The ‘I’ Written by a Machine: When AI Becomes the Ghostwriter of Our Identity

A young lecturer records a voice note after a difficult class. She sounds tired, amused and uncertain. The re is a joke that only her students will understand, a sentence that runs too long and one honest admission: ‘I may be wrong about this.’ She asks an AI assistant to turn it into a LinkedIn post.

The result is immaculate. The joke has vanished. The doubt has become ‘a key insight’. The long sentence now marches in three disciplined bullet points. The lecturer apparently feels ‘thrilled to announce’ something she had not planned to announce at all.

She reads it, winces, then publishes it. The post performs well.

Nothing dramatic has happened. No deepfake has stolen her face. No hacker has seized her account. Yet a tiny constitutional change has taken place: a system has drafted the version of her that other people will meet. If this happens every day, the interesting question is not only whether the machine learns her voice. It is whether she begins to learn the machine’s version of it.

The ghostwriter has moved inside the conversation

Traditional ghostwriters have long helped politicians, executives and public figures sound more coherent than breakfast, jet lag or human nature normally allows. They work within a social arrangement: somebody briefs them,

somebody can challenge them, and somebody knows that the final ‘I’ has been negotiated.

Generative AI makes that arrangement ordinary, private and continuous. It can rewrite an email before breakfast, soften a disagreement at lunch, prepare a condolence message in the afternoon and run a social-media account overnight. In AI-mediated communication, the technology does not merely carry a message from A to B. It may suggest, correct, modify or create the message itself (Hancock et al., 2020). As Aleksandra Skrzypiec puts it, ‘AI acts on behalf of people’ (2025, p. 32, author’s translation).

That short sentence changes the map. The machine is neither a passive pen nor a full moral agent. It is an active participant with no biography, vulnerability or responsibility of its own. Human–Machine Communication research therefore asks not only what a device can do, but how people understand themselves in relation to it. Its relational frame is especially useful here: the issue is not simply output quality, but the interpretations formed between person and system (Guzman & Lewis, 2020; Skrzypiec, 2025, pp. 35–36).

The first mistake is to imagine a clean border between ‘my idea’ and ‘its wording’. Wording carries stance. Compare: ‘I am worried this policy may fail’ with ‘This policy is fundamentally misguided’. The factual topic is similar; the public character is not. One speaker invites inquiry. The other enters the room wearing rhetorical shoulder pads.

A public self is something we practise

Erving Goffman’s account of self-presentation did not treat identity as a stone hidden inside us, waiting to be discovered. Social life resembles a sequence of performances shaped for particular audiences, with a front stage, a backstage and constant repair work when the scenery wobbles (Goffman, 1959). We are not necessarily dishonest when we speak differently to a student, an editor and a close friend. We are selecting parts of a complex self for a situation.

AI inserts a powerful dresser into that theatre. It knows the costumes favoured by professional platforms: confidence, speed, clarity, relentless usefulness and suspicious enthusiasm for ‘journeys’. It can make a hesitant person sound decisive and an angry person sound serenely strategic. Once published, however, the costume becomes evidence. Colleagues respond to it. Employers reward it. Readers expect the next performance to match.

This is where self-effects matter. Communication can affect the sender as well as the receiver. A position repeatedly expressed in public may become part of the speaker’s self-understanding, especially when others react to it and the speaker must remain consistent. With AI, that loop includes language the person did not fully generate. The sequence may be:

1. the system proposes a more polished persona;

2. the person approves it because it is effective;

3. the audience rewards that persona;

4. the person receives the reward as feedback about who they are;

5. future prompts ask the system to produce more of the same.

This is a plausible mechanism, not proof that one month of AI editing rewires a personality. Research on AI social-media clones is still early and should be treated as a warning signal rather than a population-wide diagnosis (Liu et al., 2026, pp. 1–31). Experimental work on AI-assisted social posting likewise studies bounded behaviours, not the migration of an entire self (Møller et al., 2026). The modest claim is already serious enough: repeated delegation can narrow the range of voices a person feels able to use.

How voice drift happens without anybody choosing it

Voice drift rarely arrives with a villainous notification saying, ‘Your personality will be updated tonight.’ It grows through convenience.

First comes asymmetry of suggestion. A blank page offers thousands of possible sentences; an AI draft offers one highly available path. Editing an existing paragraph is cognitively easier than recovering the unwritten alternatives. The first draft therefore sets the temperature of the room.

Second comes metric selection. Platforms reward some styles more visibly than others. A neat thread with a strong claim can outperform a careful paragraph containing uncertainty. The user concludes that the machine has found their ‘best voice’, when it may have found the voice most legible to a ranking system.

Third comes archive feedback. The assistant is asked to imitate previous posts, many of which it helped write. A synthetic style begins training the next synthetic style. It is the communicative equivalent of photocopying a photocopy and calling the fading edges a personal brand.

Fourth comes social commitment. Once an executive has published ten uncompromising posts, a nuanced eleventh post may appear weak. Once a creator has adopted breezy intimacy, reserve can look like betrayal. The audience becomes an informal enforcement mechanism for a persona produced partly by optimisation.

Finally comes agentification. The system moves from drafting to acting: replying to comments, choosing what to publish or maintaining a clone that interacts while its owner sleeps. The moral distance between person and message grows, but public attribution remains. ‘My agent posted it’ is an explanation, not an ethical eraser.

Who pays when the polished ‘I’ becomes compulsory?

The most obvious cost is embarrassment: a person sounds pompous, generic or oddly American in a message to a Polish aunt. The deeper costs concern freedom and unequal power.

Consider a hypothetical freelance reporter whose natural English is precise but visibly non-native. An assistant repeatedly removes regional turns of phrase and cautious formulations. Editors begin to expect the smoother version. The reporter now has to use AI not to gain an advantage, but to remain employable as the person the system has advertised. The convenience has become a private tax.

Or consider a local activist whose authority comes from speaking as a member of a community. If AI replaces local idiom with the global dialect of campaign communication, the message may travel further while its speaker becomes less recognisable to the people represented. Linguistic difference is not dirt on the lens. Sometimes it is the evidence that the lens belongs to someone.

The cost can also fall on audiences. A reader may believe that an apology, statement of grief or political conviction expresses a person’s considered words when it was generated, selected or posted under loose supervision. This does not mean every use of AI must be disclosed. Spell-check is not a constitutional crisis. The ethical threshold rises with the stakes: identity, commitment, vulnerability, authority and the possibility of harm.

Human dignity here includes the right to an imperfect voice: to pause, change one’s mind, use a minority variety, sound less marketable and refuse constant performance. Autonomy is not merely the ability to press ‘approve’. It requires meaningful alternatives, enough information to understand the delegation and a genuine way to withdraw it.

Responsibility cannot be delegated with the syntax

AI complicates authorship, but it does not make responsibility evaporate. A useful distinction is between causal contribution and answerability. The model may contribute thousands of tokens; the person or organisation that deploys, approves and benefits from them must still answer for foreseeable effects.

That responsibility is shared, not magically concentrated in the last person to click. A platform that designs automatic posting, an employer that demands synthetic productivity and a manager who gives staff no time to review all shape the outcome. Yet shared responsibility must not become diluted responsibility, where everybody owns 10 per cent and nobody answers the telephone.

For high-stakes speech, three questions should always have a named human answer:

• Who authorised the system to speak in this domain?

• Who could stop or alter the message before publication?

• Who will correct the record and repair harm afterwards?

If the answer to all three is ‘the user’, the user must actually have time, context and control. A ceremonial approval button placed after a thousand automated replies is not control. It is a small green alibi.

A Voice Constitution: seven clauses for keeping the self in the loop

A voice constitution is a short working document, not a solemn parchment guarded by trumpets. It can fit on one page. Its purpose is to turn vague comfort into explicit limits.

1. Define the delegation. List permitted tasks: spelling, shortening, translation, headline alternatives, tone checks. Then list excluded tasks. A newsroom might allow AI to condense a logistics email but forbid it to draft apologies, source communications, political endorsements or statements about a reporter’s personal experience.

2. Preserve a human baseline. Keep several unassisted samples from different contexts: explanatory, humorous, critical, intimate and uncertain. These are not prompts for perfect imitation. They are reference points for detecting systematic loss — for example, every hedge disappearing or every local expression becoming corporate wallpaper.

3. Keep the original beside the revision. For consequential messages, review changes in a comparison view. Ask not only ‘Is this better?’ but ‘What position, emotion or relationship changed?’ A shorter sentence may also be a harsher one. A fluent translation may quietly erase an ambiguity the source wished to preserve.

4. Establish risk tiers. Low-risk text, such as calendar notes, may be sent after a light check. Medium-risk public posts require line-by-line approval. High-risk speech — apologies, testimony, employment decisions, health claims, political positions and messages sent in another person’s name — requires deliberate human composition or documented review. No autonomous publication.

5. Mark authorship where it matters. Disclosure should be useful rather than theatrical. ‘Made with AI’ tells a reader almost nothing. State the relevant contribution: translated, substantially drafted, synthetic voice used, replies automated, or facts independently checked by a named editor. The aim is informed interpretation, not ritual confession.

6. Run a monthly drift test. Take one fresh human-written paragraph and compare it with recent assisted outputs. Look for recurring changes in certainty, warmth, vocabulary, humour, cultural reference and willingness to disagree. Ask a trusted colleague whether the public voice still resembles the person. Record the decision to continue, narrow or stop the delegation.

7. Provide revocation and repair. The user must be able to disable the agent, delete stored style profiles where the service permits, export

originals and correct messages already sent. An organisation should define who pauses automated communication during illness, conflict or reputational crisis. A clone without a kill switch is not assistance; it is tenancy.

What the constitution cannot solve

No protocol can isolate a pure, pre-technological self. Human voices have always been co-produced by teachers, editors, friends, genres and institutions. Nor is every stylistic change a loss. AI may help a person with dyslexia express an idea more clearly, give a second-language writer access to a professional arena or allow someone with speech impairment to communicate independently. Refusing assistance in the name of ‘authenticity’ can become its own form of exclusion.

The constitution also cannot correct the platform economy that rewards predictable performance, or reveal every influence embedded in a proprietary model. A user may monitor outputs without knowing which styles were privileged during training. And a monthly comparison will not measure identity with laboratory precision.

That is why the aim is not purity. It is reversible authorship: the capacity to see what was delegated, contest the result and return to a voice that has not been made economically impossible.

The lecturer from the opening scene need not throw away the polished post. She might restore the joke, return the uncertainty and delete ‘thrilled to announce’. The final version may perform less impressively. It may also perform a more important task: allowing the person recognised by her students to remain recognisable to herself.

AI can hold the pen. It should not quietly acquire the casting vote on who the ‘I’ is.