Selin has run the technology company she founded for twelve years. It began as a small software team and grew into an international business serving several countries. She is more than its chief executive: she is the public face of the brand. Customers watch her talks. Employees listen to her as each new business period begins. Investors often interpret the company's decisions through her words. Once the business operates in six countries, recording the same training repeatedly becomes difficult. Every new group of employees needs videos in English, Turkish, German, Spanish, Arabic and Russian.
The team proposes an AI avatar system. It will use Selin's face and voice to deliver approved scripts in different languages. At the first project meeting, the technical team explains: ‘Your face and voice will be used only for internal training and customer welcome videos you have approved in advance.’ Selin sees the benefit. Two days of recordings follow. The team captures variations in her tone and facial movements; she repeats sample sentences with different emotions. She signs a consent form.
The form says: ‘I authorise the generation of synthetic images and audio for corporate training and communications.’ Selin understands this to mean: ‘The training scripts I approve can be produced without my having to appear on camera again.’ The technical team reads it differently: ‘The face and voice model may be used for a broad class of content relating to company activities.’ In time, the marketing team will interpret it more broadly still: ‘Corporate communications could include customer videos, social media content and public statements.’
The first few months go smoothly. Selin sees every script and approves it before publication. The avatar introduces new employees to the company culture, shows customers how to use the platform and delivers short welcome messages in six languages. The system genuinely saves time. Viewers know the videos are synthetic. Every description states: ‘This content uses an AI-powered avatar created with Selin's permission.’ Then the system expands. A content agent can now generate scripts automatically from internal documents.
A localisation agent adapts each script into six languages. An avatar agent generates the face and voice. A publishing agent sends approved videos to social platforms, while a scheduling agent sets the publication times. Over time, obtaining Selin's explicit approval for each video begins to slow the operation. The marketing manager says: ‘Selin has already authorised general use. We don't need to ask for approval every time we make a routine announcement.’ A new system rule follows: high-impact public statements require human approval.
Routine corporate communications may be published under the existing general authorisation. The problem is this:
An agent now decides what is high-impact and what is routine.
One Friday, the company changes its data-use policy. The new policy allows anonymous patterns drawn from customer support records to be used for product development. The legal team prepares a lengthy, carefully worded document. It sets specific limits:
- Raw customer messages will not be used to train models.
- Personal data will be separated out.
- Users will be able to object to certain forms of processing.
- Technical and legal checks will be completed before implementation begins.
The content agent turns the document into a shorter public announcement: ‘To improve the customer experience, we are starting to use our support data to improve our AI systems.’ The sentence loses the legal document's essential limits. ‘Anonymous patterns’ has gone. So has ‘raw messages will not be used’. The requirement to complete the technical review first has disappeared, as has the right to object. The marketing agent classifies the text as a routine product-development announcement and seeks no human approval. The localisation agent translates it into six languages.
The avatar agent produces six videos using Selin's face and voice. The publishing agent sends them to social media and the company website. On Monday morning, Selin wakes to messages on her phone. An employee has written: ‘Why didn't you tell us we were starting to use our data for AI training?’ A customer asks: ‘Are our support conversations being used to train models?’ Selin cannot understand what has been said. She taps the link and sees herself on screen. Her face. Her voice. Her own speaking rhythm. The video is in English.
The avatar says: ‘To improve the customer experience, we are starting to use our support data to improve our AI systems.’ Selin watches again. She has never read that sentence, let alone approved it. She saw the document prepared by the legal team, but this short version is new to her. She hears it for the first time from her own mouth. She calls the marketing manager. ‘Why was this video published?’ The manager replies: ‘You gave general permission to use the avatar. The system classified this as routine communication.’ Selin says: ‘I did not authorise those words.’
The technical team investigates the avatar system's records. Permission to use her face is active. Permission to use the voice model is active. Corporate communications is an active purpose. The content agent derived the script from an approved policy document. The marketing agent classified the video as routine. The publishing agent used a valid general publishing token. In its own records, every technical component appears authorised. Yet nobody showed Selin the final sentence. She instructs them: ‘Stop all the videos immediately.’ The central avatar agent stops generating new content.
The dashboard displays:
Avatar production paused.
But three more videos are already scheduled on external platforms. One is due to go live in ten minutes. The external social media queue does not check the central avatar agent's stop status. A second video is published. A third goes live in another language. The technical team intervenes on each platform separately and deletes some videos. But customers have already downloaded copies. An employee has made a screen recording. A video has been shared in an industry group. A news site has quoted Selin's statement.
The company can take the videos down. It cannot undo the fact that people have heard words Selin never said. She sits silently in the meeting room for a long time. Then she says:
‘The face is mine. The voice is mine. The company is mine. But whose words are these?’
Nobody in the room can answer alone. The content agent wrote the sentence. The marketing agent classified it as low-risk. The avatar agent voiced it. The publishing agent published it. People granted general authority. The tools worked. The queues ran as scheduled. Yet the sentence has no real owner. More precisely, there are people who ought to answer for its effects, but the system has divided that responsibility into small technical decisions until it is no longer visible. Someone can say, ‘The AI generated it.’ But the AI did not change the company's data policy.
People wrote that policy. The AI did not create Selin's face and voice model on its own: people recorded her and connected the model to the system. The AI did not open its own publishing account: people granted access. Nor did it derive the limits of ‘routine content’ from a legal order of its own. People defined that category, or failed to define it. The problem is not simply that AI generated the wrong sentence. The deeper problem is this:
A person's identity has been separated from her right to the final say.
Selin's face, voice, earlier speeches and corporate role are all present in the system. Her current wishes are not connected to it with the same force. Broad permission given once has become a standing mandate to produce new statements. Permission to generate content has become permission to publish it. General consent to corporate communications has been treated as a blank cheque for any public statement. The stop request reached the central agent but not the external queues. The system resembles Selin.
At the critical moment, it does not obey her.
The problem extends beyond synthetic faces and voices
A person may discover that a loan application was rejected because of a risk score they have never seen. When they challenge the decision, the same system may review the same data and reach the same conclusion. An employee may object to an earlier permission to contact them being used for a new sales campaign. The central record changes, but old queues keep running. A patient may say they no longer want an appointment and follow-up agent to act on their behalf. Their app account closes, but external calendar and notification tokens remain active.
A customer may want to end a free trial. The agent could start it with a single call, yet sends them to a human support line to cancel. A job applicant may point out that a date on their CV was parsed incorrectly. The system replies: ‘The decision was made automatically.’ The technologies, institutions and people differ. The founding question does not:
If a machine can affect a person's life, which right must that person never lose?
NOMOS 13 answers in thirteen articles. One principle comes before them all:
The final say remains human.
This does not mean people are always the best informed or invariably right. Nor does it mean one person can instantly set aside every safety rule. It means this:
Giving a machine a task does not give it sovereignty.
A system can help people act more effectively in the world. It cannot turn their final say over their own identity, rights and future into a technical function delegated to itself.