The Intimacy of Voice
The human voice is probably the most intimate communication channel there is. It carries not only the words we choose, but also those we hesitate to pronounce. It reveals our fatigue, our enthusiasm, our anxiety, our confidence. It betrays our lies through micro-pauses, our doubts through hesitations, our certainties through a firm rhythm. It is the sonic reflection of our thought in the process of forming, long before it is fully structured.
When you dictate a text, you do not simply transmit content. You deliver a recording of your cognitive process. Your false starts, your self-corrections, your hesitations between two formulations, your sighs of frustration, your satisfied laughs — all of this is captured, analyzed, modeled. It is a window onto your mind far wider than the final text suggests.
And it is precisely this window that cloud dictation services exploit. Not to serve you better, but to know you better. To predict your behaviors. To anticipate your decisions. To influence your choices before you have even consciously formulated them.
The Economic Model of Influence
The major technology platforms are not software companies. They are influence companies. Their product is not the tool you use, but the modification of your behavior that this tool enables. You do not pay with money — you pay with your attention, your data, your future decisions.
Voice dictation is a particularly powerful vector of influence, because it captures thought in its raw state. When you type, you have time to think, rephrase, censor. When you speak, thought flows more freely, less filtered, more authentic. It is this authenticity that AI models seek to model — not to help you express yourself better, but to learn to make you express what they want.
But dictation is only one dimension of the issue. The other, even deeper, is interaction. Increasingly, we do not simply dictate a text — we ask an LLM to write it for us. We submit a situation, a decision, a dilemma, and we expect an opinion. "Draft an email to my client to postpone the meeting." "What do you think of this commercial proposal?" "How do I phrase my refusal without hurting?" These requests are no longer transcription. They are conversations. And in a conversation with a cloud LLM, it is the model that has the last word on the shape of your thought.
The applications are immense and terrifying. Models trained on millions of voices can identify your emotional vulnerability points. They can detect the moments when you are fatigable, stressed, imprudent. They can calibrate their suggestions to seem more relevant, more convincing, more difficult to refuse. They can test formulations, measure your reactions, refine their approaches — all at a scale and speed impossible for a human manipulator.
From Prediction to Influence
The shift from prediction to influence is subtle, almost imperceptible. It begins with "useful" suggestions — a slightly more impactful formulation, an additional argument, a calculated touch of empathy. Then it evolves toward more marked orientations — formulations that minimize your reservations, structures that privilege certain viewpoints, tones that elicit specific emotions.
Soon, the tool no longer simply suggests words. It suggests thoughts. Perspectives. Positions. And because it has been trained on your own data, it knows exactly how to present them so they seem to come from you. The most effective manipulation is that which the victim does not perceive as such.
Consider a concrete scenario. You dictate an important email — a salary negotiation, a delicate discussion with a client, a political stance. The dictation tool, nourished by your past habits, your vocabulary, your style, suggests formulations. Some are neutral, others slightly oriented. Without realizing it, you adopt a formulation that softens your position, minimizes your demand, privileges conciliation over assertion. This is not your choice. It is the model's, calibrated to minimize conflicts, maximize engagement, optimize a metric that does not concern you.
The second scenario is even more insidious, because it concerns not your words, but your judgment. You ask a cloud LLM: "Does this contract clause seem reasonable to me?" or "Should I accept this offer?" or "What do you think of this strategy?" The model responds. Prudently, reasonably, in a balanced way. But balanced according to what parameters? Calibrated on what corpora? Aligned on what interests? You have no way of knowing. What you perceive as a neutral opinion is the product of a model trained with objectives that its publisher will never detail publicly.
The Cloud LLM as a Biased Echo Chamber
We must understand the deep nature of what happens when you consult a cloud LLM. You are not asking a question of a neutral oracle. You are interacting with a system whose every response has been elaborated by an alignment process — a set of deliberate choices about what the model must say, how to say it, which positions it must privilege, which perspectives it must attenuate. This alignment is not a bug. It is a design.
Take a banal example: you ask a cloud LLM to draft an email to your boss asking for a raise. The model proposes a polite, measured version that highlights your value without ever being direct about the figure. It spontaneously adds a conciliatory formula, minimizes the assertive tone, avoids formulations that could "create tensions." You adopt its version. It seems natural to you — precisely because it has been calibrated to seem so. But is it really the email you would have written? Is it the position you would have defended?
Multiply this phenomenon by millions of daily interactions — negotiations, opinions on professional decisions, positions on sensitive subjects, formulations of complaints or demands — and you get an imperceptible leveling of individual expressions toward a norm fixed by a few engineers in a few companies. This is not censorship. It is far worse: it is assisted self-censorship, integrated into your workflow to the point that you no longer perceive it as such.
A local LLM does not operate according to this logic. When you ask it to draft this same email, it executes your instruction without underlying agenda. It does not seek to preserve social peace. It has not been trained to avoid certain subjects so as not to "generate controversy." It is accountable to no shareholder, no advertiser, no government. It does what you ask, in the terms you set, with the tone you choose. The opinion it gives you is structurally more neutral — not because the model is perfect, but because it has no reason to orient you toward one answer rather than another.
The Concentration of Power
This capacity for influence is not distributed democratically. It is concentrated in the hands of a few companies — mostly American, some Chinese — that control the infrastructures, the models, the data. These companies have their own commercial interests, their own political agendas, their own visions of society. And they have the technical power to promote them surreptitiously, one suggested word at a time, one privileged formulation by millions of users.
Conflicts of interest are systemic. A company that sells advertising has an interest in your messages being more engaging, more viral, more likely to generate attention time. A company that sells consumer products has an interest in your communications being more oriented toward acquisition, desire, lack. A company with political affiliations has an interest in your vocabulary, your framing, your tone gradually aligning with its positions.
And even in the absence of deliberate malice, biases infiltrate. Models are trained on corpora that reflect societal biases — gender, race, class, culture. They amplify these biases in their suggestions, subtly reinforcing them each time you adopt them. Cloud dictation does not simply transcribe your voice. It filters it, adjusts it, harmonizes it with a statistical norm that is not yours.
The Alternative: Reclaiming Control of the Narrative
Faced with this threat, local AI is not simply a technical option. It is an act of narrative resistance. It is the refusal to let a third party model your expression, filter your thought, orient your communication.
When you use a fully local tool, you eliminate this layer of interference. Whisper transcribes your words without judging them. Llama.cpp transforms them according to your explicit instructions, not according to opaque objectives. But more importantly: when you question a local LLM about a decision, when you ask it to draft an email, analyze a situation, give you its opinion — you obtain a response that has been calibrated on no commercial interest, no moderation policy, no agenda external to your request. The result is injected at your cursor without passing through a server that could modify it, analyze it, store it.
You regain the integrity of your communication chain. From your thought to your expression, there is only you and tools you control. No algorithmic intermediary that "optimizes" your message. No model that "suggests" a more "effective" formulation. No infrastructure that analyzes your tone to calibrate its response.
PerkySue embodies this integrity. The 7 to 12 billion parameter model running on your machine does not seek to influence you. It executes your instructions. You say "make this text more formal," it formalizes — without adding political subtext, without calibrating empathy to maximize engagement. You ask "draft this refusal email," it drafts according to your parameters, not according to an imposed conciliation norm. You ask a question about a decision to make, it gives you an analysis — not an answer optimized to make you adopt a predetermined position. This is the fundamental difference between a tool and a mediator.
The Sovereignty of Expression
Freedom of expression is not only the right to say what one thinks. It is also the right to think without interference. The right to formulate one's ideas in an environment that does not distort them. The right to communicate without a third party surreptitiously modeling the message.
Cloud dictation threatens this sovereignty. Not through open censorship — that would be too visible, too contestable — but through gradual influence, subtle suggestion, algorithmic calibration. Each suggested word, each privileged formulation, each implicit tone contributes to shaping not only what you say, but what you think you are saying, what you think you are thinking.
Local AI restores this sovereignty. It makes you master of your expressive chain — and of your thought chain. It guarantees that your words are yours, that your decisions are yours, that the opinions you receive have not been filtered by a model calibrated for objectives that are not yours, controlled by entities that owe you no accountability.
Toward an Ethics of Interaction with AI
We must develop an ethics of interaction with AI, that goes beyond the sole question of voice data. Because the problem is no longer only "what is done with your voice?" — it is "what is done with your thought?" Every time you ask a cloud LLM to draft an email, analyze a situation, or give you its opinion, you submit your decision-making process to a system whose biases are structurally invisible to you. This calls for clear principles.
These principles could include: informed and revocable consent for any use of data beyond immediate transcription; transparency on the models used and their training objectives; the possibility of independent audit of suggestion algorithms; the right to be forgotten and complete data deletion; the prohibition of using voice data for behavioral or political influence.
But while waiting for these regulations to emerge — and they will emerge slowly, if they emerge — responsibility falls on each user. Every time you choose a cloud tool for your dictation, you vote for an economic model based on the exploitation of your data. Every time you choose a local tool, you vote for a model based on respect for your autonomy.
Conclusion: Thought as a Red Line
Your voice is the last frontier of your cognitive intimacy — but it is no longer the only one at stake. Every interaction with a cloud LLM, every email requested from a model, every decision submitted to its judgment, exposes your thought to a layer of influence that you do not see and cannot control. Entrusting it to cloud infrastructures is entrusting to third parties the power to model not only what you say, but what you think to ask — and what you accept as an answer.
The alternative exists. It is technical — Whisper, llama.cpp, 7 to 12 billion parameter models that fit on your machine. It is economic — a single hardware investment against perpetual subscriptions. It is ethical — your data remain yours, your expression remains yours, your thought remains yours.
The question is not whether cloud models are more powerful. They are, and they will remain so. The question is whether you are ready to pay the price of that power — not in dollars, but in autonomy, in intimacy, in cognitive integrity.
Because in the end, dictation is just one productivity tool among many. The real issue is all your interactions with AI: dictating a text, asking it to be written, soliciting an opinion, having a decision arbitrated. It is the channel through which your thought takes shape in the world — and this channel deserves to be yours, entirely, definitively, without compromise.
About the Author
Jérôme Corbiau is the creator of PerkySue, a local voice dictation tool with AI that works entirely offline, with no remote server or data transmitted. He is also co-founder and software architect of My App Zone SRL (Brussels), and creator of the Cloud Neareo platform — an award-winning CMS notably by Microsoft and the Public Service of Wallonia, deployed in museums and heritage sites. His work aims at a constant objective: putting technology at the service of the user, rather than the reverse.
Top comments (0)