"Let Me Introduce You to My Avatar": Who Speaks for You After You Build an AI?
As AI doubles evolve from static chatbots into agent supported digital personas, the challenge shifts from preserving knowledge to determining who controls, updates, and ultimately speaks for the artificial version of a human being.
Part I: The Rise of the AI Double
A recent article in The New York Times described a growing practice among executives, professors, consultants, and entrepreneurs: the creation of artificial intelligence versions of themselves. Trained on years of interviews, presentations, books, blog posts, and correspondence, these systems answer questions, provide advice, and interact with employees, students, or clients on behalf of their creators.
The attraction is easy to understand. Expertise does not scale. A professor can hold only so many office hours. A chief executive can meet with only so many employees. A consultant can advise only so many clients. An AI double offers a way to distribute accumulated knowledge beyond the limits of time and geography.
In some respects, the concept represents the latest stage in a much older tradition. Books preserved ideas. Recorded lectures preserved voices. Websites and blogs preserved arguments. Artificial intelligence adds a conversational layer on top of those archives. Rather than searching through documents, users can ask questions directly and receive customized responses. The result feels less like consulting a library and more like engaging in dialogue with an expert.
The New York Times article highlights how rapidly the practice is spreading. Harvard Business School has incorporated AI versions of faculty into entrepreneurship programs. Executives are experimenting with assistants trained on their communications and decision making frameworks. Consultants increasingly direct prospective clients toward AI doubles capable of answering routine questions and providing preliminary guidance. Knowledge that once required direct access to an individual can now be delivered through software.
Yet the article also revealed an important limitation. Students still preferred meeting their professors. Clients still wanted access to the actual consultant. Employees remained interested in hearing directly from leaders rather than their digital substitutes. Information could be replicated. Judgment remained attached to the person.
That distinction may prove more important than the technology itself. Once an AI carries a person's name, voice, and identity, users naturally assume it speaks with the authority of the individual it represents. Whether that assumption is justified remains an open question.
Part II: ElmerGPT and the Problem of the Frozen Self
The question became more tangible when I created a personal experiment that I jokingly called ElmerGPT
Like many custom GPT projects, the goal was straightforward. Could an AI system trained on years of blog posts, research papers, presentations, enrollment forecasts, governance documents, and professional writing answer questions in ways that resembled my own thinking? The answer was largely yes. The resulting system could explain many of my views on higher education, artificial intelligence, liberal arts education, enrollment management, and data governance. In some cases, it could summarize those views more quickly and more clearly than I could. At first glance, the result appeared remarkable. Upon closer examination, however, it exposed a deeper problem.
Around the same time, reports emerged involving an AI chatbot associated with rapper Flo Rida. Users interacted with the system because they assumed it possessed authoritative knowledge about the artist. Yet conversational fluency created an illusion of expertise that the technology could not always support. The chatbot sounded informed because it sounded like the person it represented. Those are not the same thing.
The episode highlighted a growing risk for all personal AI systems. Once a chatbot adopts a person's name, voice, or reputation, users naturally begin treating its answers as authoritative. The problem is that authority does not automatically transfer from a human being to a model trained on information about that human being. An AI can confidently summarize a person's past statements without possessing reliable knowledge of their current actions, intentions, or decisions.
The same challenge applies to every personal AI model. An AI trained on my writing does not represent who I am. It represents the record I have left behind. It captures published arguments, completed projects, and previously expressed opinions. Human beings, however, continue learning after publication. We revise assumptions. We change our minds. We discover mistakes. We encounter new evidence. The person continues evolving even as the archive remains fixed.
An AI double therefore begins aging the moment it is created.
Without continuous updating, ElmerGPT would gradually drift away from the person it was intended to represent. The system might continue defending positions I no longer hold. It might explain policy frameworks that have since changed. It might describe strategies that no longer make sense under current circumstances. Over time, the gap between the living individual and the digital representation would continue to widen.
That realization suggests a different way of understanding AI doubles. Rather than viewing them as digital twins, we may be better served by viewing them as intellectual portraits.
A portrait can capture something meaningful about its subject. It can reveal habits of thought, values, interests, and perspective. Historians often learn a great deal from letters, memoirs, recorded interviews, and personal archives. None of those artifacts, however, are mistaken for the individuals who created them.
Artificial intelligence introduces an unusual twist. Unlike a book or a portrait hanging on a wall, the archive can answer questions. It can participate in conversations. It can generate new combinations of old ideas. As a result, the distinction between stored knowledge and living identity becomes harder to see.
Universities may eventually preserve AI versions of retired professors. Companies may maintain conversational archives of former executives. Future generations could interact with digital representations much as historians consult letters and memoirs today. The archive would no longer sit silently on a shelf. It would answer back.
For all the excitement surrounding AI doubles, that possibility may become their most enduring contribution. Their greatest value may not lie in replacing human expertise but in preserving access to it. Future generations may gain the ability to converse with the accumulated knowledge of the past in ways previous generations could scarcely imagine.
The challenge is remembering what such systems actually are.
An AI twin can preserve knowledge. It can imitate a voice. It can reproduce patterns of reasoning. What it cannot do is remain a living participant in the process of becoming. Human identity is not simply a collection of stored information. It is an ongoing act of learning, revising, forgetting, and changing.
An AI may tell us what someone once thought. Determining whether it still speaks for them remains a question that only humans can answer.
A student consults a digital doppelgänger while the original remains behind a closed door. The archive can answer questions. Determining whether it still speaks for its creator remains a human judgment. AI generated image, 2026.
Part III: From AI Twins to AI Ecosystems
Yet the story may not end with AI doubles. The next generation of systems could prove far more consequential.
A custom GPT is fundamentally an archive. It reflects a collection of documents, conversations, and instructions assembled at a particular moment in time. Even when the system performs well, it remains constrained by the knowledge available when it was created. Human beings continue learning. The archive does not.
Increasingly, developers are attempting to solve this problem through agents. Rather than simply answering questions, agents perform tasks. They search for information, monitor changes, summarize developments, retrieve documents, and update knowledge repositories. In effect, they function less like experts and more like researchers, librarians, and assistants.
Viewed through that lens, a future ElmerGPT might not exist as a standalone system. Instead, it could sit atop an ecosystem of specialized agents. One agent might monitor changes in federal higher education policy. Another could track enrollment trends. A third could index research papers, conference presentations, and blog posts. Additional agents might monitor institutional data, update knowledge repositories, and organize new information as it becomes available.
Such a system would address many of the weaknesses associated with static AI doubles. Rather than relying solely on historical material, the model could draw upon continuously updated knowledge. The result would be less a frozen snapshot and more a living archive.
Yet the improvement introduces a new question. If agents continuously update an AI model's knowledge, who determines what should be added, revised, or removed? At what point does the system cease to represent a historical individual and begin evolving into something else?
The challenge shifts from preserving knowledge to governing it.
A future ElmerGPT might know far more than the person who originally created it. Agents could ingest thousands of articles, reports, and documents that no individual could realistically read. The system could maintain awareness of developments across dozens of domains simultaneously. Its knowledge might become broader, faster, and more current than that of its human counterpart.
Yet knowledge alone does not resolve the problem of identity.
The critical decisions would still involve judgment. Which sources are trustworthy? Which conclusions deserve acceptance? Which assumptions should be challenged? Which values should guide interpretation? Those questions remain fundamentally human.
Perhaps the most important lesson of AI doubles is that the future will not be defined by a single digital twin. More likely, it will be shaped by a partnership between archives, agents, and human judgment. The archive preserves what we know. Agents help us discover what is new. Human beings decide what matters.
Books preserved knowledge. Libraries organized it. Universities transmitted it. AI systems may become the next institution in that long tradition. Their success will depend not on how well they imitate human beings, but on how effectively they help human beings exercise judgment.
That distinction returns us to the question posed at the beginning of this essay. An AI may preserve your writings, imitate your voice, and even learn to update itself through networks of agents. Yet every such system ultimately depends on a human decision about what should be believed, remembered, and carried forward.
The question: Who speaks for you after you build an AI?
Further Reading
- Sarah Kessler. “Talk to My A.I. Twin: Busy Executives Have a New Productivity Hack.” The New York Times, June 6, 2026.
- Sherry Turkle. "Alone Together: Why We Expect More from Technology and Less from Each Other." Basic Books, 2011.
- Vannevar Bush. “As We May Think.” The Atlantic, July 1945.