Asking an AI to write a condolence message does more than save time. In 2026, Nadia Guerouaou, a cognitive neuroscientist, publishes Our Brain Under Influence, which examines the relationship between emotional outsourcing and brain reconfiguration through mechanisms of neuroplasticity. Neuroplasticity—the brain’s capacity to reconfigure itself based on repeated use—operates silently, without the user’s deliberate choice.

The essentials

  • According to Nadia Guerouaou (Our Brain Under Influence, 2026), outsourcing emotional expressions to an AI can modify circuits involved in emotional production, through neuroplasticity mechanisms.
  • Affective communication algorithms are trained on corpora whose composition can influence the norms of emotional expression they generate.
  • Progressive standardization of emotional expressions could reduce the diversity of emotions available in the public sphere.
  • Regulating training corpora and educating people toward affective awareness are among the proposed responses.
  • Europe’s regulatory framework for digital services (DSA) offers an existing lever, but extending it to affective norms remains to be built by 2030.

Neuroplasticity does not discriminate

This mechanism has an often underestimated implication: neuroplasticity describes experience-dependent adaptations whose consequences must be evaluated based on practices and their effects. Repetition is one factor in plasticity among many, and effects also depend on the task, context, and mechanisms involved. Repeated use of affective assistance tools could, according to certain hypotheses, produce effects on the circuits concerned. The user’s intention does not act as a brake on cortical reshaping. The gap between occasional emotional outsourcing and habitual outsourcing is a threshold the user crosses without perceiving when it happens.

The human brain reconfigures itself based on what it practices. This is an established fact since pioneering work on cortical plasticity: London taxi drivers develop denser hippocampi after years of mental navigation, musicians strengthen motor zones related to their instruments. Similar principles of brain reconfiguration apply to affective circuits. Depending on the task and training, modifications in prefrontal-limbic activity or connectivity can occur. When this production is externalized to an algorithm, the recruitment of these circuits could be altered.

Guerouaou examines what happens when writing a condolence message, a breakup letter, or a professional disagreement is delegated to a language model. LLM-assisted writing has been associated with differences in EEG connectivity in a test task, though without evidence concerning the emotional circuit. In ordinary conversation, the user reformulates their own affect into words; direct generation could reduce the share of formulation produced by the user, without necessarily demonstrating that it bypasses all personal reformulation. Studies on digital media examine links with mood and well-being, but the literature does not establish a specific trace of brain activation caused by outsourcing emotion to AI.

The critical point is spontaneous irreversibility. Neuroplasticity works both ways, but returning to an abandoned practice requires deliberate effort. One does not automatically recover a capacity one has stopped exercising.

The emotions that train algorithms

The second level of the problem is that of corpora. Large language models learn to produce emotional expressions from human texts. These texts come largely from culturally and linguistically unequally represented spaces. The emotional categories they encode, the degrees of intensity they judge appropriate, the formulations they favor all reflect this training base.

A condolence generated by a model trained on this corpus is not neutral. It carries a particular way of naming loss, of calibrating emotional distance, of balancing consolation and silence. This way is specific. It excludes other ways, equally legitimate, of traversing the same emotional register. When a significant number of users adopt these formulations for their own messages, progressive standardization of expression is observed.

Guerouaou calls this phenomenon “techno-moral normalization.” The algorithm does not dictate what one must feel, but it sets parameters on what one expresses. The relationship between feeling and expressing could be affected by progressive delegation of emotional expression.

The technological optimism argument

The most substantive competing view comes from innovation economics. Tyler Cowen, among others, defends the idea that tools reducing communication costs increase the quantity and quality of human exchanges. The printing press standardized writing and enabled its massive distribution. By this logic, AI tools that help formulate one’s emotions could enable people with limited writing ability or from linguistic minorities to communicate what they feel with greater precision and impact.

The argument deserves to be taken seriously. It is true that someone whose native language is not French or English can find in an AI assistant a lever for expression that vocabulary deficiency denied them. It is true that partial standardization of a practice does not equal its destruction. The printing press produced literary canons, but it also made a world literature possible.

The difference lies in the mechanism. The printing press standardized the material form of text, not its affective content. The use of language models constitutes a cognitive experience susceptible to being studied within the general framework of plasticity, without any demonstrated specific neuroplastic effect, and it operates at a scale and speed without historical precedent. The depth at which this standardization operates is at issue, more than standardization itself.

Democracy needs heterogeneous affects

This dynamic is all the more difficult to challenge politically because its effects remain individually invisible. Each user perceives an immediate benefit without perceiving their contribution to collective convergence. Affective heterogeneity cannot be defended as an individual good: it is only visible at the scale of interactions between citizens, where disagreements must remain productive. This asymmetry between individual perception and collective cost is precisely what makes it difficult to constitute social demand for its preservation.

The democratic question lies here. Deliberative democracy rests on citizens’ capacity to feel differently, to name that disagreement, and to bring it into the public sphere. Indignation, mistrust, solidarity, civic pride are drivers of mobilization that precede rational argument. They make that argument possible by creating the motivation to engage.

Progressive normalization of affects by algorithms could reduce expressive diversity before public debate. Homogenization of emotional expressions risks reducing the emotional diversity available to debate. Such convergence could limit the diversity of affective positions in the public sphere. The vote remains formal, but its emotional substrate has already been filtered. This is what Guerouaou calls “emotional democracy”: the condition for political emotions to remain a heterogeneous common good, not a standardized product.

This concern echoes work published on the pitfalls of digital governance. The state capable of regulating these dynamics is missing on two counts: through lack of regulatory power, but also through the absence of shared understanding of the mechanism to be regulated.

Two scenarios by 2030

Without regulation of training corpora, progressive convergence of affective patterns would constitute a risk. Intensive use of AI assistants for affective communication would likely be accompanied by modifications in patterns of emotional expression. This convergence is not catastrophic in itself: shared cultures always emerge from common practices. But a reduction in the reservoir of heterogeneous affects would pose questions for democratic deliberation.

The second scenario assumes an extension of existing regulatory frameworks. The European regulation on digital services (DSA), whose 2025 European Commission implementation report documents initial effects on content moderation, offers a mobilizable architecture. The idea would be to impose on platforms and generative AI tools an obligation to disclose the cultural corpora used to train affective communication modules. Making visible the homogeneity that is encoded is the minimum condition for users, regulators, and citizens to be able to judge it.

The bifurcation that AI operates on labor markets offers a useful precedent: inequalities settle into invisible slides that data suddenly makes legible. Affective regulation follows the same logic: acting before the mechanism becomes irreversible requires measuring what changes, not only naming it.

Two signals would make it possible to distinguish trajectories before 2030. The first is measuring the lexical and emotional diversity of political and civic messages produced with AI assistance compared to those produced without, on representative corpora. This type of measurement is technically feasible today with available linguistic analysis tools. The second is whether national regulators adopt or do not adopt criteria for diversity of affective corpora in algorithm audits provided for by existing frameworks. Audits could examine systems’ biases or norms, but an extension to affective dimensions remains to be built.

Education beyond regulation

Regulating platforms is necessary. It is not sufficient. Guerouaou advocates for programs of education in affective autonomy within school systems, centered on recognizing one’s own emotional patterns before and after using AI tools. The idea is simple: if students learn to identify what they feel before resorting to algorithmic assistance, the emotional circuit would remain more involved in the process. The AI tool enriches expression; the emotional circuit continues to be engaged.

Several prototypes of this type of program exist in North American contexts, notably in Quebec, connecting neuroscience and education policy. The challenge is to move from prototype to scale, with the public resources that requires.

The comparison with media literacy is instructive. Learning to read images, identifying a reliable source, detecting rhetorical manipulation—all these skills required decades of integration into school curricula, and they remain incomplete. Affective literacy in the face of AI will have to traverse the same path, with less time available. Early signals to watch: the inclusion of this competency in national media and information literacy programs, and public funding of longitudinal research on cognitive effects of school use of AI assistants.

The challenge consists in collectively deciding how deep these tools can operate and on which circuits, without refusing those that improve communication. Debates on automation and employment have shown that benefits and costs of a technology are rarely distributed in the same way. The benefits of affective AI are immediate and individual: a better-formulated message, a difficult conversation made easier. The costs are deferred and collective: a potential attenuation of affective diversity and emotional richness of deliberation.

This temporal imbalance is precisely what makes regulation difficult, and necessary.

Sources

  1. Maison de l’Innovation Sociale Québec, Discover the inspiring journeys of the 7 winners of the 2026 Social Innovation Personalities Awards, La Presse, July 10, 2026: https://www.lapresse.ca/xtra/2026-07-10/maison-de-l-innovation-sociale/decouvrez-les-parcours-inspirants-des-7-laureats-des-prix-personnalites-de-l-innovation-sociale-2026.php
  2. Nadia Guerouaou, Our Brain Under Influence, 2026 (publication associated with work awarded by the Maison de l’Innovation Sociale Québec)
  3. European Commission, DSA Implementation Report 2025 (implementation report on the regulation on digital services)
  4. Nature Neuroscience, research on neuroplasticity and digital media use, volumes 2024–2026