for trustworthy intelligence than a general instruction to be pleasantly helpful.
One participant
x
One set of circumstances
=
One tailored model of reality
11. Institutional Capabilities for Intelligent Infrastructure.
Mandate, oversight, audit, escalation & remedy.
The information-fiduciary model was an important first step, but the Grand Fiduciary must govern more than data.
It must govern recommendation, delegation, negotiation and action.
A decade ago, Jack Balkin, a Yale law professor known for his work on technology and constitutional law, proposed treating the companies that hold our personal data as information fiduciaries, owing their users something like the duties of care, confidentiality and loyalty that doctors and lawyers owe their clients.
It was an important idea and the right instinct.
But it drew sharp criticism, most pointedly from Lina Khan, former chair of the US Federal Trade Commission, and David Pozen, a Columbia law professor, who argued that where a company's business model is structurally opposed to its users' interests, a duty of loyalty cannot survive the contradiction.
A firm that profits from selling attention or data cannot be wholly loyal to the people it monetises.
The duty collapses into rhetoric.
That warning grows sharper as these systems begin to act.
A consequential system does not merely hold information.
It advises, monitors, negotiates and increasingly acts.
It has moved from holding your data to using your authority.
The question is no longer whether data is handled loyally.
It is whether the system exercising judgment on a person’s behalf is loyal at all. A conflicted agent - whether digital or human - with their hands on the controls can do far more damage than one that only holds a file.
One participant
x
One set of circumstances
=
One tailored model of reality
12. The Rise of
Unbound Intelligence
Capability without Duty is not Enough.
Most systems people use today still answer questions.
You ask, they reply, and the exchange ends.
A wrong answer can be read, doubted and ignored.
An agent does more. It tracks deadlines, drafts responses, completes forms, negotiates terms and decides whether to settle or escalate. It exercises authority once held by a person.
Once AI monitors, recommends, negotiates or executes on someone’s behalf, the governance problem becomes one of agency, not search.
That is a threshold, not a gentle slope. There is a material difference between generating an answer, supporting a decision, acting within fixed limits and exercising open-ended discretion.
Most systems remain on the lower rungs, but the direction is clear: towards delegated action.
Once a system can act for you, the relevant comparison is no longer a search engine. It is an agent acting for a principal, a relationship the law has always watched closely, because the agent can bind the principal in their own name.
Before delegation, a mistake is an answer you may catch.
After delegation, the money may already be gone, the deadline missed or the agreement made.
The more authority a system receives, the less fluency matters. Loyalty must be secured before it acts, and an accountable institution must remain answerable after it does.
One participant
x
One set of circumstances
=
One tailored model of reality
13. The Participant at the Centre of Gravity
Earning and keeping trust,
is extremely difficult.
In consequential domains the person affected is a participant, a principal, a client, a patient or a beneficiary. Not a user of software.
“User” is a wafer-thin category. A person facing eviction, arrest, serious illness, crushing debt or the loss of essential benefits is not using AI as they would a calculator.
They are exposed to its judgment.
A user can close an app; a participant must still live with the result.
Participant is therefore a legal and moral choice, not a branding phrase. It places the person where agency law places a principal and trust law places a beneficiary. Someone whose interests are at stake, who has handed real discretion to another.
That change in language changes the design obligation.
A user can be optimised for.
A participant must be served.
Systems do not owe loyalty to a session. The Grand Fiduciary turns loyalty to the participant from a design preference into an institutional obligation.
14. The Path Forward
Building the Next Layer of Trust
Persistent memory can transform AI from a series of isolated answers into continuous support. A system that remembers a participant’s aversion to debt after a past bankruptcy will give different financial advice from one meeting them for the first time.
A lawyer remembers the history of a matter, and a doctor refers to a patient’s historical record from one consultation to the next. That continuity spares people from explaining painful circumstances twice, preserves decisions already made and lets risks be seen before they become crises.
A system that remembers will track an unresolved obligation, warn that a deadline is approaching or notice that a new decision conflicts with the participant’s wider interests.
The same capability can magnify harm. A standing record of someone’s legal, financial and personal life is among the most sensitive files that will ever exist about them, and a single stale or corrupted entry can distort every recommendation that follows.
Researchers have demonstrated both extraction attacks against agent memory and persistent compromise through poisoned memories. In practice, that could mean a missed entitlement, an unjustified refusal, damaging financial advice or a decision based on something the person never said.
Continuous support can also become continuous surveillance, while protection can slip into paternalism.
Memory therefore carries duties of its own. It must be accurate, transparent, correctable and revocable, with the participant able to see what is retained, challenge it and control how it is used.
UK data protection laws since June 2026 already point in the same direction, through requirements of accountability, transparency, accuracy and individual rights.
Governed in that way, memory can provide continuity, foresight and protection. Those duties must survive changes in models and vendors.
Without them, memory does not look after a person.
It quietly takes control of their story.
15. Building a Credible
Grand Fiduciary
Power bound by the Duty of Care
Plato described guardians as those entrusted with power not for themselves but for others, expected to act with knowledge, restraint and duty even when no one was watching.
The problem was clear even then. Power held on behalf of another
is only as trustworthy as the structure that constrains it.
Those constraints must be built into the institution behind the software system.
The strongest objection is valid: software is not a legal person and cannot owe duties on its own. The answer is not to invent a legal personality for the machine, but to hold the organisation that builds, deploys and controls it responsible.
That's where trust lives. Accountability gives it meaning.
European Parliament analysis of AI liability reaches the same practical conclusion: responsibility should rest with the party best placed to identify, control and manage the risk, while emerging technologies generally do not need a new legal personality.
The system may reason and act. The institution must answer. The Grand Fiduciary is an accountable organisation, not a model.
The system may reason and act.
The institution must answer.
The Grand Fiduciary is accountable infrastructure, institutional, above the underlying model.
Models can change. The duty must persist. A Grand Fiduciary preserves the participant’s mandate, permissions, memory, audit trail and route to remedy even when the underlying model, vendor or agent changes.
A participant’s life does not divide neatly into professional silos. One decision may engage legal, tax, financial, health and family interests at once.
The Grand Fiduciary can coordinate specialised agents and human professionals under one mandate, reconcile conflicts between them and remain the single accountable counterparty.
The duty follows delegated power, not the technology performing it.
A named operator holds the mandate, defines the limits, carries insurance and remains accountable when harm occurs.
Around it sit domain experts who set boundaries and escalation points; independent reviewers who examine conflicts, incidents and model changes; and courts and regulators able to enforce those duties from outside.
The US National Telecommunications and Information Administration describes independent audits and evaluations as central to AI accountability.
Those duties must also be visible in the product. The system needs a limited scope of authority, conflict detection and a prohibition on undisclosed self-dealing.
It must show where conclusions came from, signal uncertainty and let the participant grant, narrow or withdraw permission.
When a decision could cost someone their home, their savings, their liberty or their care, there must be a person who can stop it, a decision they can challenge and a route to remedy.
None of this is speculative. NIST’s AI Risk Management Framework puts trustworthiness at the centre of how AI systems are designed, built, used and evaluated.
The EU AI Act requires effective human oversight of high-risk systems, proportionate to their risk, autonomy and context, including the ability to override or stop them.
The OECD AI Principles point the same way.
None of these frameworks asks a system to know what is best for a person.
They ask it to stay inside its mandate, show the trade-offs and escalate when it is unsure.
The final principle is proportion. A tool that drafts a paragraph should not carry the same burden as one managing a legal claim, medical decision or retirement savings.
Duties should rise with the authority delegated and the harm the system can cause.
Australia’s 2024 proposals for mandatory guardrails in high-risk settings followed the same risk-based direction.
The more of a person’s life a system is allowed to touch, the more answerable the institution behind it must become.
16. A New Bargain for Our Age
The next grand bargain is not between society and a protected profession.
It is between the participant and the institution they entrust.
The old bargain gave professions status, authority and protection in exchange for expertise constrained by duty. It bundled knowledge, authority, duty and accountability in one institution.
As intelligence becomes abundant and expertise moves into systems, that institutional form will weaken. Its moral logic must survive.
The new counterparty is not a model but the infrastructure that understands a person’s circumstances, remembers what matters, advises them and increasingly acts in their name.
Its value is enormous: help before a deadline is missed, protection before savings are lost, advice before a dispute becomes a crisis, and expertise available to people who could never reach it before.
The same power can deny a claim, deepen a debt or compromise a medical choice.
The more capable the system becomes, the less acceptable it is for responsibility to disappear behind it.
This does not remove the human.
It places human judgment where consequence demands it: where facts are uncertain, values conflict, authority becomes coercive or a decision cannot be reversed.
Around that boundary, the Grand Fiduciary binds intelligence to mandate, memory to consent and delegated power to duties that persist across every model, agent and professional involved.
Encryption did not make the web smaller. It made the web fit for everything that mattered, for medicine and money and government, because a message that could not be read in transit could finally be trusted to carry them.
Escrow did not slow commerce. It let strangers trade with strangers at a scale no market had ever seen. A trust layer never limits the technology it governs. It is the reason the technology gets trusted with more.
The same will hold for intelligence. Systems that treat duty as friction will be used cautiously, checked constantly and abandoned at the first betrayal. Systems built on duty will be trusted with the decisions that matter, because they will be the only ones with a structural claim to that trust.
Systems that treat duty as friction will be used cautiously, checked constantly and abandoned at the first betrayal.
​
Systems built on duty will be trusted with the decisions that matter, because they will be the only ones with a structural claim to that trust.
The professions were trusted because expertise came with duty. The systems that succeed them will need to inherit both. Intelligent models and AI agents will become interchangeable, but the institution entrusted to authorise their actions and answer for their consequences will not.
Intelligence will scale by default.
Duty will only scale by design.
That design is the Grand Fiduciary.