Recommendation in the era of GenAI: from an attention economy to a trust economy

1 Introduction

With the proliferation of online cultural content services, catalogs and the stock of available content have become massive. To limit ourselves to the video and music content to which this article is dedicated, for the past decade, offerings have been found on platforms as diverse as Digital Service Providers (DSPs) like Spotify or Netflix, on social networks, or on User-Generated Content (UGC) platforms like YouTube or TikTok. More recently, a new market shift has been taking place; the explosion of Generative AI (GenAI) models and the profusion of their consumer tools are leading to a flood of synthetic content inundating all types of platforms. This article aims to understand to what extent this new market shift leads to a transformation of the role of recommendation in the digital economy, a major objective of European public policies.

First, we review the role of recommendation systems through a short literature review. For many platforms, guiding users within a plethoric offering through appropriate recommendation systems has become a weapon of competitive differentiation. In this attention economy, attracting the consumer to a given content is essential. Furthermore, the study of recommendation systems has shown their contrasting effects on cultural diversity.

The tsunami of synthetic offerings, which we examine in the second part, weakens the existing ecosystem and challenges consumer trust. Indeed, the supply of synthetic content entirely generated by AI has increased considerably with the widespread availability of tools offered to the general public from 2022 onwards.

In the final part, we propose a forward-looking analysis of the evolution that this tsunami could lead to, in order to restore a digital economy of trust. While the emphasis has long been placed on developing the best interfaces and recommendation algorithms, it is now also necessary to acquire tools that allow platforms to detect the difference between human content and AI-generated content. Once this detection is achieved, each platform would choose to adjust its recommendation guidelines in order to modulate the promotion of human-made works.

2 Recommendation: a major stake in an attention economy

2.1 Navigating content abundance: discoverability and recommendation

The profusion of content and the fact that it is all available simultaneously and permanently is profoundly transforming the audience experience. In the music sector, the number of releases has been steadily growing since the 2000s (Waldfogel, 2015), and more than 100 million tracks are now accessible on Spotify. Every day, tens of thousands of new songs continue to enrich this considerable mass. In the audiovisual sector, the logic is similar. Online consumption has also developed, and subscription video services have multiplied (Arcom, 2024, Hadopi and CSA, 2021), growing from 63 platforms in 2017 to 78 in 2021, and then to 125 in 2024. Even limiting the count to seven services available in France (Netflix, Prime Video, Disney+, Max, Universciné, Mubi, and Arte), there are more than 5,700 unique available films.¹ Faced with such vast catalogs, the abundance of content does not necessarily facilitate access to culture; on the contrary, it often makes choice more complex. Users find themselves confronted with a form of information overload, where the difficulty is no longer finding a specific work, but identifying which work we want to consult. In this context, the volume of immediately available content can reduce their well-being by increasing search and decision costs. This dynamic is part of what is known as the attention economy, already outlined by sociologist Gabriel Tarde (1901), who emphasizes that the overabundance of goods requires systems capable of capturing consumer attention.

The indefinite increase in options increases the difficulty of the decision-making task (Chernev, Böckenholt and Goodman, 2015) and induces uncertainty about preferences, which can lead to choice paralysis (Larceneux and Berger, 2006). This hyperchoice generates a high psychological and cognitive cost (Haynes, 2009), exacerbated in the digital era by a digital fatigue linked to hyper-information. The consumer must then rely either on traditional market cues when the choice concerns the consumption of goods (Larceneux et al., 2007), or on recommendation algorithms of the platforms to preserve decision-making autonomy (El Gozmir and Chouhbi, 2026). In the specific field of cultural industries, the effects of overabundant offerings also manifest in consumer choices. For example, a study conducted on the Netflix streaming platform shows that the extent of the content catalog increases search time, increases and selection effort in the face of choice overload (Romero Meza and D'Urso, 2024).

Consequently, in a context of information abundance, audience attention becomes a rare and valuable resource at the heart of the attention economy and competition among cultural industry players (Citton, 2014; Simon, 1996).

In such a context, recommendation systems make it possible to guide audiences through the abundance of available content. Users are thus guided, in a more or less direct way, toward content likely to match their preferences, thereby reducing their search costs (Bakos, 1997; Bourreau and Gaudin, 2022). Recommendation in the cultural industries has therefore established itself as an instrument of guidance, facilitating the discovery and selection of works within ever-expanding catalogs.

For platforms, these systems are no longer simple technical features, but real levers for competitive differentiation and data acquisition tools. To lock in subscriber loyalty against the constant risk of unsubscribing (churn), continuously adjust the relevance of their algorithms, and improve user experience, streaming services rely on dedicated Key Performance Indicators (KPIs). These evaluation tools are divided into several categories: algorithmic performance KPIs, interaction and engagement KPIs, and finally catalog structure and commercial experience KPIs. Recommendation is therefore studied through companies’ internal tools.

The gradual development of recommendation tools has given rise not to a single form of recommendation, but to a diversity of recommendation systems and mechanisms, according to the strategies defined by companies. These recommendation systems can schematically be classified into four ideal-typical categories, often subject to hybridization: editorial, personalized, self-sustaining, and contributory (Farchy et al., 2017).

Editorial recommendation is based on a selection made by experts or teams of curators, who promote certain content for all users of the service. Personalized recommendation, on the contrary, relies on algorithms that exploit users’ past consumption behaviors to propose works likely to match their individual preferences. Self-sustaining recommendation refers to mechanisms that reinforce the visibility of already popular content, favoring those that benefit from a high audience level, which can accentuate concentration dynamics. Finally, contributory recommendation relies on the activity of the users themselves, through reviews, ratings, shares, or playlists, thus integrating a social dimension into the guidance of choices.

No recommendation system can ever be neutral since, by definition, it refers to a choice. This is the case with selections by curators or decisions made by algorithm programmers. These trade-offs, more or less visible, can direct the promotion of certain content and, consequently, influence cultural diversity. In the case of Netflix, a study shows that of the approximately 5,272 titles available in the French catalog in May 2022, a maximum of 20% of them appear on subscribers’ homepages (Farchy et al., 2022), with most of the catalog remaining largely invisible unless actively searched for by the user. There is therefore significant power on the side of actors who occupy a dual position, acting both as content distributors and as decision-makers in promotion mechanisms

Benabou and Farchy (2021). emphasize that recommendation systems indeed influence the discoverability of works and, consequently, potentially the cultural diversity of consumption. The authors insist on the need for transparency in selection criteria and algorithms, to guarantee that certain types of available content are not systematically eclipsed by popularity or commercial logics, and that audiences can access a plurality of creations.

Nevertheless, these recommendation systems essentially function as choice architectures designed to guide decisions without imposing formal constraints on the user, who retains the freedom to follow or deviate from them. In reality, consumers combine multiple sources of content discovery, combining organic research, word-of-mouth, social networks, and recommendations offered by the platform (Frey, 2021). On music streaming platforms such as Spotify, user habits reflect a high degree of personal choice. Listeners are far from being passive receivers subjected to recommendations; they articulate algorithmic suggestions with their own listening routines and maintain significant autonomy in exploring catalogs (Beuscart et al., 2019).

However, while in the consumption of music, films, and series, users never rely exclusively on the platform's recommendation, this is not the case in all kinds of digital consumption. Other types of platforms, due to their design and business model, guide users much more broadly toward the use of recommendation. For example, TikTok places the user in a logic of passive discovery, where the consumed videos are mainly proposed by the algorithm rather than actively searched for. The platform thus deploys particularly aggressive attention-capturing strategies, where the continuous flow generates compulsive engagement dynamics, or even forms of behavioral addiction (Montag et al., 2021).

A distinction must therefore be made between catalog diversity, exposure diversity, and consumption diversity. A diversified offering does not guarantee that promotion will be diversified, and diversifying promotion does not guarantee that the viewer will follow this recommendation to the letter. Public authorities endeavor to impose diversity quotas on catalogs; however, these measures will only be effective if these contents are also highlighted and consumers are receptive to them. Just because content is available does not mean it is effectively visible and consumed by users.

2.2 The contrasting effects of recommendation on diversity

2.2.1 User lock-in

The high personalization of recommendations, through the analysis of past behaviors by algorithmic tools, feeds fears of locking users into limited content categories. In his book The Filter Bubble, Pariser (2011) proposes a definition of this algorithmic lock-in, which he describes as an opaque bubble, a sort of prison that excludes the possibility of serendipitous discoveries. Applied to the domain of information, this thesis raises the question of access to the pluralism of opinions, essential in a democratic society. Transposed to cultural services, it gives rise to the fear that audiences may find themselves restricted to narrow tastes, to the detriment of genuine cultural diversity, reinforcing monomaniacal consumption habits.

To evaluate the existence of a filter bubble in recommendations within the cultural industries, several empirical studies have been conducted. One of these focused on the Netflix homepage (Farchy and Et Tallec, 2023). The study simulated monomaniacal users using robots and observed the evolution of their homepages over eight days. The results show that, while the page adapts to the user's tastes, it does not systematically reject other genres. Thus, despite highly targeted preferences, Netflix continues to suggest content from different categories, suggesting that no binding filter bubble forms systematically on this platform.

In the music and podcast domains, several studies on Spotify show that algorithmic recommendations influence consumption diversity. Anderson et al. (2020) analyzed the listening habits of millions of users and compared consumption guided by the algorithm to that guided by the user's choices. They find that listening diversity is greater when users make their own listening choices rather than relying on an algorithmic suggestion. Another result of the study is that users who expand their consumption over time do so mainly by exploring on their own rather than by following recommendations. This suggests that, while recommendations are effective at meeting immediate user needs, they also tend to reduce individual consumption diversity. A complementary study conducted by Holtz et al. (2020) on podcasts compared personalized recommendations, based on each user's listening history, with recommendations based on popularity within the demographic group. Personalized recommendations increased engagement, with nearly 29% more podcasts listened to per user, but reduced individual-level diversity by 11.5% and slightly increased group-level diversity (5.96%). Without necessarily concluding on the existence or absence of a filter bubble, these studies show that personalization influences the diversity of content actually consumed by users.

2.2.2 Popularity bias

Another frequently expressed concern regarding the effects of recommendation systems is popularity bias Kapwing (n.d.), which leads to favoring content that is already widely consumed. By systematically favoring the most popular works to the detriment of the “long tail”, these systems can reduce the visibility and discoverability of more confidential productions, risking an impact on the dynamism of artist discovery from the consumers’ perspective.

In the music sector, Kowald et al. (2020), for example, studied this phenomenon by comparing six of the most widely used recommendation algorithms based on a Last.fm dataset. The results show that, with the exception of a random algorithm, all algorithms tend to recommend popular artists more than less-known ones, regardless of the kind of user group considered. However, other research nuances this finding. Beuscart et al. (2019) observe, from the analysis of millions of plays, that most recommendation systems guide listeners toward more confidential artists than those they initially listened to. Similarly, a study conducted in partnership with Deezer (Villermet et al., 2021) distinguishes two user profiles: those who use the service as a personal library, listening only to tracks they choose, and those who rely on recommendations, viewing the service as a kind of radio that allows room for discovery and chance. The authors conclude that the latter develop a more varied listening experience, less concentrated on popular artists.

Regarding editorial recommendation, studies also nuance this fear of popularity bias. In the case of Spotify, Aguiar et al. (2021) show that the editorial curation of the New Music Friday playlist does not seem biased in favor of major labels. On the contrary, the curation of tracks tends to favor music from independent labels. Similarly, Laqueyrerie and M'Barki (2025) emphasize that, even though established artists are largely overrepresented in the platform's editorial playlists, suggesting a certain popularity bias in playlist creation by curators, the effects of recommendation on streams remain comparable for emerging and established artists. Indeed, inclusion in a playlist generates similar proportional gains for both groups. Thus, while recommendation systems reinforce certain imbalances, they do not necessarily lead to the systematic exclusion of niche content or less-known artists.

3 The tsunami of synthetic offerings

3.1 The emergence of a market

Within a few years, with the rise of generative artificial intelligences accessible to the general public, artistic creation has become automated, or at least greatly simplified. It is now possible, from a simple text prompt, to produce books, images, videos, or even music. However, AI should not be viewed as a single, uniform entity. AIs are diverse, models are specialized in specific fields, and final tools have distinct functionalities and objectives. Thus, a work can be entirely generated by an AI, from A to Z, or designed in a hybrid manner, relying on AI as a tool for assistance and support in creation.

Regarding the “ghost creations” (images, audio files, or videos) that interest us here, which are mass-generated by AI without significant human intervention, several categories can also be distinguished. Some productions do not seek to rely on proven resemblance to what exists, but simply to be perceived as authentic by the public. These “ghost” creations, even if they do not look like any existing people or elements, indeed replicate a human appearance (artists, music groups, influencers, all “ghost personalities”..), elements of human appearance (voice, face..) or references to creators or authors that can be perceived as human and deceive the public. Other productions that show almost exact resemblances to people (their complete appearance or simply their voice or face) are classified as deepfakes by the AI Act (Article 3, point 60).

The rapid expansion of companies developing these technologies has been made possible by large fundraising rounds that are becoming common practice (Le Monde, 2025, Mistral AI, 2025). The European market for AI related to content creation is estimated at $3.7 billion in 2024, with an expected average annual growth of 31.8% until 2030, reaching $19.6 billion (Grand View Research, 2025). To better understand the framework of this ecosystem, research work (Butraud-Assathian, 2026) established an original database based on Crunchbase². This generalist database shows a set of 89,061 AI-related companies operating at different levels of the value chain: development of generation models, provision of APIs, inference services, and creation and dissemination tools. To map this ecosystem, the 89,061 companies were classified using a text factorization algorithm on the descriptions of these companies, grouping them based on their similarity. The market was thus structured into 32 thematic clusters. In order to isolate specifically the actors in the manufacture of synthetic content, 5 clusters (out of the 32) were extracted: Gaming (video games), media and entertainment, design and visualization, virtual assistants (XR), and marketing. These 5 clusters gather 7,942 companies distributed across 112 countries. Content generation thus represents approximately 10% of the global AI ecosystem in terms of number of companies.

Geographically, in terms of the overall number of content generation companies, observations confirm an overwhelming domination of the United States, which accounts for 3,022 companies dedicated to synthetic content compared to 1,371 for Europe. The Asian continent stands out as a secondary hub, led by India with 460 companies, followed by Japan (333 companies) and China (225 companies).

The research conducted highlights differentiated funding across clusters. Although content generation enjoys overwhelming media visibility, it is structurally underfunded compared to the rest of AI. As an example, the median funding for a US synthetic content generation company drops to $2.5 million, compared to $3.75 million for other AI companies. Companies thus seem more cautious to invest in content generation than in the rest of AI (agrifood, medical devices, training, data analysis, etc.).

In their interactions with the cultural industries, we find both foundation model developers who themselves offer consumer tools and services, and other companies that build on foundation models developed by the former to offer their own services. Some of these companies focus on the consumer market with simple, automated interfaces. Others position themselves in the professional market by developing services for studios, platforms, or agencies. Specialization hubs are emerging: This development, which is taking place at a lightning-fast pace, seems, however, to rely on practices that lack transparency, and these companies are widely suspected of having trained their models on a large number of copyright-protected works. Despite everything, content generation tools are thriving. Individuals generate synthetic productions of all kinds, circulate them, and share them on various platforms, whether social networks, Digital Service Providers like Spotify or Deezer for music, or UGC platforms like TikTok or YouTube for videos.

• Text generation: ChatGPT (OpenAI's tool), Le Chat (Mistral's tool), Claude (Anthropic's tool).
• Image and video generation: Midjourney, Runway, DALL-E, Pika, Luma, Synthesia, Rephrase, Sora (OpenAI's tool), Hedra, Nano banana Pro (Gemini's tool - Google), Veo 3 (Google), Vibes (Meta's tool).
• Music generation: Suno, Udio, Mubert, AIVA, Boomy.
• Voice cloning and speech synthesis: ElevenLabs, (n.d.), Respeecher, Wellsaid, Voicemaker.
• Post-production and dubbing tools: Flawless, Papercut, HeyGen, Deepdub.

A press release from Deezer announces that in 2026, 60,000 AI-generated tracks were uploaded on the platform each day, representing 39% of all daily uploads. In other words, more than a third of the incoming tracks feeding the catalog no longer come from a traditional human creative process, but from automated generation services. A study also revealed in 2025 that 21 to 33% of content broadcast on YouTube consists of AI-generated videos. Furthermore, in terms of shorts, AI videos (AI Brainrot) represent 33% of the first 500 short YouTube videos that appear in a new user's feed.

This multiplication of synthetic content has significant consequences both for traditional human creation in the music and audiovisual industries and also for the public’s trust in what they see, watch or listen to.

3.2 Weakening of human creation and non-sustainability of synthetic production models

The observed growth of the markets does not always benefit the upstream of the cultural sector, that is to say, the creators. Certain professions already feel globally threatened: translators, photojournalists, illustrators, extras (who serve as a gateway to the acting profession), designers, and graphic designers see their current practices challenged. Regarding writing professions, such as journalism, a study estimates that the jobs of reporters and journalists are among the most exposed to AI systems (Eloundou et al., 2023). Similarly, for performing artists, content reproducing all or part of the attributes of their personality and talent (appearance, voice, attitude, artistic style, etc.) can be mistakenly perceived as authentic.

These challenges are exacerbated by the ability of AI models to generate realistic but purely synthetic content, likely to be perceived as representing artists, authors, or celebrities without their consent. Any unauthorized use of the personality attributes of a performing artist to create AI-generated content thus has consequences for their reputation, by capturing the attention of fan communities, disrupting planned promotional campaigns, and generating revenue based on their notoriety, but without their knowledge. Increasingly realistic content therefore creates confusion in the public's mind regarding the official image of an artist, a group, or a brand by leading people to believe in the authenticity and actual attribution of the performances, which directly harms their reputation along with the associated economic consequences.

For certain professions, crowding-out effects are felt first through price competition, insofar as AI allows for faster production at a lower cost than humans. Over time, synthetic content (images, archive sounds, etc.), less expensive and faster to produce, devalues the work of human creators, leads to a drop in rates for creative services, and, ultimately, threatens certain occupations. Companies like Suno, Udio, or Mubert allow for the instant creation of royalty-free tracks, effectively replacing some composers with AI generators. This dynamic is no longer limited to independent startups; very large groups are now deploying their own generative creation tools (such as Google with Gemini), while others, conversely, are abandoning these market segments deemed unprofitable (see OpenAI's abandonment of the video generation tool Sora in March 2026). This situation directly raises the question of the viability of specializing in content generation and relates back to the relative caution of investors in this AI segment (see supra).

Crowding-out effects are also exerted through quantities. The overabundance of content generated by AI systems indeed risks saturating the market and, consequently, reducing the discoverability of human works. This is particularly the case with the proliferation of unauthorized voice clones, especially on YouTube and TikTok, which diverts attention from official recordings. This practice diverts public attention and, consequently, a portion of the revenues that could have benefited human creators (Josan, 2024).

Thus, on music streaming platforms, “ghost” artists are flooding in, with recordings entirely generated by AI under fictitious names, offering a semblance of authenticity and real identity (AI generation of album covers, fictitious bios, credible profile pictures, etc.). Revenues from ghost artists on music platforms (Deezer, Spotify, etc.) and audiovisual platforms (YouTube) have become a worrying phenomenon due to their scale and expansionist nature. The proliferation of ghost content is often accompanied by another phenomenon, that of “click farms” which attempt to simulate streams made by human listeners. Initially, this practice of click farms relied on precarious human labor (mainly in Southeast Asia) whose goal was to artificially inflate engagement (likes and views) to generate advertising revenue. Today, the model is becoming automated, and workers are being replaced by bots and autonomous agents.

An emblematic case illustrates the mechanics of this fraud. American musician Michael Smith was indicted in 2024 for organizing a massive royalty fraud; he used AI to create hundreds of thousands of tracks, then used robots (“bots”) (around 10,000 accounts) to generate approximately 661,440 streams per day, diverting over $10 million in royalties (Prada, 2024). The millions of dollars diverted were subtracted not from the platforms—at least not directly, as these practices erode the trust relationship with majors and artists—but from the global pool intended for human artists.

This case also illustrates the ease with which generative tools make it possible to industrialize music fraud, from the generation of “slop” content (low quality and produced in large quantities) to the creation of identities, all the way to artificial consumption by robots capable of simulating human behavior to bypass anti-fraud systems (creating pauses, playlists, etc.). However, 80% of detected fraud is located not in the top streams, but in a kind of “artificial long tail” (Centre national de la musique, 2023). Instead of propelling a single track to millions of streams, which would trigger alerts, it becomes more sensible to generate massive quantities of tracks streamed only a few thousand times, which overall has a massive crowding-out effect to the detriment of human creators. Even more recently, AI-generated content has appeared on the pages of genuine artists, on streaming platforms or, most often, on YouTube and TikTok, usurping the identity, name, image, and voice of these artists.

Beyond the immediate economic and social effects on the cultural sectors, Zolynski et al., (2026) emphasize that AI-generated creations also raise sustainability issues for generative AI models themselves. This is referred to as “model collapse” or “model degeneration”. The authors highlight that if AI systems gradually replace human creation with artificially generated content, they find themselves faced with an internal contradiction. Indeed, these models, initially trained on human works, end up being retrained on their own productions. This phenomenon thus leads to a model collapse. The quantity, diversity, and richness of content will undeniably degrade as training data itself becomes synthetic. By feeding on their own outputs, models will progressively lose their ability to produce varied and relevant content. The sustainability of AI performance depends on maintaining a continuous flow of original human works to feed the training databases. Protecting and valuing human creation is not only a cultural, economic, or social issue, but also a prerequisite for the future proper functioning of AI systems themselves.

3.3 A threatened trust economy

Researchers have empirically established that the trust a society inherits from previous generations predicts its long-term economic development by creating the conditions for sustained cooperation, long-term investments, and stable institutions (Algan and Cahuc, 2010). In societies where institutional trust is already precarious, destabilization can occur in a particularly rapid and acute manner. This macroeconomic analysis finds an equivalent at the sectoral level: trust in a country's cultural industry (its reputation, the authority of its creators, the credibility of its institutions) is a stock accumulated over several generations, whose erosion can be rapid but whose reconstitution is slow.

However, converging empirical data in recent years confirm the cognitive limits of the public in the face of synthetic content. The vast majority of listeners and viewers are unable to distinguish AI-generated content from that produced by humans, either by ear or with the naked eye. This finding is particularly clear in the musical and vocal domain, and the distinction is even more difficult to make when the tracks are human-AI co-creations (Rigole et al., 2025).

A survey conducted jointly by Deezer and the Ipsos Institute among 9,000 people in 8 countries shows that 97% of listeners fail to distinguish music entirely generated by AI from a human composition during a blind test. Academic studies have investigated the question of public recognition of AI-generated content. One of these consists of a survey conducted among 1,276 participants, who were tasked with distinguishing authentic media from synthetic media (images, audio, video, and audiovisual formats). The results show that individuals encounter significant difficulties in making this distinction (Cooke et al., 2024). The study also highlights that detection becomes even less reliable when the content includes hybrid elements, human faces, a foreign language, or a single sensory modality.

Similarly, a meta-analysis dedicated to human detection of deepfakes confirms this difficulty in recognizing synthetic content (Diel et al., 2024). By aggregating 137 statistical effects from 56 empirical studies conducted among more than 86,000 participants, the authors show that overall detection performance does not significantly exceed the level of chance: the average accuracy reaches around 55%, with variations depending on the type of media (slightly higher for audio and video, but close to chance for image and text). Overall, these works confirm that, without supporting devices, human ability to recognize synthetic content remains limited and fragile (Hashmi et al., 2024).

Perceptual blurring is therefore now established. The vast majority of the public fails to distinguish synthetic content from human content under ordinary reception conditions, and this inability worsens with the most recent AI models.

Furthermore, various works reveal a structural tension between declared norms and actual uses. On the one hand, opinion surveys reveal strong expectations regarding transparency (New Industry Focus, nd., Next, n.d.), and a persistent distrust of the quality and legitimacy of entirely synthetic productions (Deezer and Ipsos, 2025). In some cases, the simple fact of being informed that a work was produced using this technology tends to reduce both their appreciation of the work and their willingness to pay for it (Friedrichsen et al., 2026). On the other hand, behaviors observed in context and certain experimental studies suggest a pragmatic acceptance, or even a conditional valuation, particularly when the content sparks curiosity or triggers narrative transport (Bagratuni et al., 2025).

Finally, the central cognitive stake is not so much the specific content of each synthetic piece, but the systemic effect of uncertainty that its mere existence introduces into the information ecosystem. This transformation corresponds to the “impostor bias” (Casu et al., 2024). The simple existence of synthetic content alters the credibility heuristics that subjects mobilize in the face of synthetic content, establishing a generalized skepticism even toward human content (Weikmann et al., 2025). This bias is not an isolated error of judgment but a structural transformation of informational trust that affects the evaluation of evidence, the reception of messages, and the dynamics of persuasion. The central threat therefore lies less in their ability to make us believe in the virtual than in their ability to make us doubt the real—a major epistemic shift

Cardon (2026). concludes that what ensures the credibility of an image is no longer the image itself, but the person or institution that addresses it to us, with whom we enter into a reading contract of the image. In other words, epistemic authority shifts from content to the mediator, a displacement that could renew the role of recommendation.

4 Promoting human offerings: a new stake for recommendation to restore trust

In the face of the proliferation of AI-generated content, the AI Act (European Artificial Intelligence Regulation) provides for transparency obligations. Its Article 50.4 thus indicates that “providers of AI systems that generate synthetic content shall ensure that the outputs of the AI systems are marked in a machine-readable format and identifiable as having been generated or manipulated by an AI.” Article 50 therefore places these transparency obligations on AI providers, that is to say, upstream, at the time of synthetic content generation.

For platforms that make online cultural content available to the public, no detection obligations are prescribed. For a number of them, however, it is not legal obligations but the survival of their business model that is at stake and pushes them to deploy synthetic content detection techniques. For many of these platforms, synthetic content may appear, at least in the short term, positive by generating engagement and advertising revenue, whereas detection is a significant cost center. However, with the multiplication of synthetic content, these platforms face a form of generalized pollution of their services. While they may have sometimes been tempted to accommodate it, most are now conducting detection policies to consolidate their economic models. Indeed, in the longer term, detection appears to be the only path to restore an economy of trust for a public disturbed, as we have seen, by generalized uncertainty. Furthermore, the economic partners of online distributors—on whom the supply of their platforms depends—include creators and producers who are fighting for their survival, and the suffocation of human creation by synthetic pollution is unsustainable in the long run (see supra). Finally, advertisers, upon whom many services depend, demand guarantees of human audience, rather than streams and views generated by bots. This economic requirement for detection is, however, far from simple to implement.

4.1 Available techniques for detecting synthetic content

While humans are incapable of distinguishing synthetic content, machines appear to be much better at it. Two main categories of detection mechanisms are distinguished (for a review, see Farchy et al., 2026). On the one hand, identification (or downstream detection without watermarking), which refers to forensic analysis and seeks to determine the origin (généralement the generative model used) in the absence of any prior mark, relying solely on the characteristic artifacts left by the generator. On the other hand, detection when watermarking has been carried out upstream—watermarking refers to the proactive insertion of information signaling the AI origin of content by the system producing it—consists of verifying the presence or integrity of this mark in a given content (Serra et al., 2026).

4.1.1 Automatically distinguishing between synthetic and human content circulating downstream

In the absence of upstream watermarking, detection techniques will compare the statistical properties of human and synthetic content with the objective of identifying detectable artifacts in the latter.

However, technical tools for detecting synthetic content downstream, as it circulates, face a classic “cat-and-mouse” problem because they must constantly adapt to new AI models. Generation models progress faster than detection tools, and some researchers believe that precise and generalizable detection of AI-generated content might be fundamentally impossible as these models improve. The downstream deepfake detection strategy, in the absence of watermarking, is therefore still occasionally effective but seems unsustainable as models advance. Forensic detection techniques without watermarking tend to degrade as generators progress, which disqualifies them as a single solution but maintains them as a safety net for content not marked by AI providers. This is why watermarking, from the production stage, which the AI Act addresses, is expected to offer decisive perspectives.

4.1.2 Making synthetic content detectable from generation via upstream watermarking

Upstream techniques allow for the embedding of metadata or watermarks in generated images, making their human or artificial origin detectable by automated tools. The idea is to provide each content piece with an unforgeable digital birth certificate upon creation, rather than trying to detect synthetic content already in circulation. It should be noted that watermarking can be applied to synthetic content, but also to human, authentic content. A complementary strategy is therefore not to detect the synthetic, but to certify the authentic, by attaching a proof of provenance to the content from its creation.

However, watermarking is technically available and operational for images and audio in closed contexts with a cooperative producer. It is fragile for text and bypassable for open-source models. While the principle of preventive upstream watermarking is established, its effectiveness will depend on the standardization of watermarking and signaling methods. The risk is the emergence of a multiplicity of distinct systems developed by different actors, without interoperability or consistency. These upstream techniques therefore require the cooperation of AI providers to develop, which constitutes their main limit when faced with uncontrolled open-source models or in-house models. They require cooperation between developers of technological solutions, tool deployers, broadcasters of cultural content, and public institutions, notably to distribute the burden of the necessary investments.

The economic stakes of the concrete distribution of the compliance burden across the entire value chain are indeed considerable. In parallel with the development of the content generation market, a mirror market is emerging around defensive tools: automatic detection, cryptographic watermarking, provenance authentication, and usurpation monitoring. Estimates place this market at around $1.5 billion in 2025, with an expected growth rate of over 40% per year. The main players include specialized companies (Reality Defender, Sensity AI, Pindrop, Truepic, DeepMedia), services integrated by major platforms (Deezer Radar, YouTube Content ID adapted for AI, Meta watermark detection), and open standard initiatives (Coalition for Content Provenance and Authenticity – C2PA, Content Authenticity Initiative). This market remains for now mostly downstream (post-production reaction) rather than upstream (prevention at generation).

4.2 What new role for recommendation on online cultural offering services?

Once technical detection tools are put in place despite all the difficulties we have just mentioned, platforms have a role of transparency and information toward the public. Some companies have chosen to take measures and communicate their approach to the public.

In the music industry, the strategies of streaming players are diverse. Deezer has developed and patented its own in-house technology capable of detecting content from Suno, Udio, and other major generators. The tool does not merely compare the audio to a database (classic “fingerprinting”) but analyzes the intrinsic structure of the signal to determine a probability of synthetic origin. Systematic labeling of detected tracks is carried out (the mention “AI” on certain tracks) (Afchar et al., 2024). Deezer also commercializes its “radar” tool to other organizations such as SACEM. Deezer has already excluded AI tracks from its editorial and algorithmic playlists. The tool thus aims to exclude them from recommendations and monetization, rather than to delete them. As soon as a track is identified as being entirely generated by an AI, it is neither integrated into editorial playlists nor pushed by recommendation systems.

Apple, which claims to be a premium brand, focuses on human curation to differentiate itself from the low-quality content proliferating elsewhere with AI. Spotify, for its part, plays the card of market segmentation; it has industrialized specific spam filters for AI-generated tracks by acquiring specialized startups (Spotify, n.d.), and in parallel offers in 2026 a premium subscription of high technical quality and guaranteed genuinely human origin.

In this new universe saturated with synthetic content, the role of recommendation could evolve accordingly. It is no longer just a matter of highlighting a human work among others, but also of adopting methods for promoting human content and ensuring transparency about what is seen or heard.

5 Conclusion

With the tsunami of possibilities offered by GenAI, value is shifting from an economy of capturing public attention to an economy of restoring trust in a world saturated with synthetic content.

To restore this trust, a first step of transparency toward the public seems indispensable. However, the concrete ways in which information about the AI origin of content should reach the end user are far from clear. The consultation conducted by the European Commission with stakeholders in 2025 devotes a substantial part of its proposals to this aspect. Moreover, transparency on how the content was created, whether by AI or not, cannot constitute a sufficient proxy to restore public trust. Knowing whether content is AI-generated is not equivalent to knowing whether it is reliable and trustworthy. Although these two questions may overlap, not all content originating from generative AI is deceptive or problematic, and not all deceptive or problematic content is generated by AI. This invites us to abandon the idea that transparency could be guaranteed by a single technical system (Zhao et al., 2025): it is the coherence of an ecosystem—technical, legal, economic, and educational—that will condition the ability to restore public trust.

Furthermore, for online cultural content services, the survival of their business model requires not only restoring public trust but also not mortgaging³ the future of human creation. In the long term, as we have seen, only economic models that allow human production to exist are indeed likely to prevent the degeneration of the GenAI models at the origin of synthetic content.

In this new world, the future will undoubtedly involve a different way of approaching recommendation. Indeed, the question is no longer just to guide the user in their choice toward a specific piece of content, but also to inform and orient them toward human content in the face of the abundance of synthetic content.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Translation.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Antoine Henry, Université de Lille, France

Reviewed by: Jannick Kirk Sørensen, Aalborg University, Denmark

Manuel Damasio, Lusofona University, Portugal

Footnotes

¹Farchy, J., Butraud-Assathian, A., and Laqueyrerie, L. (2026). Positionnement concurrentiel de plateformes de vidéo à la demande sur le marché français. Unpublished manuscript.

²An international database that compiles detailed information on companies (industry sector, funding, acquisitions, location, etc.) and their economic ecosystems. Working paper.

³SPECTRA, Analysis of the Stakeholder Consultation on Transparency Requirements for Certain AI Systems under Article 50 AI Act.