No One Is Training a Model on You. What They Are Doing Instead Is Better.

Three components do three different jobs, and the useful question is which one holds your part.

• A personal AI trained on your own data is never trained from scratch on one person. Every commercial product in this category runs on a general-purpose base model trained on public and licensed data. That model is the same one everybody else is using, and you never contributed to the training.

• Your material sits in a private retrieval layer. Your recorded answers are stored as individual passages, indexed, and pulled back at the moment a question is asked rather than baked into weights.

• Your voice is a third, separate model. An acoustic model built from your own recordings, used to synthesise answers and nothing else.

• Two of the three layers are yours to delete. The retrieval layer comes out in full, including the index. So do the voice model and the source recordings behind that model. The base model cannot be deleted and holds nothing of yours in the first place.

• Retrieval is what makes an answer traceable. A retrieved passage has a date and an address, so a wrong memory can be corrected the same day and the correction takes effect on the next question, with no retraining.

Training is the wrong tool for this particular job, and the reason is not the invoice.

OpenAI's own fine-tuning documentation says plainly what the technique is for: classification, "generating content in a specific format", "correcting instruction-following failures", holding a consistent response pattern. All of that is teaching a model how to behave. None of it is teaching a model what is true about one specific human being.

A fine-tuned fact has no address. You cannot point at the fact. You cannot show your daughter where an answer came from. You cannot correct one sentence without another training run, and you cannot prove afterwards that anything was removed.

A retrieved fact has an address. Ask about your father and the system pulls the particular passages where you talked about your father, and those passages can be displayed, edited, dated or destroyed. That is why your knowledge belongs in a store a person can read rather than in weights nobody can read, whatever a training run happens to cost this year.

The voice layer has been cheap for years. The VALL-E paper from Microsoft researchers, posted in January 2023, demonstrated high-quality personalised speech from "only a 3-second enrolled recording of an unseen speaker". Three seconds. Audio of you is abundant in 2026, and nothing about the technology is scarce.

What is scarce is audio with a dated consent record attached, held by a company that will synthesise the voice only for people you named. That scarcity is the whole of the protection, which is why the practical steps for keeping your voice out of a cloning pipeline matter more than the state of the art does, and why who legally owns a clone of your voice is worth settling before you record with anybody. If you are still deciding whether to record at all, the case for and against voice cloning is the page for that decision. If you would rather hand somebody the gentler version of this argument, what a personal AI is covers the same ground without the architecture, and what a digital twin actually refers to untangles the other term people use for the same idea.

A context window is not storage. A context window is the input to a single forward pass, refilled from scratch on every turn, and everything inside competes for the model's attention by position as much as by relevance.

Liu and colleagues measured the effect and published the result as "Lost in the Middle: How Language Models Use Long Contexts" (Transactions of the ACL, 2023). Across multi-document question answering and key-value retrieval, performance was highest when the relevant information sat at the beginning or the end of the input, and degraded significantly when the model had to reach into the middle of a long context. Their own phrasing is that current language models "do not robustly make use of information in long input contexts". They found the same pattern in models built specifically for long contexts.

That is the thing you felt around the folder, described properly.

Consumer memory features solve a smaller problem than the word memory implies. They persist a handful of salient facts across sessions so you stop retyping your job title. What they are not is a structured record of how you decide, organised by subject, reviewable as a whole. They were built to improve a product experience rather than to survive you, and that second part is literal rather than rhetorical: what actually happens to your saved ChatGPT memories when you die has a specific and disappointing answer, and the same answer covers the memory your Claude conversations accumulate.

The architecture that works is dull. Break what you said into passages. Index the passages. At answer time, retrieve only the ones that bear on the question in front of you. Your record can then be as large as your life, because nothing enters the prompt except the relevant part. And because every retrieved passage is a real thing you said on a real date, an answer can carry its own provenance instead of asking to be trusted. If you already keep notes, what changes when your own notes become a corpus that answers back is this argument applied to material you have already written.

Joon Sung Park and colleagues, including Percy Liang, Robb Willer and Michael S. Bernstein, first posted work on building LLM agents from self-reports in November 2024 and substantially revised it in April 2026 (arXiv:2411.10109). The figures here come from the revised version. They recruited a diverse national sample of 1,052 Americans and built agents three ways: from two-hour semi-structured interviews using the American Voices Project schedule, from structured surveys including General Social Survey items and the Big Five inventory, and from the two combined.

On held-out General Social Survey items, interview-only agents reached 83% of the participants' own two-week test-retest consistency. Survey-only agents reached 82%. Combined agents reached 86%. Agents built from demographics alone reached 74%.

The ceiling in that study is the fairest one anybody has proposed. Not perfection. How consistently the actual human agreed with themselves two weeks apart.

Two conclusions fall out of those four numbers. A two-hour conversation carried roughly the same predictive weight as a battery of validated psychometric instruments, and combining both added very little, which the authors read as evidence that predictive gains asymptote once a model has seen enough evidence within a domain. And elicited self-report beat demographic inference by a wide margin, which is the formal version of what you already suspected about being profiled. Elicitation is the whole trick, and why a journal that asks you questions produces better material than a blank page is the same finding in a domestic setting.

Now the archive you were going to point the model at. Your email is a record of your obligations. Your photo library records where you were standing. Your Slack history records the register you use at work and nothing whatever about why you took the job. All of it is artefact, the output of a life rather than an account of how the life was lived. Almost none of it contains the sentence "I chose that because", and the reasoning is the transferable part.

A model with your calendar can tell you what you did on 4 March. A model that knows why you left that job in 2011 can answer a question nobody has asked you yet.

There is a legal argument sitting under the practical one. On 20 December 2024 the Italian data protection authority fined OpenAI EUR 15 million, finding among other things that personal data had been processed to train ChatGPT without an appropriate legal basis having first been identified, alongside breaches of transparency obligations. Scraped personal data carries an unresolved lawfulness question. Answers you gave deliberately, to a specific question, for a stated purpose, sit on firmer ground. Consent is the difference between a dataset that can be defended and one that has to be argued about.

Storing what you said is trivial. Knowing what you never said is the difficult half, because an ungrounded region is exactly where a general model's priors move in and produce something plausible, warm and generic.

So the interview is organised rather than open. Eleven dimensions. Identity and core beliefs. Values and principles. Relationships and family. Life events and stories. Work and contribution. Health and wellbeing. Adversity and growth. Joy and delight. Legacy and future messages. Estate and bequests. Family instructions.

The list is a coverage map rather than a taxonomy for its own sake. The list exists so a system can separate a subject you covered thoroughly from a subject you touched once from a subject you never raised at all. A hole in your own record is invisible from inside a conversation. A hole shows up only when something is counting.

Coverage decides the failure mode. Where a region is well covered, the answers come out of your own words. Where a region is thin, a grounded system has nothing to retrieve and stays with what is actually present instead of manufacturing a specific fact to cover the silence. An ungrounded system fills the hole with a confident, average-sounding paragraph, and that is the failure that costs a family something real. Whether a system also announces its own thin regions out loud is a separate design decision from whether the system invents, and any provider should be made to answer the two questions separately.

If eleven headings read as intimidating rather than useful, what you would actually say to an AI of yourself on the first day is the smaller place to start, and the practical knowledge only you know how to do is the region people underestimate most reliably.

Start building your legacy Free build, 50 memories, no card required.

From 2 August 2026, Article 50 of the EU AI Act (Regulation (EU) 2024/1689) requires providers of AI systems that generate synthetic audio, image, video or text to mark those outputs in a machine-readable format and make them detectable as artificially generated or manipulated, and requires anyone deploying a deepfake to disclose that the content is artificially generated or manipulated. Article 50(1) requires that people be told they are interacting with an AI system unless that much is obvious to a reasonably well-informed person. Article 50(5) requires the disclosure to be clear and distinguishable, at the latest at the time of the first interaction or exposure. Under Article 99(4)(g), breaches attract administrative fines of up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher, reduced for SMEs to the lower of the two.

One correction to the headlines, checked 29 July 2026. Regulation (EU) 2026/1744, the Digital Omnibus on AI, was adopted on 8 July 2026, published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. The Omnibus deferred the Annex III high-risk obligations to 2 December 2027 and 2 August 2028, and gave providers of generative systems already on the market until 2 December 2026 to meet the Article 50(2) marking requirement. Article 50 itself was not postponed. If you read somewhere that the AI Act had been delayed, that is the delay being described. The plain-language walkthrough of Article 50 sets out which duties land on a provider and which land on you as a deployer.

For a build like the one described here, the duties are concrete rather than abstract. A synthesised answer in your voice is synthetic audio, so the machine-readable marking duty in Article 50(2) applies. Anybody talking to your build is a natural person interacting with an AI system, so the Article 50(1) notification duty applies too. Both bite from 2 August 2026, which is days away as this page goes up, so treat any provider's claim to be compliant already as something to verify rather than accept, this one included. Marking is what keeps a consented, disclosed replica legally distinguishable from an unconsented one, so marking is worth having rather than routing around.

Then the limit of the law. Article 50 regulates the honesty of the output and says nothing at all about where the input came from, or whether anybody agreed to hand the input over. A perfectly compliant, correctly watermarked clone of your voice, made without asking you, is still a clone made without asking you. Consent, and who gets to grant consent once you are gone, is handled elsewhere: who controls your identity after you die is the governing question, the state-by-state position on digital replicas is where the American answer currently lives, and the NO FAKES Act in plain terms is the federal proposal that would move the picture. If your worry runs the other way, toward proving that a recording really was you, the methods for proving a voice is genuinely yours cover what evidence holds up.

Ask these of every company in this category, this one included, and treat a vague answer as an answer.

Who owns the recordings and the text? You should. The company should hold a licence to operate the service on your instructions and nothing wider than that.

What does export produce, exactly? Ask for the format rather than for a yes. GDPR Article 20 sets the standard worth borrowing wherever you happen to live: personal data in a "structured, commonly used and machine-readable format", transmissible to another controller. Transcripts and audio files clear that bar. A proprietary bundle that opens in one product does not.

What does deletion actually remove? Under GDPR Article 17 you can require erasure without undue delay on the grounds listed there. Push past the policy language to the derived artefacts: the vector index, the voice model, the backups, and the retention window on each of them.

Does my data train a shared model? For Afterlife.ai™, no. Your memories and your voice operate your own build, for you and for the people you nominate, and they are not used to train any model that serves anybody else. Ask any provider to say that in one sentence without conditions, and notice if they cannot.

Where does inference run, and under whose law? Ask for the deployment type as well as the region, because those are two different answers. Ours: conversation runs on an Azure OpenAI resource provisioned in Australia East, but the model deployments are Microsoft's Global type, and Microsoft's own documentation states that a Global deployment may process prompts and responses in any geography where that model is deployed, with data at rest held in the designated geography. Retrieval embeddings are generated through Google's Gemini API. Voice synthesis runs through a separate specialist provider. All three sit under commercial terms that bar the provider from training on what passes through. Jurisdiction decides which regulator you can complain to, so a company that will not tell you has already told you something.

How is synthetic output disclosed? See Article 50 above, and expect machine-readable marking and first-interaction disclosure rather than a promise.

What happens if you cease trading? The only answer with any weight is exportable material in open formats, in your possession, before the question turns urgent.

Open a plain text file tonight. Write down ten decisions that changed the direction of your life, and under each one, two or three sentences on why you chose that way. Not what happened afterwards. Why you chose.

That file is worth more to any future model than your entire email archive, for exactly the reason the interview study found. The file is also portable to anything, because plain text outlives every company in this industry, mine included. If the drill works and you want the longer version, the self-interview method sets out how to run it properly, the question set people actually use gives you three hundred prompts to work through at your own pace, and the step-by-step build of an AI version of yourself is the procedural companion to the architecture above.

We built the same drill at Afterlife.ai™ with somebody asking the follow-up question, and the narrow version of the pitch is the only one I will make. A guide called Idy runs the interview. Idy asks, you answer in text or out loud, and the next question follows what you just opened up rather than the next line on a list. Coverage across the eleven dimensions is tracked while you work, so the thin regions are visible to you rather than to us.

Every answer is stored as a memory: your words, with a date, individually readable and individually editable. We call the result a "Persona". Personas retrieve the memories that bear on the question being asked, and where nothing relevant exists, they are instructed never to manufacture a specific fact and to speak from what is actually present.

Be exact about what that buys, because two different things get run together everywhere in this industry. The protection against invention is real and written into the system prompt. The announcement is not. Personas do not say "I have nothing on that", because the same prompt deliberately stops them from narrating missing memory aloud, on the view that people do not audit their own recall mid-conversation. If what you want is a system that raises its own silence out loud, that is a reasonable thing to want, and this one does not behave that way today.

Your voice is cloned from recordings you make. That voice model speaks only inside your own build, and only for the people you have named.

Access is the part most systems leave undefined. You nominate specific people by name. While you are alive, you decide what each of them can reach. After you die, release runs through Executor Lock™, the mechanism that verifies a death before anything opens: a nominated executor makes the request, evidence of death is verified, and a seven-day waiting period runs, during which the request is visible and can be contested. The lock takes a complete snapshot of your memories at the trigger date. Nothing is pruned or reweighted at lock time, and after the lock nobody can change your personality.

Your material comes out while you are alive as a single JSON file: your memories with their dates, your chat transcripts, your nominated people and your profile. Audio is the gap in that file today, because your recordings stay in the app rather than coming down with the export. Deleting your account removes your memories, the retrieval vectors built from them, your uploaded media, your voice recordings and your login. The cloned voice held at the synthesis provider is not torn down by that same flow yet, which is a gap being closed rather than a policy, and if that matters to you, ask us for written confirmation rather than taking this paragraph's word.

Now the limits, written from building the system rather than from selling access to one. Coverage tracking can tell you a region is thin and cannot tell you a region is wrong, so a memory you recorded carelessly stays in the record at full weight until you edit the memory yourself. Retrieval picks the passages the system judges relevant, and on a question sitting across two dimensions the system sometimes pulls the wrong side of you. Editing the offending memory fixes that in seconds, but you have to catch the error first, and the memories behind a given answer are not currently shown beside the reply. The voice is good and stops short of perfect, particularly on laughter and on the way people trail off mid-sentence. And the base model underneath is a general-purpose model with its own register, so where your record is silent you are hearing a well-behaved average rather than a person. No amount of engineering closes those gaps. More answers do.

The first 50 memories are free, no card, and the free build does not expire, so an unfinished record waits for you instead of lapsing. If you want to know what the paid plans cover, read what each plan includes. Current pricing lives on that page rather than in an article, because a stale number here is worse than no number, and the question worth pressing on with any provider is what happens at the end of a plan: whether your memories stay exportable, and whether anything is deleted automatically.

Start building your legacy Free build, 50 memories, no card required.

The architecture claims are drawn from primary sources rather than summaries: the three papers read in full, the EU AI Act and the Digital Omnibus regulation as published in the Official Journal, the GDPR articles themselves, the Italian regulator's own published decision, and the cloud and model providers' own documentation and terms. Product statements were checked against running code rather than against marketing, and three of them are deliberately narrower than the easy version: export is JSON and excludes audio, deployment region and deployment type are stated separately, and a build that will not invent is not the same as a build that raises a gap out loud. What is unsettled: no accepted benchmark exists for how closely a build of this kind resembles a specific person, so no fidelity figure here or anywhere else should be read as measured, and the 2 December 2026 marking deadline for systems already on the market has not yet been tested by any regulator.

• Park JS, Zou CQ, Kamphorst J, Egan N, Shaw A, Hill BM, Cai C, Morris MR, Liang P, Willer R, Bernstein MS. LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals. arXiv:2411.10109, v1 posted 15 November 2024, v2 posted 22 April 2026. The 83%, 82%, 86% and 74% figures quoted above are from v2. <https://arxiv.org/abs/2411.10109>

• Liu NF, Lin K, Hewitt J, Paranjape A, Bevilacqua M, Petroni F, Liang P. Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 2023. arXiv:2307.03172, 6 July 2023. <https://arxiv.org/abs/2307.03172>

• Wang C, Chen S, Wu Y, et al. Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers (VALL-E). arXiv:2301.02111, 5 January 2023. <https://arxiv.org/abs/2301.02111>

• OpenAI, Model optimization (fine-tuning) developer documentation. <https://developers.openai.com/api/docs/guides/model-optimization>

• EU Artificial Intelligence Act, Article 50, Regulation (EU) 2024/1689, applies from 2 August 2026. <https://artificialintelligenceact.eu/article/50/>

• EU Artificial Intelligence Act, Article 99(4)(g) (penalties), Regulation (EU) 2024/1689. <https://artificialintelligenceact.eu/article/99/>

• Regulation (EU) 2026/1744 (Digital Omnibus on AI), adopted 8 July 2026, OJ 24 July 2026, in force 27 July 2026. <https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng>

• European Commission, Regulatory framework for AI application timeline, checked 29 July 2026. <https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai>

• Regulation (EU) 2016/679 (GDPR), Article 20 (data portability). <https://gdpr-info.eu/art-20-gdpr/>

• Regulation (EU) 2016/679 (GDPR), Article 17 (erasure). <https://gdpr-info.eu/art-17-gdpr/>

• Garante per la protezione dei dati personali (Italy), decision of 20 December 2024 fining OpenAI EUR 15 million. <https://www.garanteprivacy.it/home/docweb/-/docweb-display/docweb/10085432>

• Microsoft, Data, privacy, and security for Foundry Models sold by Azure, covering Global and DataZone deployment types and the commitment that prompts and completions are not used to train the base models. <https://learn.microsoft.com/en-us/azure/ai-foundry/responsible-ai/openai/data-privacy>

• Google, Gemini API Additional Terms of Service, Paid Services section. <https://ai.google.dev/gemini-api/terms>

Last reviewed: 30 July 2026. Regulatory and platform-policy claims are on a 30-day recheck cycle until the Article 50 position settles. Found an error? Report it through the contact page and we will correct the text and note the change at the next review.