Expressive voice filters modify listeners’ emotional inferences. This is not a hypothesis: it is what research published in Philosophical Transactions of the Royal Society B, one of the world’s oldest scientific journals, demonstrates. When a voice passes through a filter that makes it sound more “smiling,” listeners conclude that the speaker is in a better mood than they actually are. They adjust their own affective response accordingly. The mediation no longer conveys emotion: it manufactures another one.
This mechanism is no longer theoretical. Voice filters that modify timbre and character are already integrated into certain platforms — Discord offers native filters — and third-party tools work with Zoom or WhatsApp. However, sophisticated emotional expressive filters, those that subtly adjust the affective content of the voice in the sense of Guerouaou’s research, remain largely at the prototype or limited deployment stage. Millions of people communicate every day through filters that adjust their voice in real time, without having explicitly chosen it and without their interlocutor being informed. The question is not whether technology modifies communication: all technology does. The question is what happens when the intervention bears on affect itself, before it reaches the other person.
The Essentials
- Research published in Philosophical Transactions of the Royal Society B establishes that expressive voice filters modify listeners’ emotional inferences, independently of the speaker’s actual emotion.
- While timbre modification filters are already deployed on certain platforms, advanced emotional expressive filters remain largely in the experimental stage — without this distinction being visible to users.
- Neuroscientist Nadia Guerouaou and the Human Technology Foundation identify a structural risk: the progressive normalization of a single emotional norm, defined by platforms.
- No existing regulatory framework — neither GDPR, nor the European regulation on AI, nor debates on attention economy — explicitly addresses the modification of affective expression in real time.
- The critical horizon is a generation: if expressive filters follow the same adoption trajectory as photo filters, the gap between experienced emotion and transmitted emotion will become structural before being recognized as a problem.
What Voice Carries That Text Does Not
Voice is not merely a channel for transmitting information. It is the primary vector of what psychologists call para-verbal communication: timbre, rhythm, variations in intensity, micro-hesitations that signal doubt, inflections that betray fatigue or anxiety. Decades of research in communication psychology show that these signals are processed by listeners rapidly and largely unconsciously. They precede and condition the interpretation of verbal content.
It is precisely this register that expressive filters intervene in. Making a voice sound more “smiling” does not consist of changing words: it consists of modifying the acoustic structure of para-verbal signals. And this is where the result of research published in Philosophical Transactions B takes on its full scope. Listeners do not perceive a filtered voice as “a voice modified by a tool.” They perceive a voice, and they infer an emotion from it. The filter disappears in perception. The manufactured emotion takes the place of the experienced emotion.
The comparison with other technical mediations is instructive — and it reveals how this one is different. The telephone distorts frequencies, MP3 compression cuts certain sound ranges. These alterations are unintentional losses, artifacts of the channel. Expressive filters are the exact opposite: intentional additions, signal corrections aimed at a specific emotional effect. The intention is not to transmit the voice as it is. The intention is to transmit a voice as it should, according to a parameter, be perceived.
The Invisible Norm That Configures Default Settings
When a filter makes a voice sound more “smiling,” it implies a prior judgment: smiling is preferable to neutral. This judgment is not natural. It reflects a specific cultural norm, that of so-called WEIRD cultures (Western, Educated, Industrialized, Rich, Democratic) and more particularly of major technology platforms calibrated for the American market. In many cultures, vocal neutrality signals seriousness and respect; excessive vocal enthusiasm is perceived as insincere or superficial.
This is the heart of the thesis developed by Nadia Guerouaou, neuroscientist specializing in cognitive biases of AI systems, and synthesized in the work of the Human Technology Foundation on emotions and sociability in the age of filters. Expressive filters are not culturally neutral. They encode an affective norm and propagate it at the scale of platforms. When this norm becomes the default setting for hundreds of millions of users, it no longer remains one option among others: it becomes the emotional spirit of the times of digital communication.
This dynamic has a precedent in media history. Television in the 1950s progressively imposed particular prosody on journalists: a speech rhythm, a management of silences, an emotional register that became the norm of professional communication in many countries. Radio journalists from Africa and Asia trained in Western schools often adopted this imported register. The difference with voice filters is one of degree: where television standardized through imitation and training, filters standardize in real time and without visibility. You cannot see the filter. You therefore cannot resist its influence.
What Regulators Have Not Yet Looked At
The European regulation on AI, adopted in 2024, classifies systems according to their level of risk and imposes transparency obligations for systems with significant impact. It explicitly targets emotion recognition systems in workplaces and educational institutions. But it does not address filters for modifying affective expression in real time in everyday communication. This is not a minor blind spot: it is precisely the most diffuse and most massive mechanism.
GDPR, for its part, protects data relating to emotional state insofar as it constitutes sensitive data — but only their collection and processing. The modification of emotional expression before transmission escapes this framework: no data is collected on the user’s emotion, the filter operates on the acoustic signal, not on a stored inference.
Debates on attention economy, which have structured criticism of platforms for the past ten years, focus on the capture of attention and its effects on mental health, democracy, market concentration. They have produced important works and partial regulations. But they presuppose a sender whose signal arrives intact to the listener: the problem is overload, not alteration. Expressive filters shift the problem: it is no longer the quantity of signal that needs to be regulated, but its affective quality.
This is where cross-reading with Daron Acemoglu’s work on technology and power sheds useful light. Acemoglu distinguishes technologies that “augment” human capacities from those that substitute for or reconfigure them to the benefit of those who deploy them. Expressive filters clearly belong to the second category: they do not give the user better control of their expression, they transfer that control to the platform. The question of governance is not technical — it is political.
The Analogy of Photo Filters, and Why It Is Not Reassuring
The most common argument for downplaying the phenomenon is the analogy with photo filters. Instagram generalized automatic image retouching starting in 2012. Users learned to read filtered photos as filtered photos. Society absorbed this new visual norm without perceptual collapse. Why would voice filters be different?
There are at least two serious reasons. The first is the nature of the signal. A retouched photo remains visually a photo: its contours, its composition, its staging are readable as cultural artifacts. Voice, on the other hand, is processed by older and less analytical cognitive circuits. Research on prosody processing shows that emotional inferences from voice are largely preconscious and difficult to correct through deliberation. The voice filter disappears in perception in a way the visual filter does not.
The second reason is reciprocity. Vocal communication is a loop: what I hear in the other’s voice modulates my affective response, which in turn modifies my own expression. If both interlocutors communicate through filters that adjust their respective emotions, the affective dynamic of the exchange is entirely constructed by the algorithms of two platforms. What we still call “a conversation” is actually an exchange between two filtered representations. The shared emotion is an emotion manufactured by algorithmic consensus.
This trajectory deserves to be calculated seriously. Photo filters took approximately ten years to go from gadget to implicit standard. If voice filters follow a comparable adoption curve, users born after 2015 will never have known any other standard of digital voice communication than the filtered kind. The norm will not be chosen: it will be inherited.
What Intelligent Regulation Could Do — and What Remains Open
Two positions are emerging, which it would be misleading to reduce to an opposition between pro-technology and anti-technology. The first, close to Acemoglu and Johnson’s approach to technology governance, calls for notification obligations: any active expressive filter must be notified to the sender and, in professional or institutional contexts, to the listener. Not a ban, but minimal transparency that restores to users awareness of the mediation.
The second position, more liberal in its instruments, relies on competition and differentiation. If expressive filters become an object of public debate, some platforms could make their absence a commercial argument. Users aware of the dynamic can choose tools that do not filter. Market pressure can correct what regulation struggles to reach — provided that information circulates.
Both positions converge on one point: the central problem is not the existence of filters, it is their invisibility. A visible filter can be contested, deactivated, discussed. An invisible filter configures reality without offering any leverage. It is this information asymmetry that regulation should address as a priority.
There remains a question that neither one settles: who defines the emotional norm against which the filter adjusts the signal? This decision is not technical. It assumes a choice about what is “normal,” “appropriate,” “desirable” emotional expression. It is a normative question that belongs to public debate, not engineering. It has not yet taken place.
One can imagine institutional safeguards analogous to those that frame advertising or image modification in media. One can imagine audit labels for expressive filtering systems. One can imagine access rights to default parameters. What is more difficult to imagine is how to re-politicize a question that the speed of technological adoption has transformed into an accomplished fact before we had time to pose it.
The debate on digital identity took twenty years to enter law and institutions. The debate on the modification of communicative affect has not begun. This gap in tempo between the speed of platforms and that of democracies is the real lever to watch — more than any particular filter.
Sources
- Human Technology Foundation / Nadia Guerouaou, “Emotions and Sociability in the Age of AI Filters” : https://www.human-technology-foundation.org/news/emotions-and-sociability-in-the-age-of-ai-filters
- Philosophical Transactions of the Royal Society B — reference journal for biological and cognitive sciences (The Royal Society, London)
- Daron Acemoglu and Simon Johnson, Power and Progress (PublicAffairs, 2023)
- European regulation on artificial intelligence (AI Act), Official Journal of the European Union, 2024
- Guerouaou et al. (2021), “Voice modulation: from origin and mechanism to social consequences”, Philosophical Transactions of the Royal Society B : https://royalsocietypublishing.org/rstb/article/376/1840/20200386/108749/Voice-modulation-from-origin-and-mechanism-to
- Human Technology Foundation — “Emotions and Sociability in the Age of AI Filters” : https://www.human-technology-foundation.org/fr-news/emotions-et-sociabilite-a-lere-des-filtres-ia
- European Commission — AI Act (prohibitions) : https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
- Future of Privacy Forum (FPF) — “Red Lines Under EU AI Act: Unpacking the Prohibition of Emotion Recognition in the Workplace and Education Institutions” : https://fpf.org/blog/red-lines-under-eu-ai-act-unpacking-the-prohibition-of-emotion-recognition-in-the-workplace-and-education-institutions/
- CNIL — Questions and Answers on the European Regulation on AI : https://www.cnil.fr/fr/entree-en-vigueur-du-reglement-europeen-sur-lia-les-premieres-questions-reponses-de-la-cnil
- PubMed — Preconscious Processing of Emotional Prosody : https://pubmed.ncbi.nlm.nih.gov/17964813/
- Nadia Guerouaou — IRCAM : https://www.ircam.fr/person/-98/
- Nadia Guerouaou, Notre cerveau sous influence (Eyrolles, 2026) : https://www.amazon.fr/Notre-cerveau-sous-influence-g%C3%A9n%C3%A9ratives/dp/2416023489