Loading
A digital twin is a virtual model of a physical asset, process, or system, a machine, a factory line, a power grid, even the planet's climate, that is continuously synchronized with real-world sensor and operational data so it mirrors what the physical counterpart is actually doing right now, not just how it was originally designed. The US National Institute of Standards and Technology (NIST) describes manufacturing digital twins as "synchronized virtual models that help manufacturers represent, diagnose, predict, and optimize their operations," a concept formalized internationally through ISO 23247, the Digital Twin Framework for Manufacturing, published in 2021. What separates a digital twin from an ordinary simulation is that live data link: as the physical system changes, degrades, or fails, the twin is meant to reflect that in near real time rather than running once against assumed inputs. Digital twins matter for AI governance because they increasingly embed machine-learning components, predictive maintenance models, anomaly detectors, generative-AI interfaces, and because their outputs are used to trigger real-world action, from automated shutdowns to national climate-policy decisions. That combination raises concrete governance questions: how do you know the model is still faithful to reality, and who, or what, holds the authority to act on what it says.
Run the free AI Health CheckDigital Twin, a virtual representation of a specific physical asset, process, or system that is continuously synchronized with real-world data so it mirrors the current state and behavior of its physical counterpart, as opposed to a static, one-off simulation.
Think of it as a live, data-driven mirror of one particular machine, building, or process, not a generic model, and not a simulation you run once and file away. NIST defines manufacturing digital twins as 'synchronized virtual models that help manufacturers represent, diagnose, predict, and optimize their operations,' formalized internationally through ISO 23247. The governance stakes appear when organizations let a twin's outputs trigger real-world action automatically (a 'closed-loop' twin) or inform high-stakes decisions: if the underlying model has drifted out of sync with reality, or if AI components inside the twin haven't been validated, errors can propagate into physical or policy consequences before a human ever reviews them, which is exactly why NIST is now developing a verification-and-validation standard (ISO 23247 Part 7) specifically for digital twins.
Source: NIST, Digital Twins for Advanced Manufacturing program; ISO 23247-1:2021, Digital twin framework for manufacturing
A traditional simulation is typically run once, or occasionally, against assumed or historical inputs to answer a specific question. A digital twin is different in kind: it is instance-specific (modeling one particular machine, building, or process rather than a generic type) and is kept current through an ongoing feed of sensor and operational data across the asset's lifecycle. Microsoft's Azure Digital Twins platform, for example, represents each physical entity as a 'twin' connected to live telemetry from IoT-managed devices, with an event system that keeps the model's state current as conditions change, the twin is queried for real-time insight, not just run as a one-off report.
This live connection is often called the 'digital thread', the linkage of data and models across design, production, and maintenance stages of an asset's life. NIST's own manufacturing digital-twin program is explicitly building on this idea, contributing to a new part of the ISO 23247 standard (Part 5) focused on digital threading so that a twin built during design can keep being useful and accurate once the physical system is actually built, operated, and maintained.
ISO 23247, 'Automation systems and integration, Digital twin framework for manufacturing,' is the primary international standard in this space, published in 2021. It lays out reference architecture and principles for how a manufacturing digital twin should be structured and how data should move between the physical system and its virtual counterpart. NIST is actively contributing new parts to extend it, including a guideline standard on 'Verification, Validation and Uncertainty Quantification (VVUQ) Framework for Digital Twins in Manufacturing' (Part 7). This is the piece that speaks most directly to governance: it is meant to establish how an organization tests and demonstrates that a twin's model is still accurate enough to be trusted before its outputs are used to make a decision.
That work exists because the gap is real. NIST itself acknowledges that 'building digital twins correctly is challenging, especially for small and medium-sized enterprises,' and identifies the lack of standards and common validation procedures as a major barrier to using digital twins with confidence. In other words, most digital twins deployed today were not built against a shared, external trust baseline, validation has largely been left to each vendor or operator.
Model fidelity and drift
A twin's virtual model can silently diverge from the real asset as physical conditions change, sensors degrade, or the equipment is modified without the twin being updated to match, the same underlying problem as model drift in any predictive system, but with a physical asset on the other end.
Decision-authority gap in closed-loop twins
Some digital twin deployments are 'closed-loop', the twin's output feeds directly back into automated control of the physical system rather than to a human for review. If the model is wrong, an error becomes a real-world action with no check in between. NIST's own human/machine-teaming research treats this as an open problem, deliberately emphasizing human-centered collaboration over autonomous hand-off.
Accountability when a twin-informed decision goes wrong
When an AI-augmented digital twin recommends (or triggers) a shutdown, a maintenance call, or a policy decision that turns out to be based on a stale or flawed model, existing safety and liability frameworks generally still apply to the operating organization, but few sectors yet have a twin-specific standard of care to point to.
Cybersecurity and OT attack surface
A twin that is bidirectionally synced to a physical control system, receiving live telemetry and, in closed-loop setups, sending commands back, turns the twin's data pipeline into a bridge into operational technology (OT). Compromising the twin's inputs or model can compromise the plant it represents.
Sensitive operational data exposure
Continuous synchronization means a digital twin often carries a live, detailed record of proprietary processes, safety-relevant data, and (for AI-enhanced twins) production data used to train or fine-tune models, data that wouldn't otherwise leave the facility in that form.
Immature standards base
As NIST notes, the lack of common validation standards is itself a governance risk: without a shared VVUQ baseline (still under development as ISO 23247 Part 7), organizations have limited external means to compare how rigorously one twin was validated against another.
Semiconductor manufacturing
In May 2024, CHIPS for America announced roughly $285 million to establish a Manufacturing USA institute dedicated to developing and validating digital twins for chip fabrication, advanced packaging, assembly, and test, defining digital twins as 'virtual models that mimic the structure, context and behavior of a physical counterpart.'
Smart infrastructure and IoT
Platforms like Microsoft Azure Digital Twins model buildings, factories, energy networks, railways, and even entire cities as connected 'twin graphs' driven by live IoT data, used to monitor operations and optimize performance.
Climate modeling and public policy
The European Commission's Destination Earth program is building high-accuracy digital twins of the Earth system (developed by ECMWF) to simulate climate change, extreme weather, and disaster risk, explicitly to support evidence-based EU and national policymaking.
General manufacturing operations
NIST's Digital Twins for Advanced Manufacturing program supports twins that help manufacturers represent, diagnose, predict, and optimize production, aimed particularly at helping small and medium-sized manufacturers close the efficiency gap with larger firms.
What is a digital twin in simple terms?
It's a virtual model of a specific physical thing, a machine, building, process, or system, that stays continuously updated with real data from that physical thing, so the model reflects its actual current state rather than just its original design.
How is a digital twin different from a computer simulation?
A simulation is usually run once or occasionally against assumed inputs to answer a question. A digital twin is instance-specific and kept synchronized with live sensor and operational data across the physical asset's lifecycle, so it changes as the real thing changes.
What does ISO 23247 actually standardize?
ISO 23247, 'Digital twin framework for manufacturing,' published in 2021, sets out reference architecture and general principles for structuring a manufacturing digital twin and exchanging data between the physical system and its virtual counterpart. NIST is contributing further parts, including a Part 7 guideline on verifying and validating a twin's model before it's relied on.
Is a digital twin a form of AI?
Not by itself, a digital twin is fundamentally a modeling and data-synchronization architecture. But many modern digital twins embed AI/ML components, such as predictive maintenance models, anomaly detectors, or generative-AI interfaces for interacting with the twin, which is why AI governance frameworks increasingly apply to them.
Can a digital twin control real-world equipment automatically?
Yes, in 'closed-loop' configurations the twin's output can feed directly back into automated control of the physical system. NIST's own research on human/machine teaming in manufacturing treats the resulting decision-authority question as unresolved, favoring human-centered collaboration over full autonomous hand-off.
Who is accountable if a digital twin's model is wrong and causes harm?
There is no single global standard yet assigning this responsibility specifically to digital twins. In practice, existing safety, liability, and risk-management obligations fall on the organization operating the twin, the same way they would for any other automated decision system it deploys.
Last reviewed July 2026
This page is general information about What Is a Digital Twin?, not legal, regulatory, or professional advice, and does not capture every nuance or exception. Requirements change and can be fact-specific. Always verify against primary sources and your own qualified legal counsel before relying on it.