The Human-Referential Fidelity Test asks how faithfully an AI system represents one specific source human under a declared evaluation protocol. It measures fidelity in decisions, preferences, language, values, and reactions while keeping resemblance separate from identity, rights, and authority.
The test at a glance
Source Person
The real human whose identity is being represented
AI Twin
The system claiming to represent the source person
Informed Judges
People who know the source person and evaluate fidelity independently
Judges compare held-out decisions, preferences, and reactions between the source person and the AI twin — using evidence the twin was never trained on.
The twin matches the source person's patterns on held-out decisions within the declared scope, and judges cannot reliably distinguish their responses.
Does this system match this particular person?
A familiar voice is only the beginning
Maya authorizes a digital twin built from interviews, messages, decisions, and stories. Her friends recognize its humor, but recognition alone is not enough. HRFT asks whether the twin makes the kinds of choices Maya would make across situations that were not simply copied into its construction data—and whether the test states where the resemblance stops.
A claim that can be tested and challenged
HRFT evaluates referential fidelity: how well a human-referential AI system represents the behavioral and cognitive patterns of a named, consenting source person. The reference is not humanity in general. It is one person, in a stated period, domain, and context.
Held-out decisions
Questions, dilemmas, and preferences that were not used to build the twin reveal whether it generalizes beyond memorized examples.
Source-person judgment
The represented person can judge whether answers fit, where they fail, and which parts should never have been inferred.
Informed observers
People who know the source person can compare responses independently instead of being guided toward agreement.
Provenance and consent
Every evaluated claim needs a legitimate origin, and the protocol must define what data and relationships were authorized.
How could researchers run HRFT?
- 1
Name the source person, the consent boundary, the period being represented, and the domains in scope.
- 2
Reserve real decisions, autobiographical details, and preference questions for evaluation rather than construction.
- 3
Collect answers from the source person and the candidate under comparable conditions.
- 4
Ask the source person and informed observers to rate fidelity independently and explain important mismatches.
- 5
Publish the test limits, disagreement, and failure cases instead of reducing the identity claim to one persuasive score.
HRFT vs. the classical Turing Test
| Dimension | Classical Turing Test | Human-Referential Fidelity Test (HRFT) |
|---|---|---|
| Question | Can this machine appear human? | Does this system match this particular person? |
| Reference | A broad idea of human conversation | One named and consenting source person |
| Evidence | A conversation transcript | Held-out choices, biography, preferences, observers, provenance, and consent |
| What success supports | Human-like presentation | A bounded claim of referential fidelity |
Use plain language, not one magical score
Strong fidelity in scope
The candidate reliably matches the source person within the declared domains and time period.
Partial or domain-limited fidelity
The candidate represents some patterns well but should not generalize beyond the evidence.
Persuasive imitation
The candidate feels familiar but relies on style, stereotypes, or memorized phrases rather than deeper fidelity.
Invalid or unconsented representation
The construction or evaluation lacks legitimate provenance, consent, or an honest reference boundary.
What should HRFT expose?
Generic plausibility
The twin gives reasonable answers that could fit almost anyone.
Data recitation
The twin repeats construction records without generalizing to new situations.
Stereotype filling
Missing evidence is replaced with demographic or personality stereotypes.
The authority leap
Behavioral resemblance is treated as permission to inherit the person’s accounts, rights, or relationships.
What HRFT does not prove
Passing HRFT does not prove that the twin is the source person.
HRFT does not transfer consciousness, legal personhood, property, relationships, or authority.
HRFT does not establish that later versions remain continuous after migration or branching; that is CCT’s job.
A high-fidelity twin must still pass SIT so it does not invent a life that neither the human nor the digital system lived.
HRFT turns an intuitive identity question into a mathematically grounded, verifiable protocol. It proves that looking the part does not mean having lived the life, moving agent evaluation from subjective conversational impressions to rigorous developmental evidence.
Questions about HRFT
What is the Human-Referential Fidelity Test?
The Human-Referential Fidelity Test is PCI’s protocol family for evaluating how faithfully an AI system represents one particular consenting source human across declared tasks, evidence, evaluators, time intervals, and domains.
How does HRFT evaluate referential fidelity?
HRFT uses held-out decisions, informed observers, and domain-specific evidence to evaluate how accurately an AI twin represents the decision-making, preferences, and communication style of a specific source human.
How is HRFT different from the classical Turing Test?
The classical Turing Test asks whether a machine can appear human. HRFT asks how faithfully a human-referential system represents one specific source person using held-out decisions, informed judges, provenance, and explicit scope.
Does passing HRFT make an AI the same person?
No. Passing HRFT supports a bounded fidelity claim. It does not establish consciousness, legal identity, subjective continuity, or inherited authority.
Who should evaluate an HRFT candidate?
The source person should participate whenever possible, supported by informed observers and independent evaluators. The protocol should record disagreements rather than forcing consensus.