Syntheticauthenticity.

The term was coined by Prashant Saxena, a Singaporean practitioner-scholar and PhD researcher at Nanyang Technological University. It is formalised with Andrew Prahl in the threshold model of synthetic authenticity, published in the International Journal of Human-Computer Interaction in 2026.

The formal definition

Synthetic authenticity has three parts, and all three have to hold at once.

• Awareness stays intact. The persona is built, trained, prompted and deployed, and the audience knows it. A social robot is a robot, a virtual influencer is rendered, a companion app is software. This is not deception.
• A role is granted anyway. People permit the persona to shape what they believe, buy, learn, disclose, decide or delegate. They follow the recommendation, accept the explanation, disclose something private, or hand over the decision.
• The grant is bounded and revocable. Legitimacy attaches to a role rather than to the system's capability, it is threshold-shaped rather than linear, and it can be withdrawn. Past a point, added realism buys scrutiny instead of trust.

The distinction that matters most: AI capability determines what a system can do. Synthetic authenticity determines the role people will allow it to hold. As AI personas gain memory, continuity, identity and the capacity to act, the second question becomes the commercially decisive one. For brands, governments and institutions it reads: which AI persona will people permit to speak and act on our behalf?

Why the term exists

Existing vocabulary forced a false choice. Research on anthropomorphism and the uncanny valley asks how human a system seems. Research on deception asks whether people were misled. Neither describes the ordinary case, which is a person who knows perfectly well they are talking to a machine and decides to act on it anyway.

That case is not deception and it is not naivety. It is a judgment made under known artificiality, and it needed a name.

The threshold model

The threshold model of synthetic authenticity was built from a map of 2,685 articles and 111 AI-persona studies. It holds that human cues help up to a threshold and then start breaking trust. Below the threshold, a persona reads as too thin to engage with. Above it, added realism invites inspection.

The design implication is calibrated sufficiency: carry the cues the task requires, and deliberately stop there. This reframes persona design from a realism race into a calibration problem, where the right amount of human signal depends on the role the persona occupies and the stakes of the decision.

Where synthetic authenticity applies

The construct is tested across six settings, because the same judgment behaves differently depending on what is at stake.

• Social robots and care. How human a helper should feel, and why the threshold matters more than realism.
• AI influencers and disclosure. Disclosure lifts engagement and taxes felt authenticity at the same time. See the transparency tax.
• Comment threads. How doubt about a persona spreads through a crowd, via crowd forensics and warrant braiding.
• Advice communities. What groups criticise when a recommendation came from an AI rather than a person. See scrutiny composition.
• Companion AI. Why people confide in something that asks nothing back. See burdenless listening.
• Classrooms and institutions. What all of this does to marketing pedagogy and public confidence.

What synthetic authenticity is not

The term is narrow on purpose, and four adjacent things get mistaken for it.

• Not deepfakes or impersonation. Those turn on concealment. Synthetic authenticity is about judgments people make when artificiality is known.
• Not synthetic data. Unrelated. Synthetic data means machine-generated training records. This is about personas and audience judgment.
• Not authenticity theatre. Performed candour by a human brand or executive is a different problem, though the two interact.
• Not AI capability. Capability is what a system can do. Synthetic authenticity is the role people will let it occupy. The two come apart constantly: highly capable systems are refused consequential roles, and modest ones are granted them.
• Not a claim that AI personas deserve trust. The construct describes when people extend it. Whether they should is a separate, and often uncomfortable, question.

Related terms

Eleven further constructs come out of this research program, including the transparency tax, collaborative persuasion knowledge, burdenless listening, the public advice object and ontological load. They are defined in full on the concepts page.