I just learned about digital twins when I was applying for a company that does simulation of consumer behavior. For this post, I cloned myself from five of my most recent blog posts and gave my digital twin (GPT 5.6 Luna) a chance to write to see how it compares to my normal writing.
I have been reading about digital twins lately, and the phrase "digital human twin" sounds like the sort of thing that arrives with a glossy landing page and a suspiciously confident demo. The basic idea is more interesting than the marketing, though.
A digital human twin is a generative model that tries to act as a proxy for a particular person. Feed it enough information about someone's demographics, preferences, interviews, survey answers, and behavior, and it can answer questions or predict how that person might respond to a new situation.
That is different from a generic synthetic user. A synthetic user represents a group, such as "healthcare workers in Latin America." A digital human twin is supposed to represent one actual person. The boundary is not perfectly clean. Give a model only a few broad attributes and it behaves like a synthetic persona. Give it a long history of one person's responses and it moves toward a twin.
The distinction matters because "one person" creates an illusion of intimacy. A model that knows my job, my writing, my projects, and my opinions may sound like me. That does not mean it understands me. It means it has found patterns in the information I made available.
From Mirror to Agent
The Nature Computational Science article, The evolution of digital twins from reactive to agentic systems, describes a wider change in digital twins. They are moving from systems that observe a thing and report its current state toward systems that combine models and data with human interaction. The long-term direction is self-learning and autonomous behavior, with artificial intelligence and computational reasoning involved in decisions and actions.
The word "agentic" is doing a lot of work there. A reactive twin might tell an engineer that a machine is overheating. A predictive twin might estimate when it will fail. An agentic twin could reason about possible interventions, recommend one, and perhaps carry it out.
A digital human twin applies that progression to a person. It might start by summarizing what I have said before. Then it might predict what I would choose. Eventually, it might interact with other systems on my behalf: answering a survey, filtering opportunities, negotiating a calendar, or representing my preferences in a product test.
This is where I become cautious. An agent acting on behalf of a machine is already a difficult engineering problem. An agent acting on behalf of a human has an extra problem: the target keeps changing. Machines do not have a bad week, discover a new interest, change their mind after one conversation, or deliberately answer a question differently because the question is badly framed.
People are not static datasets. They are stateful, inconsistent, and occasionally irrational. Those are not bugs waiting to be cleaned out.
The Useful Part
The Nielsen Norman Group article describes several practical uses. A twin could fill in skipped survey answers, shorten a survey by inferring some responses, help study people who are difficult to recruit repeatedly, or predict reactions to an interface. Many of these uses are sensible because they treat the twin as an estimate that helps researchers decide what to investigate next.
That is the right level of ambition, in my opinion.
If a model has access to a detailed interview and can predict a participant's answer to a related survey question, it may reduce repetitive work. If it can identify a likely usability problem before a design reaches production, it can give a researcher a useful hypothesis. If I am trying to understand how a product might affect several user groups, simulated responses can help me choose which questions deserve real research.
The important word is "hypothesis." The twin can tell me where to look. It cannot give me permission to stop looking.
NN/g makes the same practical point more directly: current simulated humans are useful as a desk-research tool, but they are not a replacement for talking to actual customers. I trust that conclusion because it matches my experience with AI coding agents. An agent can produce a plausible implementation and still miss the one environmental detail that breaks the application in production. Plausibility is cheap. Contact with reality is the expensive part.
Four Things We Can Learn
1. A prediction is not an explanation
If a twin predicts that I will reject a feature, I still do not know why. Perhaps the feature conflicts with a stated preference. Perhaps I dislike the wording. Perhaps I am tired, distracted, or reacting to something unrelated. A prediction can be accurate for the wrong reason and inaccurate for a reason the model never represented.
This is familiar in machine learning. Correlation can be useful without being an explanation. A human twin should make us more careful about that distinction, not less careful.
2. Personal data is a partial and historical view
The twin is built from what was collected: interviews, survey responses, logs, purchases, and other context. That data is selective. It records what someone did in a particular environment, under particular incentives, using particular interfaces.
When I write a blog post, I am creating evidence about how I think. I am not publishing my whole mind. A model trained on my posts could imitate my tone and repeat my recurring arguments. It would still lack the private context, the things I have not learned yet, and the reasons I sometimes change my position.
That limitation should be visible whenever a twin is used. Otherwise, an incomplete record quietly becomes an authoritative identity.
3. The human must remain in the loop where judgment matters
The more autonomous these systems become, the more tempting it is to hand them decisions because they are fast and available. That is backwards. Automation is most valuable when the task is repetitive and the cost of being wrong is contained.
I would use a twin to prioritize interview questions or flag a possible design issue. I would not use it as the final voice of a patient, employee, customer, or applicant without a clear reason, informed consent, and a way for the real person to correct it.
Information access does not create expertise. A digital copy of someone's previous opinions does not acquire that person's judgment. A model of a doctor is not a doctor, and a model of me is not me. This should be obvious, but technology has a strange ability to make obvious things billable.
4. Trust depends on infrastructure, not only model quality
The Nature article emphasizes interoperability and standardization as conditions for digital twins to work across sectors. That sounds less exciting than autonomous reasoning, but it is probably more important. If data formats, identity rules, permissions, and interfaces do not work together, a brilliant twin remains a clever isolated demo.
For human twins, the infrastructure also includes consent, data deletion, provenance, access control, and a record of what the twin is allowed to do. The owner should be able to inspect the source information, correct mistakes, restrict uses, and revoke the twin. A model that cannot be corrected is not a faithful representation. It is a permanent guess with an API.
My Position
I am cautiously positive about digital human twins as research instruments. They can make expensive or repetitive exploration cheaper, expose assumptions, and help researchers spend real human time on the questions that matter. They may also help individuals reflect on their own patterns, provided the system shows uncertainty instead of pretending to deliver a psychological verdict.
I am strongly against treating them as replacements for people. The closer a system gets to speaking and acting in my name, the higher the standard should be for consent, transparency, reversibility, and human review. A fluent answer is not evidence that the twin is correct. A confident agent is not evidence that it has earned authority.
The most useful lesson from the digital human twin is therefore not that we can finally simulate a person. It is that human behavior is difficult to model even when we have a lot of data. The model teaches us about patterns, gaps, assumptions, and possible futures. Reality still gets the final vote.
References
- The evolution of digital twins from reactive to agentic systems, Omer San, Adil Rasheed, Eda Bozdemir et al., Nature Computational Science, 2026.
- Digital Twins: Simulating Humans with Generative AI, Raluca Budiu, Nielsen Norman Group, 2025.