privacydeep-dive

AI Digital Twins: The Privacy Gap No One Is Talking About

Your photos, videos, and voice recordings can be used to build an AI replica of you — one that talks, responds, and mimics your likeness. The legal framework to stop this barely exists.

A new category of AI application has emerged in the past two years that sits uncomfortably at the intersection of personal data and identity: systems that build computational models of individual people — their appearance, voice, manner of speaking, and behavioral patterns — from the digital records those people have created.

These are sometimes called AI digital twins, personal AI replicas, or digital personas. The technology varies: some are primarily visual, reconstructing a person’s appearance and movement from photo and video data; others incorporate voice modeling; the most sophisticated attempt to model personality and communication style from text and behavioral data.

The data that makes them possible is the same data most people store casually in cloud services: photos, videos, voice memos, messages. Understanding how this technology works, what legal gaps exist, and what it means for how you store personal files is relevant now — not as a future concern but as a present one.


What a Digital Twin Actually Is

In industrial contexts, a digital twin is a live simulation of a physical system — a jet engine, a factory floor, a power grid — that mirrors its real-world counterpart and can be used for modeling and prediction. The concept has been extended to personal AI replicas: computational models that attempt to simulate the behavior, appearance, or communication patterns of a specific individual.

Personal AI replicas range in sophistication from relatively simple:

  • A photorealistic 3D model of a person’s appearance, animated from still photos using video synthesis models
  • A voice clone generated from a few minutes of audio, capable of speaking new text in a person’s voice
  • A text-based persona that responds to messages in a person’s apparent style, derived from social media posts or message history

To genuinely complex:

  • Multimodal models that combine visual appearance, voice, and text-style simulation into a unified conversational agent
  • “Cognitive digital twins” that attempt to model decision-making and personality, drawing on behavioral history to predict how a specific person would respond to situations they have not been asked about directly

The latter category is the subject of recent academic research on “Cognitive Digital Twins: Ethical Risks and Governance for AI Systems That Model the Mind,” published through academic preprint channels in mid-2026. The paper raises specific concerns about systems that model psychological traits and values, noting that such models introduce risks of unauthorized impersonation, overconfident inference of sensitive personality traits, and misuse for persuasion or manipulation.


The Data Inputs

What makes building a personal digital twin possible is the volume of personal data that now exists in cloud systems.

Photos and videos are the primary input for visual replicas. A collection of photos and videos spanning several years — the kind most people maintain in cloud photo libraries — contains enough data for modern video synthesis models to generate realistic animated images of a person, including facial expressions and movement.

Voice recordings are the primary input for voice clones. Consumer voice cloning tools now require only a few minutes of clear audio to generate a convincing replica voice. Voice memos, videos, recordings stored in cloud services, or audio extracted from video all qualify as inputs.

Text data — messages, emails, social media posts, journal entries — provides the behavioral and personality layer. Language models trained or fine-tuned on a specific person’s text history can reproduce their apparent communication style with enough fidelity to be mistaken for the person by people who know them.

The combination of all three produces systems that can answer messages in a person’s apparent voice and style, generate new video content with the person’s likeness, and respond to questions as the person would be expected to respond based on past behavior.


Who Is Building These and Why

The applications currently being developed span several categories.

Grief technology. Companies building AI-based memorials or companions for people who have lost someone they loved. These services create interactive simulations of deceased individuals using their digital history, marketed as a form of remembrance. The privacy concerns here are partially covered in the existing discourse around AI griefbots.

Productivity and delegation. Tools that let a person create an AI proxy that handles routine communications — email responses, scheduling coordination, customer messages — in the person’s apparent style. The promise is efficiency; the implication is that an AI is communicating as you without the recipient knowing.

Entertainment and parasocial products. Platforms allowing fans of public figures to interact with AI personas modeled on those figures — a practice that has already generated significant legal and ethical controversy around consent and likeness rights.

Non-consensual applications. The use of personal photos, audio, and video data to build replicas of people without their knowledge, for purposes ranging from deepfake content to identity fraud. This is where the legal gap is most consequential.


Where the Law Stands

The legal framework for protecting people from unauthorized digital replicas is incomplete, and the gaps are significant.

The NO FAKES Act (US). Proposed federal legislation that would create a right of action for people whose digital replicas are created without their consent for commercial purposes. As of mid-2026, the bill has broad bipartisan support but has not yet passed. The Act would primarily address commercial exploitation — its protections are less clear for non-commercial, personal, or political uses of digital replicas.

The UK’s legal gap. The UK Government’s March 2026 report on Copyright and AI acknowledged directly that existing legal tools do not adequately address digital replica risks. The report committed to exploring a “digital replica right” — a new form of intellectual property protection over one’s likeness — but no legislation has followed. The UK GDPR’s protections for biometric data apply when a replica is built from identifiable images, but this creates legal recourse after the fact, not prevention.

The EU AI Act. The EU’s AI Act, which became applicable in stages from 2024, includes requirements for transparency when AI systems interact with humans and prohibits certain uses of biometric data. Systems built from a private person’s data without consent are constrained by GDPR rather than the AI Act specifically, but the AI Act’s biometric requirements create an additional compliance layer.

Biometric data laws (US states). Illinois, Texas, Washington, and several other US states have laws specifically governing collection of biometric identifiers — including facial geometry, voice prints, and neural data. Building a digital replica using a person’s photos or audio technically involves collecting biometric data. Whether existing biometric privacy laws apply to AI replica construction is a question that has not yet been definitively answered in court.

The consistent finding across jurisdictions is that existing law was not written with AI digital twins in mind, and the protections available are at best imperfect retrofits of rules designed for other purposes.


The Personal Files Connection

The practical implication for anyone storing personal files in cloud services is that the same data that makes cloud services useful — photos, videos, voice recordings, personal documents — is the input data for digital twin construction.

Cloud services are a convenient aggregation point. A person’s photo library, video collection, voice memos, and personal documents stored in a single cloud account represent a rich enough dataset to build a reasonably capable personal AI replica if that data were accessed — either by a breach, by an employee, or by a service’s terms of service allowing uses that are broader than the user expected.

This is distinct from the concern about AI training on your content, which is primarily about your data improving a general model. Digital twin risk is about your data being used to build something that specifically impersonates you.


What You Can Do Now

Audit what you have stored and where. If your cloud photo library and voice memos exist on a service that has had breaches or whose terms of service allow broad data use, the digital twin risk is higher than on a service with strict privacy controls.

Understand what voice data you’re storing. Voice memos, video recordings with audio, and any audio files stored in cloud services are the most sensitive from a voice cloning standpoint. Consider whether the convenience of cloud backup justifies the exposure for these specific file types.

Review platform terms regarding likeness and biometric data. Several cloud and social media services have updated their terms to allow uses of uploaded content that include AI training on visual and audio data. Terms that permit using your uploaded content “to improve our services” in combination with AI development programs may cover uses you would not have agreed to explicitly.

Be skeptical about voice assistant data retention. Voice queries sent to AI assistants are audio recordings stored on service servers. What happens to those recordings, how long they are retained, and what they can be used for varies by service and is worth reviewing.

daftei does not train AI models on user content and does not analyze uploaded files for personalization purposes. Files are stored, encrypted, and returned to you. The 30-day grace window on account deletion means your files are permanently and irreversibly erased from the service after the window closes.


The Bigger Picture

Digital twin technology makes explicit something that was always true but has become more consequential: your digital history is a dataset that describes you in detail. The extent to which that dataset can be used to reconstruct or simulate you has increased dramatically with recent AI capabilities.

The legal frameworks that would restrict unauthorized use of that dataset are still being written. The practical protections available to individuals are limited. The most direct lever available to most people is deciding where personal data is stored and under what terms.

A storage service that doesn’t analyze, train on, or commercially exploit your files is a materially different environment for sensitive personal data than one that treats uploaded content as a development resource. That distinction has always mattered. It matters more now.

Your memories deserve better than an ad platform.

Try daftei free →
← All posts