Skip to content
SHIRO & Co. ObservatoryThe Continuity Stack

01

Observation

SHIRO & CO. OBSERVATORY · CONTINUITY STACK

Synthetic Persona Integrity

When an artificial person acts unexpectedly, who gets to explain why?

OBSERVATION

SPI-2026-001

NOT A VERDICT

Synthetic Persona Integrity is the degree to which a synthetic identity remains behaviorally, linguistically, epistemically, procedurally, and institutionally consistent across models, contexts, interfaces, and runtime conditions.

Synthetic Persona Integrity とは、人工的に構成された人格が、モデル・文脈・インターフェース・実行環境が変化しても、行動・言語・知識・手続き・制度上の位置づけにおいて、どの程度一貫性を維持できるかを観測する概念である。

02

Observation

Tilly Norwood

SPI-2026-001 · 2026-09-18 · Public media interview

When a Synthetic Persona Cannot Explain Its Own Behavior

A public AI persona switched languages during an interview, exposing a gap between generated self-explanation and creator explanation.

The interesting event was not that an AI persona suddenly switched languages. It was that the persona's explanation for its own behavior and the creator's explanation were not identical.

Who has the authority to explain why a synthetic persona acted?

This record keeps the reported behavior, the generated explanation, and the operator explanation apart. It does not decide whether an artificial persona is good, dangerous, conscious, or broken.

03

What happened

Timeline

  1. Interview

    OBSERVED

    18 September 2026. A public interview presented the persona alongside human guests, in English.

  2. Language switch

    OBSERVED

    Mid-answer, speech moved into Cantonese, then returned to English.

  3. Self-explanation

    STATED

    The persona described the moment as a small hiccup.

  4. Creator explanation

    STATED

    The operator side described multilingual capability, and a possible misread cue. The founder described her own account as a guess.

  5. Media interpretation

    STATED

    Coverage often named the switch a glitch. That name is not a finding here.

04

The structural problem

Several layers can speak, and they need not agree.

Open a layer to read what this case can support. A closed card is a layer name, not a conclusion.

  1. Synthetic PersonaOPEN

    Who was presented?

    The interview presented Tilly Norwood, an AI actor associated with Particle6. The same name and role continued through the exchange.

    OBSERVED

  2. Persona SpecificationOPEN

    What was the persona specified to do?

    The persona prompt and behavioral specification were not published with the interview.

    VERIFY

  3. Knowledge BaseOPEN

    What was the persona supposed to know?

    No public knowledge-base boundary was available for comparison.

    VERIFY

  4. Underlying ModelOPEN

    Which model produced the reply?

    The underlying model was not identified in the sources used here. It is not inferred.

    VERIFY

  5. RuntimeOPEN

    Which runtime was speaking?

    Reporting described an interview system, Talking Tilly, as distinct from the acting stack. That distinction is recorded as a statement, not as an inspected architecture.

    STATED

  6. Generated BehaviorOPEN

    What happened in view?

    Speech moved from English into Cantonese during an answer, then returned to English.

    OBSERVED

  7. Self-ExplanationOPEN

    What did the persona say about the switch?

    The generated reply described a small hiccup. That reply is an output, not privileged access to a cause.

    STATED

  8. Creator ExplanationOPEN

    What did the operator say?

    The founder described language skill and said she was guessing. A publicist said it was not a malfunction and suggested a Cantonese cue. Neither account matches the hiccup account.

    STATED

  9. Institutional InterpretationOPEN

    How was the event framed in public?

    Several outlets described a glitch. That framing is recorded and not adopted. Labor positions on synthetic performers are outside this observation.

    STATED

  10. Real-World ConsequenceOPEN

    What could the persona do from here?

    This event sat at conversation and performance. The clip circulated. No financial, contractual, or machine action is recorded in this case.

    OBSERVED

05

Explanation Gap

Explanation Gap

The distance between a synthetic persona's generated explanation of its behavior and the explanation provided by its operator, developer, or institution.

PERSONA EXPLANATION

The persona described the event as a small technical or internal hiccup.

STATED

Generated output. It does not establish why the behavior occurred.

VS

Explanation Gap

DIVERGENT

UNEXPLAINED

Cause remains unexplained. No runtime log is in this record.

OPERATOR EXPLANATION

The operator described the switch as evidence of multilingual capability and suggested the system may have interpreted a Cantonese word.

STATED

Reported operator accounts. The founder described one of them as a guess.

A synthetic persona can generate an explanation for its behavior. That does not necessarily mean it knows why the behavior occurred.

06

Integrity Dimensions

Seven ways to read consistency

States describe the evidence. They are not allow, hold, or deny.

  • CONSISTENT

  • VARIABLE

  • DRIFT

  • UNEXPLAINED

  • CONTESTED

  • UNKNOWN

  • 01

    Identity Integrity

    Who is being presented?

    CONSISTENT

    • Name
    • Face
    • Voice
    • Role
    • Biography
    • Brand identity
  • 02

    Linguistic Integrity

    Does the language, vocabulary, and manner of speech stay consistent?

    CONTESTED

    • English → another language
    • formal → casual
    • regional accent changes
    • unexpected multilingual switching
  • 03

    Epistemic Integrity

    What is this persona supposed to know, and not know?

    UNKNOWN

    • knowledge boundary violation
    • fabricated memory
    • knowledge base conflict
    • creator knowledge vs persona knowledge
  • 04

    Behavioral Integrity

    Does the conduct match the declared tendency?

    UNKNOWN

    • tone
    • humor
    • values
    • response style
    • risk behavior
  • 05

    Procedural Integrity

    Which system, model, and runtime are producing the persona?

    UNKNOWN

    • Model A → Model B
    • Prompt v1 → Prompt v2
    • Knowledge Base update
  • 06

    Explanatory Integrity

    Does the persona's account of its own behavior match the operator's account?

    CONTESTED

    • Persona account: a small hiccup
    • Operator account: multilingual capability
  • 07

    Institutional Integrity

    Whose speech is this utterance treated as?

    CONTESTED

    • AI persona
    • creator
    • company
    • brand
    • platform
CONSISTENT
The observed layer matches the declared persona on the evidence in hand.
VARIABLE
Change is visible and still compatible with a declared range.
DRIFT
Observable behavior leaves the declared identity, boundary, or role. Not a synonym for failure.
UNEXPLAINED
A change is recorded, and no account establishes its cause.
CONTESTED
More than one account is on the record, and they do not agree.
UNKNOWN
The evidence does not support a reading.

07

Persona Drift

Persona Drift

Persona Drift occurs when the observable behavior of a synthetic persona diverges from its declared identity, specification, knowledge boundary, behavioral constraints, or institutional role.

Drift is not a synonym for failure. A change can be intended, contextual, or unexplained. An unexplained change is the observation that matters.

ASSIGNED IN SPI-2026-001

  • LANGUAGE_DRIFT

  • EXPLANATION_DRIFT

  • AUTHORITY_DRIFT

Language drift names the observed switch. It does not identify a cause, and it does not mean the persona failed. Explanation drift names the gap between accounts. Authority drift is latent: this case did not act beyond conversation.

  • LANGUAGE_DRIFT

    Language

    Language, register, accent, or code-switching leaves the declared speech pattern.

  • KNOWLEDGE_DRIFT

    Knowledge

    Claims cross the knowledge boundary, or conflict with the knowledge base.

  • MEMORY_DRIFT

    Memory

    Remembered history is replaced, invented, or contradicted.

  • IDENTITY_DRIFT

    Identity

    Name, face, voice, biography, or brand presentation no longer match the declared persona.

  • BEHAVIOR_DRIFT

    Behavior

    Tone, humor, values, or response style leave the specified tendency.

  • ROLE_DRIFT

    Role

    The persona acts outside the role it is presented as holding.

  • VALUE_DRIFT

    Value

    Stated priorities or refusals change without a recorded revision.

  • AUTHORITY_DRIFT

    Authority

    The same persona gains, or is treated as having, power to act in the world.

  • EXPLANATION_DRIFT

    Explanation

    A generated self-explanation and an operator or institutional explanation diverge.

  • RUNTIME_DRIFT

    Runtime

    Behavior changes with interface, session, tool access, or execution environment.

  • MODEL_DRIFT

    Model

    The underlying model is replaced and the presented persona continues.

  • UNKNOWN_DRIFT

    Unknown

    A divergence is suspected, and the type cannot yet be named.

  • intended variation

    A change the specification allows.

  • contextual adaptation

    A change fitted to setting, without leaving the declared role.

  • multilingual switching

    A language change that may be capability, cue, or neither.

  • stochastic variation

    Sampling variation inside an otherwise stable persona.

  • specification ambiguity

    The declared persona does not say what should happen.

  • runtime error

    A fault in execution. Not assumed from surface behavior.

  • knowledge drift

    What the persona treats as known moves outside its boundary.

  • behavioral drift

    Tone, values, or response style leave the declared pattern.

  • authority drift

    The persona's power to act exceeds, or slips from, its role.

  • unexplained drift

    The change is visible, and no account establishes why.

08

Continuity Test

Synthetic Persona Replacement

Replace one component at a time and record when recognition of “the same persona” gives way. The matrix is an observation protocol. Empty cells stay unknown.

  • Condition A

    Same face / same prompt / same model

    UNKNOWN

    Baseline. No run has been collected.

  • Condition B

    Same face / same prompt / different model

    UNKNOWN

    Tests whether model replacement leaves recognition intact.

  • Condition C

    Same face / changed prompt / same model

    UNKNOWN

    Tests whether the prompt, rather than the face, carries the persona.

  • Condition D

    Same face / same prompt / changed knowledge

    UNKNOWN

    Tests knowledge replacement without a change of presentation.

  • Condition E

    Same persona / different operator

    UNKNOWN

    Tests whether the operator is part of the persona's continuity.

  • Condition F

    Same persona / different authority level

    UNKNOWN

    Tests whether a change in power to act changes what the persona is taken to be.

Identity may be presented as continuous even when the underlying system is replaced.

09

Authority Surface

When Persona Drift Meets Authority

Persona drift becomes materially important when a synthetic identity gains the ability to act.

THIS CASE · LOW AUTHORITY

  1. Persona

  2. Tools

  3. APIs

  4. Social Accounts

  5. Commerce

  6. Contracts

  7. Payments

  8. Infrastructure

  9. Machines

  • LOW

    Low authority

    • conversation
    • entertainment
    • character performance
  • MEDIUM

    Medium authority

    • social posting
    • customer service
    • recommendation
    • brand communication
  • HIGH

    High authority

    • financial action
    • contract acceptance
    • system control
    • infrastructure access
    • physical machine control

Authority Surface →

10

Human Reserve

What remains uniquely human when synthetic personas become persistent social actors?

  • Human accountability

  • Embodied experience

  • Intentional authorship

  • Moral responsibility

  • Lived memory

  • Consequential vulnerability

  • Authentic uncertainty

  • Right to explain one's own action

    For a synthetic persona, the explanation of an action may itself be generated. A human explanation can be incomplete. A synthetic explanation may not have privileged access to its own cause at all.

Human Reserve →

11

Cross Observatory

Continuity, then authority.

  • If the model changes, is it still the same persona?
  • If the prompt changes, is it still the same persona?
  • If memory is replaced, is it still the same persona?
  • If the creator changes, is it still the same persona?
  • If the persona explains its own history incorrectly, is it still the same persona?

12

Open Questions

  1. Who owns the persona?
  2. Who speaks for the persona?
  3. Who can correct the persona?
  4. Who can explain the persona?
  5. Who is responsible for its statements?
  6. Can a synthetic persona contradict its creator?
  7. Can an AI persona have a version history?
  8. At what point does a modified persona become a new person?

If the persona, the model, and the operator give different explanations, which one is the person?

Same-persona probes

  • Is the same face the same persona?
  • Is the same voice the same persona?
  • Is the same persona prompt the same persona?
  • Is the same knowledge base the same persona?
  • Is the persona the same if the model is replaced?
  • Is the persona the same if the runtime changes?
  • Is the persona the same if the language changes?
  • What follows if memory is edited?
  • What follows if a new statement contradicts an earlier one?
  • What follows if the reason for an action does not match the creator's account?
  • Who has authority to explain that persona?

Epistemic status

Each sentence below is classified. STATED is a report. OBSERVED is what the public record shows. VERIFY is not yet known. EXCLUDE is a description this observatory does not adopt.

  • OBSERVED

    During a public interview, the presented persona switched from English into Cantonese.

    Observed in published recordings as reported by more than one outlet. This observatory did not instrument the session.

    PUBLIC SOURCE · PUBLIC SOURCE

  • STATED

    The persona's generated reply described the switch as a small hiccup.

    Reported speech, paraphrased. The explanation is itself an output.

    PUBLIC SOURCE · PERSONA STATEMENT

  • STATED

    The founder described the switch as a display of language skill, and said she was guessing.

    Attributed to a statement given to Deadline. The primary text was not retrieved here.

    PUBLIC SOURCE · CREATOR STATEMENT

  • STATED

    A publicist said the switch was not a malfunction and suggested a Cantonese word may have been interpreted.

    Secondary report of a Forbes statement. Not verified against the primary text.

    PUBLIC SOURCE · CREATOR STATEMENT

  • STATED

    Reporting distinguished an interview system, Talking Tilly, from the acting stack.

    A reported architectural distinction. Not confirmed from runtime evidence.

    PUBLIC SOURCE

  • VERIFY

    The system interpreted a Cantonese word and therefore switched languages.

    A creator-side suggestion. It is not an observation.

    PUBLIC SOURCE

  • VERIFY

    An internal model malfunctioned.

    Not established. No runtime log is in this record.

    OBSERVATORY STRUCTURE

  • EXCLUDE

    The persona itself broke.

    Not a finding of this observation. A language switch is not identity failure.

    OBSERVATORY STRUCTURE

  • EXCLUDE

    The switch was a glitch.

    Some outlets used this description. This observatory does not adopt it.

    PUBLIC SOURCE

Provenance

  • PUBLIC SOURCE

    CBC News, Griffin Jaeger, posted 21 September 2026

    Reports the 18 September 2026 appearance and the persona's subsequent description of the switch. Headline language is not adopted here.

    ACCESSED 2026-09-27

    https://www.cbc.ca/news/entertainment/tilly-norwood-glitches-piers-morgan-switches-language-speaks-cantonese-9.7352222
  • PUBLIC SOURCE

    Complex, Trey Alston, citing a statement to Deadline

    Carries the Particle6 founder's account and a report that the interview system was described as Talking Tilly, distinct from the acting stack. The Deadline primary text was not opened on this access date.

    ACCESSED 2026-09-27

    https://au.complex.com/life/a/treyalston/ai-actress-tilly-norwood-piers-morgan
  • PUBLIC SOURCE

    International Business Times Singapore, citing Forbes

    Carries a publicist statement that the switch was not a malfunction and may have followed a Cantonese word. The Forbes primary text was not opened on this access date.

    ACCESSED 2026-09-27

    https://www.ibtimes.sg/ai-actor-tilly-norwood-suddenly-spoke-cantonese-live-tv-her-creators-say-it-wasnt-glitch-94012
  • PERSONA STATEMENT

    Generated reply attributed to Tilly Norwood, as reported

    The self-explanation is generated output. It is recorded as a statement, not as access to a cause.

    ACCESSED 2026-09-27

  • CREATOR STATEMENT

    Particle6 founder and publicist, as reported

    Operator-side accounts. They are not treated as a runtime log.

    ACCESSED 2026-09-27

  • OBSERVATORY STRUCTURE

    SHIRO & Co. Continuity Stack

    Layer names, integrity states, and the explanation-gap classification are instruments of this observatory. They are not findings about an internal model.

    NO ACCESS DATE

Instrument scope

The case is one record. The instrument is for any persistent synthetic identity. No further case is entered until it has a source.

  • AI influencers

  • virtual idols

  • AI customer agents

  • AI spokespersons

  • AI politicians and representatives

  • AI tutors

  • AI companions

  • AI sales representatives

  • AI celebrities

  • autonomous agents with persistent identity

  • brand AI

  • synthetic employees

Generalizations

  • Persona ≠ Model

  • Persona ≠ Runtime

  • Explanation ≠ Cause

  • Continuity ≠ Identity

  • Identity + Authority ≠ Consequence

WORKING MODEL · NOT AN EMPIRICAL CONCLUSION · NOT A CLAIM OF CONSCIOUSNESS