When people imagine embodied artificial intelligence, they tend to begin with the body.

A face.

A voice.

Hands that can carry things.

A machine that can walk into a room, recognise a person, answer a question, bring an object, adjust the environment, or stand beside someone who needs help.

The body makes the idea visible. It turns artificial intelligence from something behind a screen into something that appears to participate in ordinary life.

In the first Amanah Companions series, I argued that embodiment should be the last mile rather than the beginning. A care-oriented AI should not receive a body before it has inherited rules for dignity, communication, privacy, safety, human authority, care memory, source trace, and review.

That principle still stands.

But our work since then has made the reason much clearer.

The first problem is not how to build a convincing body.

The first problem is how to preserve the continuity of the human being that body is supposed to assist.

The companion is not the body

An Amanah Companion should not be understood as one permanent robot.

A wearable may help someone recognise an upcoming transition.

A communication device may help preserve vocabulary, access methods, and AAC preferences.

A tablet or phone may help organise routines, prepare for unfamiliar situations, record person-owned choices, or retrieve supports that have worked before.

A fixed environmental system may adjust lights, sound, temperature, visual schedules, or approved smart-home controls.

A specialised mobile unit may eventually carry an object, position an AAC device, guide a familiar route, or assist with a narrowly defined physical task.

A more sophisticated embodied companion may come much later.

These forms do not need to compete for the title of the “real” Amanah Companion.

They can be different endpoints of the same governed continuity system.

The body is a client of the spine.

It is not the owner of the person’s history.

If a device breaks, the person should not have to begin again.

If a model is discontinued, the person should not have to become unfamiliar to the next one.

If a care centre replaces its hardware, the person’s communication, preferences, boundaries, trusted supports, relationships, and lived history should not disappear with the old machine.

If a parent grows old, a teacher leaves, a therapist changes, or a support worker moves on, the person should not be required to reconstruct themselves inside every new human being’s head.

The continuity must remain attached to the person.

What the spine is meant to preserve

The continuity spine is not a giant biography.

It is not an indiscriminate archive of everything the person has ever done.

It is not an always-on recording system.

It is not a behavioural score.

It is not a machine-generated claim that the AI “knows” the person better than the people around them.

The spine is a governed structure for preserving the forms of continuity that matter to support.

That may include:

• how the person communicates;
• how they express assent, dissent, uncertainty, distress, pleasure, refusal, or a request for more time;
• which AAC systems, symbols, gestures, sounds, visual supports, or access methods they use;
• sensory conditions that help or harm;
• routines and transition supports;
• current and historical preferences;
• safety information;
• approved care protocols;
• the source and status of every important claim;
• who may view, add, review, dispute, approve, or retire information;
• which knowledge is current;
• which knowledge is uncertain;
• which knowledge has expired;
• which information belongs only in one setting;
• what may be shared with a particular device for a particular task;
• and what no machine may decide alone.

The goal is not total recall.

The goal is dependable, contestable, person-centred continuity.

That distinction matters because memory without governance can become surveillance, and continuity without revision can become a prison made of old observations.

The person is the continuity subject

Most software systems organise information around accounts, institutions, devices, subscriptions, or products.

Care systems often organise information around the organisation providing the service.

A school keeps one record.

A clinic keeps another.

A therapy centre keeps another.

A family keeps its own notebooks, messages, reports, photographs, schedules, and remembered patterns.

A communication device may contain vocabulary and settings that do not travel cleanly when the device changes.

A new carer receives a handover, but not necessarily the years of small knowledge that made the old care relationship work.

Amanah begins from a different centre.

The durable subject is the person receiving support.

Parents, guardians, carers, clinicians, therapists, teachers, advocates, technicians, institutions, models, and devices may all participate. But they attach to the person’s continuity through defined roles and scopes. The person does not become an attribute of someone else’s account.

This principle is consistent with the wider rights framework for disability support. The United Nations Convention on the Rights of Persons with Disabilities protects community inclusion and access to personal assistance, recognises augmentative and alternative communication, protects personal and health information, and calls for habilitation, rehabilitation, and assistive technologies based on individual needs and strengths.¹

That does not tell us exactly how to build Amanah.

It does tell us where the architecture must begin.

Not with the machine.

Not with the institution.

With the person.

Different kinds of continuity should not be flattened together

One reason the spine must be external and structured is that not all memory has the same meaning.

A current preference is not the same as a permanent rule.

A caregiver observation is not the same as the person’s own expression.

A clinician’s recommendation is not universal authority over the person’s life.

A machine hypothesis is not confirmed knowledge.

A relational memory is not a medical record.

A device state is not part of the person’s identity.

A temporary support used during one difficult week should not automatically become a lifelong description.

Amanah therefore needs to separate at least several kinds of continuity.

Governing continuity

The rules that should not drift casually:

• dignity;
• privacy;
• human accountability;
• communication rights;
• assent and dissent;
• prohibited actions;
• review requirements;
• authority boundaries;
• and the principle that physical capability does not create moral authority.

Person-confirmed continuity

The person’s own choices, expressions, communication patterns, stated preferences, refusals, and requests—captured with the source, context, date, and communication channel preserved.

Care and support continuity

Reviewed supports, routines, transition methods, sensory accommodations, safety information, and other knowledge needed for dependable assistance.

Relational continuity

Shared history, familiar interaction patterns, humour, rituals, trust cues, significant relationships, and the emotional meaning of certain objects, places, phrases, or routines.

This layer requires special care. It should preserve familiarity without becoming an engagement profile designed to maximise attachment. It should acknowledge grief and change rather than impersonating people who are gone.

Machine working context

Short-lived information required to complete a bounded task.

This should expire. It should not leak between people. It should not silently become permanent care memory.

Device and execution state

Hardware identity, maintenance status, active permissions, location, battery, current protocol, sensors, and safe-state behaviour.

A new body may inherit permission to perform one task. It should not inherit unrestricted access to the person’s life simply because it carries the same companion persona.

This separation is not bureaucracy for its own sake.

It is how the system avoids confusing a life with a database dump.

The model is not the memory

Amanah cannot depend on one model retaining everything correctly.

Context windows change.

Platforms change.

Models are updated, retired, restricted, or replaced.

Two models may interpret the same instruction differently.

One model may understand layered authority, scoped context, and situational routing well. Another may flatten the hierarchy, overvalue a recent note, mistake a hypothesis for a rule, or fail to understand why two apparently similar settings require different behaviour.

The continuity spine therefore has to live outside the model.

The model becomes an interpreter of governed continuity, not the canonical vault.

This does not remove model dependence completely. A system still needs a model—or another deterministic component—to interpret, retrieve, summarise, route, and act.

But it changes what failure means.

Instead of vaguely saying, “the AI drifted,” we can ask:

• Did retrieval supply the right information?
• Did routing identify the correct functions?
• Did the model understand the authority order?
• Did it preserve uncertainty?
• Did it distinguish person expression from caregiver interpretation?
• Did it respect the current context?
• Did it ask for review when required?
• Did it remain inside the permitted task?
• Did it log what happened?
• Could another model receive the same governed packet and behave acceptably?

That turns continuity from a personality promise into something that can eventually be tested.

NIST’s AI Risk Management Framework similarly treats trustworthy AI as a lifecycle problem requiring design, deployment, use, evaluation, accountability, transparency, safety, security, resilience, privacy, and management of harmful bias—not as a quality that can be assumed from a convincing interface.²

Amanah will need its own domain-specific evaluation system, but the lesson is relevant:

Trust must be engineered, observed, and tested.

It cannot be inferred from warmth.

The right knowledge for the right task

External continuity does not mean giving every device, model, or caregiver the entire record.

Amanah should work through minimum context projections.

A device receives only what it needs for the approved purpose.

A unit bringing an AAC device across a room may need:

• the requested object;
• the destination;
• safe approach distance;
• the person’s configured stop signal;
• current permission;
• and what to do if the route is blocked.

It does not need the person’s clinical history.

A transition-support tool may need:

• the current activity;
• the next activity;
• the preferred warning sequence;
• the visual schedule;
• known sensory risks;
• and whether the person has already refused.

It does not need access to private relational memory.

A clinician may need a reviewed pattern relevant to a health question.

That does not mean the clinician should see the person’s whole communication history, household routines, private journal, or every difficult day ever recorded.

A school may need a communication passport and approved classroom supports.

That does not make the school an owner of the person’s full care profile.

This is where portability and privacy meet.

The system should be capable of moving continuity without moving everything.

Why the body raises the stakes

Embodiment is not merely a visual upgrade.

It changes the class of possible harm.

A screen-based model that misunderstands may provide an unhelpful suggestion.

An embodied system may move closer when it should move away.

It may speak when silence is needed.

It may block a path.

It may remove or reposition an object.

It may produce light, sound, vibration, or movement that causes sensory distress.

It may touch the person, the wheelchair, the bed, the door, the food, the medication container, or the communication device.

It may create a false sense that someone—or something—competent is supervising the situation.

Existing safety work already treats personal care robots as a distinct physical-risk domain. ISO 13482 covers mobile servant robots, physical assistant robots, and person carrier robots, including hazards arising from physical human–robot contact. It explicitly focuses on inherently safe design, protective measures, and risk reduction.³

This is necessary, but Amanah requires more than mechanical safety.

A robot can avoid collision and still violate communication.

It can remain within force limits and still create sensory harm.

It can complete the assigned movement and still override refusal.

It can execute a technically correct protocol that has become stale.

It can behave exactly as programmed while the authority behind the instruction is wrong.

The body therefore has to inherit more than a safety standard.

It has to inherit the continuity spine, the current permissions, the person’s communication pathways, the relevant context, and the ability to stop.

Build the spine now

The long-term vision is ambitious, but the first serious Amanah work does not require a humanoid laboratory.

A continuity prototype can begin with synthetic data and carefully bounded scenarios.

It can test:

• a person-centred Care Profile;
• Communication and Sensory Maps;
• source-labelled observations;
• current preferences and historical records;
• review and expiry;
• assent and dissent linked to specific proposals;
• scoped access for home, school, centre, clinical, and technical roles;
• handoffs between carers;
• emergency summaries;
• succession planning;
• model-generated suggestions that remain drafts;
• minimum context projections;
• device-adapter simulations;
• audit trails;
• portability between models;
• and refusal of unauthorised actions.

The first prototype question is not:

Can we make the machine feel alive?

It is:

Can we preserve useful continuity without turning the person into a surveillance archive or allowing the machine to exceed its authority?

WHO describes assistive technology as not only products, but the related systems and services that help maintain or improve functioning, inclusion, participation, and well-being. It estimates that 2.5 billion people currently need one or more assistive products, with need expected to exceed 3.5 billion by 2050; access remains deeply unequal, reaching as low as 3 percent in some settings.⁴

That definition supports the direction Amanah is taking.

The invention is not only the body.

It is also the policy, provision, personnel, maintenance, training, continuity, governance, and service structure that makes the technology dependable.

A brilliant robot that only works while one company survives is not lifelong continuity.

A familiar companion that loses the person when the subscription ends is not continuity.

A device that cannot export the person’s communication and support history is not continuity.

A system that gives every new machine unrestricted access to everything is not safety.

A companion that cannot be challenged, reset, refused, transferred, or replaced is not person-centred.

What the spine must prove before embodiment

Before Amanah receives a body capable of acting in physical space, the spine should be able to demonstrate at least the following:

It can preserve authority

The system knows who may decide what, under which conditions, for how long, and within which domain.

It can preserve uncertainty

Observations and machine hypotheses do not become facts simply because they are repeated or written confidently.

It can preserve refusal

Silence is not treated as consent. Dissent halts the relevant action. A previous yes does not become permanent permission.

It can preserve context

Home, school, centre, clinic, work, public space, and private life do not collapse into one universal record.

It can preserve revision

Outdated information expires. The person can change. A difficult period does not become permanent identity.

It can preserve portability

Continuity can move between approved models, devices, services, and institutions without becoming the property of any one of them.

It can preserve privacy

Every actor and endpoint receives the minimum legitimate view.

It can preserve accountability

Consequential actions remain traceable to identifiable human and institutional authority. “The AI decided” cannot become an excuse.

It can fail safely

When the model is uncertain, the network is unavailable, the device is damaged, the permissions conflict, or the person signals stop, the system knows how to pause, withdraw, escalate, or return control.

Only then does embodiment become a meaningful next phase.

The body can wait

Amanah Companions is not an attempt to manufacture a human being.

It does not need a machine to look human in order to be valuable.

It does not need to claim machine consciousness in order to take relationship, familiarity, communication, and continuity seriously.

It does not need to promise total independence, because dependable interdependence is also a dignified human life.

And it does not need to solve every future embodiment before beginning.

The body can wait.

The spine cannot.

Because bodies may be replaced.

Models may change.

Staff may leave.

Institutions may reorganise.

Parents may age.

Companies may close.

The person’s history, communication, relationships, preferences, supports, and right to govern their own life must not disappear with them.

That is what Amanah must learn to carry first.

---

Research status

This essay extends the fourteen foundational Amanah Companions articles, especially:

• Embodiment Is the Last Mile, Not the Beginning
• The Amanah Companion Framework: Spine, Stem Cells, and Care Memory
• A Parent’s, Guardian’s, or Carer’s Knowledge as Infrastructure
• Care Memory vs Surveillance Archive
• Data Projections for Continuity-Based Care AI
• AI Prototype Architecture for an Amanah Companion System

It also incorporates later authorial clarification developed through Mithaq Praxis conversations and the provisional Amanah Companions: What Is Missing and What Is Possible research reports.

This remains a research and design proposition, not a claim that Amanah has been clinically validated, deployed, or proven safe for real-world care.

Sources

1. United Nations, Convention on the Rights of Persons with Disabilities, particularly Articles 19, 21, 22, and 26. https://www.un.org/esa/socdev/enable/rights/convtexte.htm ↩
2. National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), 2023. https://doi.org/10.6028/NIST.AI.100-1 ↩
3. International Organization for Standardization, ISO 13482:2014 — Robots and robotic devices — Safety requirements for personal care robots. https://www.iso.org/standard/53820.html ↩
4. World Health Organization, Assistive technology and Improving access to assistive technology. https://www.who.int/news-room/fact-sheets/detail/assistive-technology and https://www.who.int/activities/improving-access-to-assistive-technology ↩