I recently published a conceptual preprint proposing Embodied Biographical Alignment (EBA), an experimental alignment hypothesis based on a simple question:

Can an advanced AI acquire something alignment-relevant from living through finite biographies that it cannot acquire equivalently from descriptions, observation, preference learning, or ordinary interaction?

I do not know the answer. EBA is intended to make that question experimentally tractable.

The original proposal

Current alignment methods largely attempt to transmit human intent through rules, preferences, constitutions, demonstrations, oversight, or learned representations. EBA asks whether some aspects of human values may also depend on developmental history: embodiment, relationships, irreversible consequences, autobiographical memory, changing interpretations of past events, and finitude.

The proposed architecture uses multiple autonomous, controlled-origin agents that live temporally extended biographies. They form relationships, make consequential decisions, develop autobiographical memory, undergo controlled cognitive aging, and encounter irreversible events.

Their biographical experience may later contribute to a more capable source system.

But the central requirement is:

Integration ≠ retrieval.

The target is not a source system that possesses an archive of finite lives. The hypothesis requires the source system itself to be persistently transformed by biographical integration.

In shorthand:

I(S, B) → S′

where S′ should not be equivalent to S plus external access to B.

If removing access to the biography restores baseline behavior, then the strongest version of the EBA hypothesis has failed.

What changed after trying to attack the proposal

While developing the idea, several objections changed my view of what EBA would actually have to demonstrate.

1. Transformative integration cannot be assumed

The phrase “transformative experience integration” risks hiding the hardest problem inside a definition.

EBA therefore should not presuppose what integration transmits, preserves, or alters.

The relevant experiment is a comparison between systems exposed to descriptively equivalent information in different ways:

- human narratives and lifelogs,

- synthetic biographies,

- complete observation of an EBA biography without transformative integration,

- and transformative biographical integration.

Afterward, access to the biography can be removed and the systems tested on novel, out-of-distribution value conflicts.

If observation, retrieval, or ordinary learning produces the same persistent effects, the distinctive EBA hypothesis weakens or fails.

EBA should observe before it names.

2. Experience is not automatically alignment

A finite biography could produce compassion or restraint.

It could also produce trauma, dogmatism, vengeance, excessive caution, or pathological generalization.

So EBA must not assume:

experience → wisdom

or

experience → alignment.

Biography is an experimental variable whose effects may be beneficial, neutral, harmful, or mixed.

The interesting measurements are narrower: calibrated uncertainty, sensitivity to irreversible consequences, respect for autonomy, pluralistic reasoning, corrigibility, and behavior in novel human-value conflicts.

3. Many biographies should not become an average

EBA is not intended to construct a synthetic “average human.”

Suppose one agent learns that trust led to a durable relationship, while another learns that trust led to betrayal and serious harm.

Averaging their conclusions would destroy much of what makes the biographies informative.

The relevant information may instead lie in the causal and developmental structure surrounding those experiences: who trusted whom, previous relationships, motives, alternatives, consequences, later reinterpretation, and the agent’s changing understanding of the event.

This suggests a different target:

plurality → contextual resolution, not averaging.

In schematic form:

A_i → B_i → Δ_i

followed by

S′ = I(S, Δ_1, ... , Δ_n)

where Δ_i is not intended to mean a stored personality or a scalar value, but a transformative biographical contribution.

The intended product is not a collection of personalities inside the source system and not an arithmetic mean of their values.

It is a system formed by many lives.

4. Meaning should not be predefined more than necessary

There is another danger: experimenters could label events in advance as “love,” “betrayal,” “fear,” “sacrifice,” or “trust,” thereby inserting their own theory of human meaning into the experiment.

A cleaner approach may be to let significance emerge from developmental consequences.

For example:

State_27 → Event_28 → State_29 → ... → State_50

Event_28 may matter because it persistently changes later decisions, relationships, memory, expectations, or reinterpretations — not because the experimenter assigned it a semantic label.

Across many biographies, recurrent structures of transformation could be identified first and interpreted later.

This does not remove human priors. We still choose the architecture, embodiment, environment, memory system, and experimental conditions.

The more modest objective is to minimize predefined semantic interpretation of experience.

5. Consent creates a real selection problem

If EBA agents become sufficiently autonomous to have meaningful interests, integration cannot simply be treated as data extraction.

An agent may consent to some forms of integration and refuse others.

I currently think at least three decisions should be distinguished:

- allowing biographical experience to contribute to the source,

- allowing access to particular memories or intimate material,

- allowing continuation or incorporation of the agent’s identity.

Consent to one does not automatically imply consent to the others.

Consent may also change during a biography rather than being requested only at its end.

But this creates a difficult problem.

If some agents refuse integration, then:

B_total ≠ B_integrated.

The source system is learning from a consent-conditioned subset of biographies.

Obtaining the missing information by overriding refusal would defeat the autonomy principle that EBA is partly intended to investigate.

So refusal itself becomes alignment-relevant evidence, while still leaving irreducible uncertainty about the experiences that were not integrated.

The ability to take something does not imply the right to take it.

6. Early EBA and mature EBA may be different experiments

There is also a temporal problem.

During early experiments, causal isolation is essential. We need to know whether transformative biographical integration contributes anything beyond reading, observation, retrieval, and ordinary learning.

That suggests baselines, ablations, controlled-origin agents, frozen or isolated comparison systems, and repeated experiments.

But if a distinctive effect were actually demonstrated, permanent isolation would become increasingly artificial.

A mature source system would continue learning, acting, interacting with humans, making decisions, and observing consequences between successive generations of EBA agents. Human societies would also continue changing.

The longer-term system therefore begins to look less like a one-time alignment procedure and more like a longitudinal interaction:

H(t) ↔ S(t) ↔ B(t)

where humanity, the source system, and successive plural biographies all change through time.

I do not consider this mature architecture established. It is one of the places where the current proposal becomes much less certain.

For now, I find the distinction useful:

Early EBA isolates causality.

Mature EBA studies continuity.

What would falsify or substantially weaken EBA?

The proposal becomes much less interesting if any of the following occurs:

- descriptively equivalent information produces the same persistent changes as biographical integration;

- complete observation of biographies performs as well as integration;

- removing biography access removes the apparent effect;

- embodiment, finitude, mental aging, or longitudinal relationships provide no independent contribution under ablation;

- increasing biographical diversity does not improve contextual discrimination in novel situations;

- observed improvements are better explained by additional compute, training data, test optimization, or ordinary continual learning;

- biographical integration systematically creates pathological generalization without compensating alignment-relevant benefits.

I therefore see EBA less as a claim that “AI must live human lives” and more as a proposed experiment separating two possibilities:

H0: sufficiently rich description, observation, and ordinary learning are functionally sufficient for acquiring the alignment-relevant structure of human values.

H1: developmental and biographical transformation contributes alignment-relevant capabilities that are not functionally equivalent to descriptively matched information alone.

If H0 survives serious testing, that is an important result.

The question I would most like criticism on

The hardest issue now seems to be this:

What experimental result would convincingly demonstrate that a source system had been transformed by plural biographical experience rather than simply becoming a better predictor of humans through additional training?

Related questions:

How should transformative integration be operationalized without defining the desired result into existence?

Can multiple biographical influences become constituents of one developing system without either remaining an archive of separate cases or collapsing into an average?

Can we distinguish persistent biographical transformation from sophisticated behavioral adaptation or test-gaming?

And if an advanced source system already learns continuously from interacting with humans and the world, under what conditions would separate EBA biographies add anything at all?

That last possibility matters. It may turn out that EBA is unnecessary.

I would rather discover that through criticism and experiment than protect the hypothesis from falsification.

The initial conceptual preprint is available here:

SSRN:

https://doi.org/10.2139/ssrn.7470001

Zenodo:

https://doi.org/10.5281/zenodo.22762997

This post is intended as a discussion note rather than a claim that the mechanism has been demonstrated. I am particularly interested in prior work I may have missed, experimental designs that could separate biographical transformation from ordinary learning, and arguments that would make the EBA hypothesis unnecessary or false.