What is a digital twin, and how is it different from an AI replica?

Nobody wants to shut down a wind turbine to find out whether its gearbox is wearing out. A digital twin solves that problem with a live virtual copy of the turbine, fed by its sensors, that engineers can inspect from a desk while the real machine keeps spinning.

Michael Grieves sketched the idea in 2002, per Stanford's historical account, and a 2012 NASA paper formalized it for flight vehicles. The term has since drifted into conversations about AI copies of people, where it describes something different.

This guide covers what a digital twin is, how it works, and why a conversational AI replica of a person needs its own name.

What a digital twin is

A digital twin is a virtual model of one specific physical object, system, or process that stays connected to its real-world counterpart and updates as that counterpart changes. The ISO 23247-1 standard from the International Organization for Standardization defines it as a "fit-for-purpose digital representation of an observable manufacturing element with synchronization between the element and its digital representation."

Synchronization is the key word. A 3D model of a pump shows the pump as designed, and a one-off simulation estimates how it might behave under a specified load. The digital twin of that same pump receives its actual temperature and vibration readings continuously, so it describes the pump as it is right now. If the data link breaks, the twin becomes a snapshot of a past state.

How a digital twin works

Every digital twin, whatever it models, runs the same loop: observations flow from the physical thing into a virtual model, operators use the model to predict what happens next, and decisions flow back. The wind turbine shows each stage.

• Physical sensors and live data collection: Sensors on the turbine's gearbox, main bearing, and generator capture vibration, temperature, oil pressure, and rotational speed. Those readings stream continuously to an edge device on the tower or to a cloud pipeline, where they're cleaned and time-stamped for the model.  
• Virtual model and real-time synchronization: The telemetry feeds a software model of the turbine that carries its current state, including load on each bearing, gearbox temperature against its rated range, and accumulated fatigue on the blades. An engineer at an operations center hundreds of miles from the site opens the twin and sees what the turbine is doing right now, without climbing the tower or stopping the rotor.  
• Simulation, prediction, and feedback: Operators can ask how gearbox wear progresses at full output through a week of high winds, or whether a bearing that's trending warm will cross its failure threshold before the next service. The answer goes back to the operations team as a decision: pull the service forward, curtail output for a few days, or leave the machine alone.

That loop can be built at different scales, from a single part to a whole workflow. It's also why the term travels so easily into other domains, including conversational AI systems built to represent people, where the promise of a live, responsive model sounds appealing but doesn't quite match how those systems actually work.

Where digital twins operate today

Four industries show current deployments and active research built on the same loop.

• Manufacturing: Automakers trial a new station layout and robot path inside a process twin before physical retooling begins.  
• Renewable energy: Wind and solar operators watch component twins of gearboxes and inverters for early wear signatures.  
• Smart-city infrastructure: City operators feed traffic and power-grid telemetry into a network twin to catch congestion or load imbalance before it cascades.  
• Clinical research: Researchers are working toward patient-specific twins that model how an individual might respond to a drug or dose before it's given. A 2025 NIST report on digital twin security includes "living things" among the entities a twin can represent.

Those clinical twins are the closest the field comes to a human digital twin, and they still follow the turbine pattern: measured data from one body goes in, and predictions come out.

Digital twin vs. simulation vs. Replica

The three terms often get used interchangeably, but each describes a different relationship to the real world.

The separation matters most in medicine, where "digital twin" already carries a technical meaning tied to a synchronized model of one patient's physiology. Applying the same label to a conversational Replica invites the wrong expectation, since a Replica trains once from recorded video and never resyncs with the person it depicts.

Why "digital twin" gets applied to people

People reach for "digital twin" when they mean an AI that looks and sounds like a specific person, partly because the phrase sounds precise and partly because the product category is new enough that common vocabulary hasn't settled.

The borrowed label causes two problems. It implies a live link to the real person that doesn't exist, which can mislead buyers about what the system knows. In healthcare, it collides with patient-specific physiological twins, where the term already carries a technical meaning tied to clinical data.

More precise terms separate the pieces. A Replica supplies appearance, voice, and mannerisms, and a Personified Application Layer (PAL) is the finished conversational system built on it: a real-time application you talk to face-to-face and build a relationship with. This is exactly what Tavus, the human computing company, is building today.

How PALs hold face-to-face conversations

A digital twin watches its counterpart, while a PAL watches the person it's talking to in real time. That live perception is where the engineering happens. Each Tavus PAL runs a closed-loop behavioral stack with four parts.

• Conversational flow: Sparrow-2 continuously models who has the floor and separates backchannels from real interruptions, so the PAL responds quickly without cutting people off. It scores 92.4% end-of-turn recall and 97.4% interruption recall on TurnBench's public dev split.  
• Perception: Raven-1 fuses tone, expression, and gaze into natural-language descriptions of the person's state, so the PAL can respond to hesitation or confusion as it happens.  
• Reasoning: The LLM layer decides what to say and do next, drawing on the Knowledge Base for grounded answers.  
• Facial behavior: Phoenix-4.5 renders behavior across 10+ controllable emotional states, including micro-expressions and attention signals while the other person is still speaking.

Picture Priya, a new pharmaceutical sales rep rehearsing with a PAL that plays a skeptical cardiologist. When the PAL asks about dosing in renal patients, Raven-1 fuses her suddenly clipped tone with the glance down at her notes, catching the mismatch between what she says and how she says it, and Sparrow-2 holds the floor open through her mid-sentence pause. The LLM layer presses the objection once more while Phoenix-4.5 renders a doubtful tilt of the head, then softens it when she finds the answer.

Where PALs are deployed today

Just as digital twins help engineers simulate and predict how physical systems will behave, PALs help teams simulate and shape human interactions in situations where face-to-face conversation drives the outcome. The difference is that a PAL's "sensor input" is the person on the other side of the screen, and its output is a real-time response that adjusts as the conversation unfolds. That makes them a fit anywhere consistent, high-quality face-to-face interaction matters and hiring more people isn't realistic.

• Sales training: Reps rehearse pitches and objection handling with PALs that role-play different buyer personas, from skeptical clinicians to technical evaluators, without booking time with a coach.  
• Patient intake: Clinics use PALs to conduct pre-visit interviews, collecting symptoms and history in a conversational format before a clinician takes over.  
• Candidate screening: Recruiters deploy PALs to run first-round interviews, asking consistent questions and adapting follow-ups based on each candidate's responses.  
• Customer onboarding and education: Companies use PALs to walk users through product setup, complex policies, or guided demos, answering questions as they come up.

Across each of these, the value comes from real-time responsiveness. A PAL that can read hesitation, hold the floor through a pause, and adjust its tone accordingly changes what a face-to-face interaction can accomplish when no human is available on the other side.

Different jobs deserve different names

A digital twin earns its name by staying synchronized with one machine, system, or process, which lets engineers understand its current state without touching it. A Replica and the PAL built on it do a different job, carrying a person's appearance and voice into conversations that respond to whoever is on the other side of the screen. Keeping the terms separate gives engineers, buyers, and clinicians an accurate picture of what each system actually knows.

Tavus is the human computing company, building PALs that see, hear, remember, and respond face-to-face in real time. Teams deploy them across sales training, patient intake, and candidate screening, wherever face-to-face conversation quality changes the outcome and hiring more humans isn't an option.

See it for yourself. Book a demo.