I spend a lot of time thinking about what makes a piece of writing sound like me. There are the subjects I keep returning to, the way a wine question turns into a history question, and the occasional detour that seemed perfectly relevant when I started the paragraph. Put enough of those pieces online and there’s a recognizable pattern. I can understand why a system might learn to imitate it, which makes me curious about how far the resemblance actually goes.
The phrase “digital DNA” is appealing because it suggests that a whole person might be hiding inside the traces they leave behind. A playlist, a search, a photograph, a message. Gather enough fragments and perhaps a second version of us starts to take shape. But the phrase is a metaphor, and several very different technologies can disappear inside it if we let the story move too quickly.
An advertising profile, a synthetic voice and a chatbot built around someone’s documents can each resemble an aspect of a person. They do different jobs, use different information and deserve different questions. I’m interested in all three, especially where the resemblance becomes convincing enough that we start giving it credit for understanding something it may only be predicting.
A Profile Can Be Useful Without Being a Person
A recommendation system can use signals from our activity to estimate what we might choose next. The particular signals and how they’re used depend on the service. That estimate can be useful, irritating, or oddly well timed. It’s still an estimate assembled for a purpose, such as choosing content or advertising, rather than a complete account of the person holding the phone.
The Federal Trade Commission’s 2024 staff report on major social and video services documented extensive collection and sharing of personal information. It included concerns about data from brokers, information about non-users, retention and the use of data in automated systems. Those are substantial privacy issues on their own. They don’t require the additional claim that a company can read every private feeling accurately.
A pattern can also be correct for the wrong reason. A person might research a subject for an article, buy a gift for someone else, or watch a video because they disagree with it. The action is observable while the reason can remain uncertain. I think about that when I consider the range of things I write about. A profile based on the subjects alone would have quite a job working out which ones describe my life and which ones simply caught my interest.
The Difference Between a Trace and a Guess
It helps to separate something you deliberately provided from something a system inferred. A birthday entered into an account and an estimated interest in gardening are different kinds of information. Both can influence what happens on a service, but they carry different opportunities for error. An inference can feel intimate even when it’s a rough category shared with a very large number of other accounts.
Google’s My Ad Center guidance offers a concrete example. It describes account information, activity and categories used to personalize advertising, along with controls for how certain information is used. Those settings provide a view into part of a particular system. They aren’t a display of everything every company might know, and changing ad personalization is different from deleting all records of activity.
I find that distinction helpful because the language around this subject can become enormous very quickly. “The algorithm knows me” might mean that it made a good recommendation, assigned a category, or used a piece of information I supplied. Asking which of those happened makes the question much easier to answer. It also gives me something practical to examine instead of leaving me arguing with an invisible double.
Imitating My Sentences Would Be a Different Project
A language model can be given examples of someone’s writing and asked to produce text with similar characteristics. It might reproduce favored phrases, sentence lengths, recurring topics and a conversational manner. A system could also be given documents to consult when answering questions. Those approaches can produce familiar-sounding material without reproducing the process by which the original writer arrived at a thought.
That difference matters to me as someone who writes. The words I publish are selected pieces of what I was thinking about, revised for an audience and a particular purpose. They leave out a great deal. A model supplied with the finished articles would have examples of the result, but it wouldn’t automatically acquire every experience, doubt or change of mind that happened around them.
There’s also a difference between retrieving a sentence I actually wrote and generating a new one that sounds as if I might have written it. The first can be checked against an archive. The second is a fresh output that needs its own scrutiny. NIST’s generative AI risk profile describes confabulation, where a system produces confident but false content. A familiar voice can make that problem harder to notice.
If a supposed digital version of me confidently explained a memory I never had, its fluency wouldn’t make the memory mine. I’d want a clear distinction between preserved material and newly generated responses. It’s the sort of detail that sounds technical until you consider someone else reading the answer and reasonably assuming I had actually said it.
I’d keep a small external drive for dated copies of my writing and the photographs I want to preserve. Being able to open my own version matters to me, especially when a new passage could sound so familiar.
Researchers Have Tested More Specific Simulations
There is research that goes beyond choosing the next video. In the work described by Stanford HAI’s 2025 policy brief, researchers recruited 1,052 people for approximately two-hour interviews. They paired the interview transcripts with a language model to create agents intended to simulate those participants’ responses. The agents were then evaluated on surveys and experiments.
On the General Social Survey, the agents reproduced participants’ answers 85% as accurately as the participants reproduced their own answers two weeks later. That comparison is easy to lose in a headline. It describes performance relative to the participants’ own consistency on a particular survey, rather than a claim that a model captured 85% of an entire person.
The study used extensive interviews, rather than silently assembling a complete personality from random browsing crumbs. It’s an interesting result for researchers asking how well a simulation can help study attitudes and responses. Predicting someone’s answer in a defined test is a much narrower achievement than reproducing every decision they might make, and narrower questions are often the ones we can actually learn something from.
A Familiar Voice Can Travel Without Its Owner
Voice synthesis creates another kind of resemblance. With suitable recordings, a system can generate new audio that resembles a particular speaker. There are uses someone might deliberately authorize, and there are impersonations they would never agree to. In either case, a voice that sounds familiar can carry words the original speaker never said.
The FTC’s warning about cloned-voice emergency scams explains how audio taken from online content can be used in an impersonation. Its practical advice is straightforward: verify the story by contacting the person through a number you already know. The sound of a familiar voice is a reason to care about a call, but a separate check helps establish who is actually making the request.
That makes consent a very concrete part of the discussion. Permission to hear or share a recording is a different question from permission to generate new speech in the speaker’s voice. I’d want a service to explain what material it needs, what it can create, who can use the result and how access can be withdrawn. The resemblance is the feature, so control over it should be easy to understand.
Memory Makes the Questions More Personal
The idea of an archive that can answer in a familiar manner becomes particularly complicated when the person has died. A saved recording contains something they actually said. An interactive system might combine preserved material with newly generated replies. The experience can be moving, but the distinction between those two kinds of response deserves to remain visible.
Cambridge researchers examined these questions in work on the digital afterlife industry. Their design scenarios considered risks around consent, advertising, unwanted interactions and the ability to end an encounter with a simulation. Those scenarios were ethical thought experiments, rather than a report that every service had already behaved that way. They identify choices a designer should consider before someone’s memories become an interactive product.
I understand the appeal of being able to return to a voice. I also think the person receiving that voice should have a choice about the experience. Someone might treasure a recording and feel very differently about an unexpected message generated in the same voice. There’s room for several reactions, and affection for the original person doesn’t obligate anyone to engage with a simulation.
For me, the important questions would be who agreed to the use, what the system is allowed to invent, and whether it can be stopped. A collection of letters, photographs or audio can already hold an enormous amount of meaning. Adding conversation changes the experience, and that deserves a deliberate choice rather than an automatic upgrade.
Whose Information Is in the Archive?
A personal archive rarely contains only one person. A message includes someone else’s words, a photograph includes other faces, and a story may reveal something its subject never chose to publish. Before uploading a large collection to any service, I’d want to think about the other people represented in it. Having a copy in my possession doesn’t settle every question about its next use.
The earlier version of this article gave a blanket answer about technology companies owning a digital self. The practical questions are more specific. What does a particular service say it can store, use and share? What permission does the uploader have, and what controls exist for the people involved? Those answers depend on the material, the agreement and the applicable rules, so a broad claim about someone owning “you” obscures the decisions that need attention.
I’d also distinguish permission for one task from permission for an expanding set of tasks. Helping organize a private archive, answering questions from it and speaking publicly in someone’s name are very different uses. Each step changes who might encounter the result and what they could reasonably assume about it.
A Few Choices Make the Profile Less Mysterious
I’d start with the services I actually use and the controls they offer. Review account information, connected apps and advertising preferences, then look at what can be downloaded or deleted. The available settings vary, and a change in one service may have a limited scope. It helps to read the description of the control instead of expecting one switch to cover an entire online life.
For a voice, avatar or archive-based assistant, I’d ask more directly about access, reuse, retention and generated content. Who can make it respond? Can its output be clearly labeled? Can the underlying material be exported, and what happens when the account closes? A service that represents a person should be able to explain those ordinary questions without requiring the customer to translate a marketing promise.
The original article suggested making misleading searches to confuse the system. I’d put the effort into understandable controls and thoughtful sharing instead. A deliberately strange browsing history gives you uncertain results, while reviewing the information and permissions you’ve actually provided gives you a clearer place to begin.
I’d Like the Resemblance to Stay in Perspective
What interests me most is how easily a recognizable detail can stand in for the whole person. A turn of phrase, a laugh, a well-chosen recommendation. Recognition is powerful, and it can be useful, comforting or misleading depending on what we assume follows from it. Matching those details doesn’t establish that a system contains the person’s continuing awareness or their complete life.
I’m still curious about what these tools might help us preserve and make. An organized archive or a clearly identified assistant could be useful, especially when its purpose and limits are easy to see. But I’d like the person represented by the system to keep a say in how the resemblance is used, and the person encountering it to understand what they’re hearing.
A model built from my writing might become very convincing about wine, gardens and whatever odd science subject distracted me that week. It would probably need a better filing system than the one in my head, which is admittedly a low bar. I’m interested in the possibilites, but I want to keep asking which parts were recorded, which were inferred and which were invented. Those questions leave room for the technology without letting its version of a person become the only one we listen to.
