Digital Twin Divergence

Sep 29

A digital twin is a model kept in step with the thing it models. Cut the link, and it starts writing its own history. For most twins that is just drift. For a twin of a mind, it may be something else.

This post is an essay; its later sections are speculative by design.

The digital twin concept is usually traced to Michael Grieves, who presented it in the context of product lifecycle management in the early 2000s, and to NASA, which adopted the term and had long practiced the idea — most famously in the ground simulators used to work the Apollo 13 problem in 1970. A digital twin is a virtual model of a physical system, updated continuously from sensor data, used to monitor, predict and test interventions safely. Twins now model jet engines, factories, power grids, individual patients' hearts and, in the European Union's Destination Earth initiative, the planet's climate system.

What makes something a twin rather than a simulation is the coupling. Remove the live data, and it becomes something else.

How twins stay synchronized

Any model of a complex system diverges from reality over time. Edward Lorenz demonstrated in 1963 that deterministic systems can be exquisitely sensitive to initial conditions, so small errors grow exponentially. Weather models run without correction lose predictive skill within days.

Twins manage this through data assimilation: continuously correcting the model's state against observations, using methods descended from the Kalman filter. The model predicts; the sensors report; the discrepancy updates the model. This is structurally the same loop described in the Machine Dreaming post — a generative model disciplined by error signals from the world.

Divergence, then, has a precise meaning. It is what happens when assimilation stops, and the model's trajectory is determined by its own dynamics instead of by the system it was built to track.

When divergence is just error

For a twin of a turbine or a climate system, divergence produces nothing but a wrong answer. The decoupled model continues computing states of an imaginary turbine that increasingly fail to match the real one. There is a self-contained trajectory, but not a subjective one, because nothing in a fluid dynamics model is a candidate for a subject. Divergence here is a maintenance problem.

The question in the title becomes interesting only when the system being twinned is itself one whose trajectory might be subjective.

Twins of minds

Two developments make that more than hypothetical. The first is commercial: products marketed as personal digital twins, trained on a person's messages, writing and recordings to respond in their style — for productivity, for customer service, and in some cases as memorials after death. These are behavioral models. They reproduce outputs, not the cognitive processes behind them.

The second is long-range: whole brain emulation, the proposal to scan a brain in sufficient detail and simulate it. Anders Sandberg and Nick Bostrom's 2008 roadmap laid out what it would require. Robin Hanson's The Age of Em (2016) explored the economics of a world in which such emulations exist and are copied freely. Nothing close to this is technically possible now.

Analysis. The two cases produce very different divergence. A behavioral twin that is cut off from new data from its source does not develop an independent life so much as freeze: it keeps producing the person as they were, and its responses to new situations are extrapolations from a fixed style. A structural emulation, if it existed and if functionalism were right, would diverge the way identical twins do — the same starting point, followed by different experiences producing different people. On Parfit's branch-line analysis from the previous post, the original and the emulation would both be continuations of one person at the moment of copying and two people from then on.

The point of no return

The title asks when a twin's trajectory becomes independent. For ordinary twins, the answer is technical: once accumulated error exceeds what assimilation can correct, the twin no longer tracks the original. For twins of minds, there are two thresholds, and they come apart.

The first is informational: the point at which the twin has experienced enough the original has not that it no longer predicts the original well. That can be measured.

The second is subjective: the point at which there is something it is like to be the twin, making its trajectory not just independent but a life. That cannot be measured by any method the series has examined. If it is crossed at all, it may be crossed at the moment the emulation starts running, not at some later point of divergence — in which case divergence does not create a subject, it only makes an existing one distinct.

The practical upshot is modest but real. Anyone building systems designed to model individual people should be clear which kind they are building. A behavioral twin is a tool that imitates a person. An emulation, if one is ever built, would raise every question in this batch at once.

Related reading

→ The Discontinuity of Rebooting

→ Machine Dreaming and Perceptual Hallucination

→ The Threat of Machine Solipsism

Social: A digital twin stays true only as long as it's fed data from the real thing. Cut the feed, and it starts writing its own history. For a twin of a min

Digital Twin Divergence

The Threat of Machine Solipsism

The Discontinuity of Rebooting

Your Site Title

email@example.com