A. Introduction
There are two opaque ways in which the internet’s ecosystem—encompassing commercial websites owned by small- and medium-sized enterprises, citizen journalists, weblogs, and social media platforms—is being undermined by the diversion of traffic, the lifeblood of digital publication, away from its original content.
At least at first sight, generative search results unintentionally do this, while shadow banning does so intentionally. This is of great importance, because websites and social media posts are the modern version of the public sphere, decentralized spaces where individuals can share news, or expressive information, and make a living, which should be able to counterbalance to centralized commercial news and commodified information. The problem will be exacerbated by a deluge of synthetic content on websitesFootnote 1 and social media posts, which will dilute further the human-created content.
Jürgen Habermas described how the public sphere emerged, separate from the private and state spheres, in the 17th and 18th centuries.Footnote 2 Mercantilism and early forms of capitalism contributed to the rise of the bourgeois class, which, in turn, played a critical role in fostering the Enlightenment and the development of the bourgeois public sphere.Footnote 3 Within this sphere, salons and coffee houses served as key institutions where members of the bourgeoisie gathered to engage in rational-critical debate about politics, culture, and society.Footnote 4 Mass media was becoming commodified and transformed journalism into a market-driven serviceFootnote 5 prone to manipulation,Footnote 6 causing an erosion of the true public debateFootnote 7 and changing the democratic authenticity of public opinion. This was further aggravated by the concentration of media ownership.Footnote 8
In contrast, the advent of the internet, particularly weblogsFootnote 9 and later social media,Footnote 10 has been embraced as a means to democratize news and provide new opportunities for self-expression. One could argue that these websites have evolved into “the modern public square,” as the Supreme Court described social media in 2017 in Packingham v. North Carolina,Footnote 11 constituting “the most dynamic forum for speech in the United States.”Footnote 12 Anyone with a computer and an internet connection can create a website and share their thoughts, experiences, news, knowledge, and expertise at minimal cost, or earn a living. Yet today, generative search and shadow banning threaten to replicate and intensify the very dynamics that Habermas critiqued. Generative search abstracts and repackages the expressive contributions of individual creators, severing the link between speaker and audience, while shadow banning quietly suppresses visibility without procedural transparency. Both practices displace the deliberative public from its own communicative outputs, filtering and fragmenting discourse in ways that privatize attention and algorithmically gatekeep participation. Generative search, which induces internet users “to skip the links,” significantly decreases the traffic of visitors to the websites.Footnote 13 No visitors means no revenue stream based on advertisement, subscription, patronage, or crowdfunding. This is undermining the sustainability of these business models for websites, the very foundational real estate of the internet. This Article contends that these practices are not mere technical evolutions but structural threats that demand doctrinal, policy-based, and normative engagement. This Article interrogates how generative search and shadow banning—two opaque, algorithmically mediated practices—are reshaping the information architecture of the internet in ways that threaten the visibility of lawful expression, the sustainability of independent content creation, and the integrity of the digital public sphere. By analyzing these phenomena through the lenses of intellectual property, platform governance, and comparative regulatory frameworks, the Article aims to develop a normative and doctrinal response to the vanishing of source attribution and audience access in an AI-dominated search environment. It argues for a recalibration of copyright, tort, and administrative law to restore transparency, fairness, and pluralism in the evolving informational ecosystem.
This Article unfolds in two parts, mapping how generative technologies and algorithmic moderation disrupt the digital information ecosystem.
The first part, Sections B to D, deals with the impact of generative search, particularly the use of Retrieval-Augmented Generation (RAG), on the visibility and sustainability of web-based content. Section B, “From Riches to RAG,” charts the internet’s transformation from a decentralized network of independent content creators to a more centralized ecosystem dominated by a limited number of platforms wielding control over traffic flows and content distribution. Subsection I traces the historical trajectory of search engines, culminating in the rise of RAG systems; Subsection II focuses exclusively on the RAG framework itself; and Subsection III examines generative search as a salient subtype of RAG. Section C addresses the responses available to website holders confronting the disruptive effects of generative search. Subsection I outlines technological measures, including opt-out mechanisms; Subsection II examines the practical limitations and barriers to effective opt-out; and Subsection III analyzes Dow Jones & Company v. Perplexity AI, offering further legal avenues of redress: (a) copyright infringement; (b) plagiarism and semi-plagiarism as normative and reputational tools; (c) unfair competition; (d) unjust enrichment and negligence; (e) the potential of licensing arrangements; and (f) the prospective role of agentic AI in mitigation. Section D offers a synthesis of the first part’s findings.
The second part, Sections E to G, turns to the phenomenon of shadow banning and its destabilizing effects on the digital public sphere through covert modulation of social media visibility. Section E introduces the conceptual contours of shadow banning. In the absence of comparable U.S. regulation, Subsection I analyzes China’s and the European Union’s frameworks governing algorithmic dissuasion, while Subsection II distinguishes between content de-ranking and traffic reduction. Section F focuses on the algorithmic suppression of content potentially implicated in copyright disputes. Subsection I explores the actors and consequences involved in such de-ranking, including: (a) measures to constrain recommendation algorithms; and (b) the treatment of unauthorized yet lawful user-generated content. Subsection II assesses the relevance of China’s Regulation on Algorithmic Recommendations (RAR), and Subsection III evaluates the approach of the EU’s Artificial Intelligence Act (AIA) to shadow banning practices. Section G concludes the second part. Section H provides the overall conclusion of the Article.
“Don’t bite the hand that feeds you.”Footnote 14 This saying warns against harming the source of your benefits. Applied to generative search using RAG, it reflects the tension between these systems and the websites they rely on for content. While RAG-powered generative search use websites as their foundation, they reduce traffic by offering direct, synthesized answers, depriving sites of ad revenue, subscriptions, and visibility.
B. Riches to RAG
This section examines the devolution of the ecosystem of the internet from predominantly a decentralized system, where control and content were distributed across many independent servers and networks, to a centralized system, with a few dominant platforms and corporations controlling significant portions of traffic, content, and user interactions.
I. Historical Development in Search Technology
The first internet search engine, “Archie,” was created in 1989.Footnote 15 It indexed downloadable files across the internet. In 1991 “Jughead” followed the Gopher protocol,Footnote 16 but searched a single server at a time.Footnote 17 In 1992, “Veronica” was used to browse and index information in Gopher menu items.Footnote 18 In 1994, the search engines “WebCrawler”Footnote 19—and in 1995, “Lycos”Footnote 20—allowed users to search the full text of web pages. In 1994, Yahoo! began as a web directory, manually cataloging websites, organizing content by categories to automate the indexing process. In 1995, “AltaVista”Footnote 21 introduced automatic indexing techniques and allowed users to search using natural language.
Google introduced its PageRank algorithm in 1998. It ranked websites based on the number and quality of links pointing to them.Footnote 22 Google constantly tweaked its search algorithm to improve the quality of the search and to increase AdWord pay-per-click advertising results. In 2013, Google released “Hummingbird,” which signaled the beginning of semantic search.Footnote 23 Hummingbird focuses more on context than keywords alone. In 2015, “BrainRank”Footnote 24 was created. BrainRank takes factors such as the location of the searcher, personalization based on click-through rates, and word combinations into account, to better discern the intention of the user over time.
In 2018, Google enhanced search with “BERT,” which stands for Bidirectional Encoder Representations from Transformers, to understand the nuances of language, especially word order and prepositions.Footnote 25 In 2021, Google’s “MUM” (Multitask Unified Model) shifted search toward multimodal generative results, to understand language better by integrating information from different formats.Footnote 26
II. Retrieval-Augmented Generation (RAG)
In 2017, Anish Vaswani and others wrote “Attention Is All You Need.”Footnote 27 This Article introduced the transformer model, which has become the foundation for Natural Language Processing (NLP) and deep-learning architectures, including Large Language Models (LLMs). The LLMs have made the generative AI revolution possible.
Traditional search engines, based on extractive systems, provide a list of relevant links, while AI services, based on abstractive systems, provide synthesized results. The downside of these generated answers is that LLMs have a tendency to hallucinate, leak information, and present wrong answers with the same aplomb as a right answer.Footnote 28 Moreover, LLMs do not provide any source;Footnote 29 after training, they cannot be updated or expanded in a cost-efficient way.Footnote 30
On May 22, 2020, a consortium of researchers from Facebook AI Research, University College London, and New York University submitted the article “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.”Footnote 31 This foundational article proposed a Retrieval-Augmented Generation (RAG) system that combines retrieval and generation,Footnote 32 to avoid the above-mentioned problems of LLMs,Footnote 33 at least to a certain extent.Footnote 34 The answers of a RAG system are grounded to the content in the sources of an external database to which it has access, and subsequently the LLM synthesizes this information in a representable and presentable form. To illustrate this system: If a user enters the prompt in a RAG system: “What are the causes of climate change?” In the semantic search, the vector of that prompt will be compared to the vectors of the external documents and the most similar and thus relevant documents or snippets will be retrieved. For example: “A scientific article on greenhouse gases;” “A news article about deforestation;” and “A government report on industrial emissions.”
In the enhanced context phase, the system filters out less relevant information and formats the input for example as: “The following are the main causes of climate change based on retrieved documents: greenhouse gas emissions, deforestation, and industrial activities. Use this information to answer the query.” After the semantic search and the enhanced context, the LLM is able to generate a coherent response: “Climate change is primarily caused by the accumulation of greenhouse gases, deforestation, and industrial emissions, which increase the Earth’s temperature.”
The RAG programmer inserts a system-level instruction—effectively a prompt that precedes the user’s promptFootnote 35—directing the model to prioritize this information. The user’s query (prompt) is converted into a vector, which is a numerical representation.Footnote 36 The vector captures the semantic meaning of the query, rather than just its surface-level keywords.Footnote 37 After a user’s query, the vector of the query is then compared to the vectors of the content in the database to identify which parts of the content are most similar in meaning to the query (semantic search).Footnote 38 The weights in the vectors are not adjusted, but are frozen embeddings. Subsequently, the most relevant parts of the content, those with the highest similarity scores, are retrieved and passed on.Footnote 39 Then, the information from the retrieved documents will be formatted and curated into a context that the LLM can use (enhanced context).Footnote 40 Finally, the LLM will generate an answer, based on the user’s prompt.Footnote 41
III. Generative Search
Imagine a bustling library where each book represents a website and visitors come to browse, research, and take notes. Initially, people wander through the library’s aisles, flipping through books to find information they need. One day, a librarian sets up a desk at the entrance with a sign: “Ask me anything.” Visitors can pose their questions, and the librarian provides a list of books and their locations. Over time, the librarian becomes more advanced. Instead of merely pointing visitors to books, he summarizes the content directly at the desk. Visitors still receive a list of book references, but most are less inclined to visit the shelves themselves.
Ultimately, the librarian evolves into an omniscient entity. By leveraging knowledge gained from scanning every book in the library—and even external libraries—the librarian can answer any question instantly, without referencing specific books or shelves. Visitors stop entering the library altogether, as the librarian now fulfills all their needs at the entrance.
Similarly, generative search diminishes traffic to websites by providing users with answers directly, rendering the original sources less visible and visited.
A similar development one can observe, between the ranked list of URLs of a conventional search engine, the generative search results of Google, Baidu, and Perplexity AI, and generative results of a generative AI-search result. More positively, generative search has broken up the hegemony of Google as the search champion. This provides chances for other providers.Footnote 42
A special application of RAG is generative search. Generative search results from Google, Perplexity AI, or Baidu do not merely display ranked links, but synthesize information across many sources and generate coherent, detailed responses, addressing complex queries in a conversational manner. “[T]raditional search engines that simply provide hyperlinks promote merely the discovery of copyrighted content, and not its substitution.”Footnote 43
Both Google and Bing have integrated AI to allow users to engage in multi-step conversations with the search engine, refining and narrowing down their queries. This conversational search experience, powered by natural language understanding, is a key aspect of this generative search era.
Google’s generative search results of “Google AI Overviews” refer to the integration of AI-powered language models, like Google’s LaMDA or the integration of technologies similar to OpenAI’s GPT-4, into its search engine. The results have been sometimes less than perfect.Footnote 44
SearchGPT is a temporary prototype of new AI search features that aims to give rapid answers from relevant sources.Footnote 45 SearchGPT is designed to help users connect with publishers by prominently citing and linking to them in searches, with in-line, named attribution and links so users know where information is coming from and can quickly engage with even more results in a sidebar with source links. OpenAI is planning to integrate the best of these features into ChatGPT.
Initially, on May 10, 2023, users could start using Google Search Generation Experience (SGE) by activating it via Google Labs.Footnote 46 Individual Google Search users could activate SGE if they choose to do so.Footnote 47
A year later, Google launched AI Overviews in the U.S.Footnote 48 Three months later, they announced that AI Overviews was rolled out to six more countries: Brazil, India, Indonesia, Japan, Mexico, and the UK.Footnote 49 On October 28, Google provided AI Overviews to over 100 countries.Footnote 50
In August 2022, Perplexity AI started as a Discord bot connected to Bing; with 15 million active users, it is already valued 9 billion US dollars.Footnote 51 After allegations of plagiarism,Footnote 52 Perplexity AI hedged its bets and started to conclude licensing agreements with publishers, such as Time, Der Spiegel, and Fortune.Footnote 53 Nevertheless, Perplexity AI was sued by Dow Jones & Company, who described Perplexity’s business model as “usurp[ing] content creators’ monetization opportunities for itself.”Footnote 54
C. Technological and Legal Measures Against RAG
This section explores the combined use of technological and legal measures to address the challenges posed by RAG systems. As RAG models increasingly rely on web content for training and output generation, website holders face significant risks, including copyright infringement, loss of revenue, and diminished traffic. This section outlines strategies that content providers and website holders can employ to protect their interests, emphasizing the need for targeted technical tools and enforceable legal frameworks. By integrating these measures, stakeholders can establish clear boundaries for permissible use while mitigating the exploitation of their content by RAG systems.
I. Technological Measures Opting Out
Website holders who do not want their information to be used by RAG models often still want to be indexed by search engines, and therefore hesitate to block them. If they were to do so, their findability would decrease, leading to a significant drop in website traffic, which would adversely affect their advertising-derived income. Generative search is separate from training Large Language Models (LLMs). Therefore, websites can surface in search results even if they opted out of generative AI training.Footnote 55 Most websites have been dependent on Google Search for their traffic. Therefore, in any opt-out method, a distinction should be made between blocking spiders for training LLMs and being ranked in Google Search.
If a website holder wants her website to be indexed and ranked by search engines but not used by RAG models, she could apply a combination of technical and legal methods.
The Terms of Use can make the legal boundaries of content providers clear: for example, that scraping for LLM training or retrieving data for RAG models is not allowed. The problem, however, is that bots do not have the semantic understanding of the Terms of Use text yet. Even if they had this, then the implementation thereof depends on whether the bot is programmed ethically or maliciously.
The “robots.txt” file can guide search engine crawlers to the parts of the site that can be indexed. By disallowing certain sections or pages in the “robots.txt” file, one can prevent search engines from crawling or indexing certain content. However, RAG models may ignore these rules,Footnote 56 so this approach alone would not fully prevent them from accessing the website’s data. In addition, one can use X-Robots-Tags, which are directives in Hypertext Transfer Protocol (HTTP), such as “noindex” or “nofollow,” that allows one to control the indexing of non-Hypertext Markup Language (HTML) content such as PDFs, images, or other file formats.Footnote 57
One can also prevent RAG systems from accessing or interacting with the website’s Application Programming Interface (API) or data endpoints.Footnote 58 Other ways include limiting the number of requests per IP address, user, or token within a certain timeframe. This would prevent bots or automated systems like RAG from sending high-frequency requests to harvest data.Footnote 59 It also prevents Internet Protocol addresses (IP) from white-listing,Footnote 60 or black-listing,Footnote 61 blocking known IPs associated with unwanted bots or scraping activities. By requiring CAPTCHA (“Completely Automated Public Turing test to tell Computers and Humans Apart”) or other challenge-response authentication systems, automated access to APIs can be prevented,Footnote 62 while API usage can monitor patterns through logging and analytics.Footnote 63
Website holders can use encryption of sensitive data in transit with Hypertext Transfer Protocol Secure (HTTPS),Footnote 64 and tokenization, a process that replaces sensitive or important data with tokens that are meaningless on their own but can be mapped back to the original data only by an authorized system with the proper mechanism.Footnote 65
Website holders can serve dynamic content that changes frequently or requires interaction, making it harder for RAG to retrieve and process the data.Footnote 66 They can also watermark and embed hidden metadata or other content tracking systems that can identify content when it is used in RAG models.Footnote 67
RAG system providers can prevent liability for the deployment of infringement of copyrighted works by concluding licensing deals with content providers, and respecting the websites Terms of Use and technological protection measures.
Copyright holders should be able to send platforms that use a RAG system, which is infringing their copyrighted works, a cease-and-desist notice, specifying the material that is infringing and providing proof of ownership. AI service providers have a published process for handling intellectual property complaints, similar to a notice-and-takedown procedure of online service providers, under the Digital Millennium Copyright Act of 1998.Footnote 68 Outside of the U.S. similar systems are being used.Footnote 69 Service providers that use a RAG system, might file a counter-notice if they believe their use is fair use.Footnote 70
II. Legal Claims Against Generative Search
In October 2024, Dow Jones & Co. sued Perplexity AI, Inc. for the following claims:Footnote 71 copyright infringement,Footnote 72 for ingesting plaintiffs’ copyrighted works to feed its RAG Index (count one); and to generate outputs for user queries (count two).
Count three refers to the claim of false designation of origin and dilution of plaintiffs’ trademarks,Footnote 73 in case of hallucinated output after requests by users to provide for example the main news at the Wall Street Journal (owned by Dow Jones & Co.).
Subsection 1 goes deeper into the claim of copyright infringement and looks at fair use. Subsection 2 deals with the non-legal norms against plagiarism and introduces “semi-plagiarism.” Subsection 3 provides an additional claim, overseen by the plaintiffs, of unfair competition, where plagiarism and semi-plagiarism can play an auxiliary role.
Subsection 4 deals with the possible claims of unjust enrichment and negligence. The more generative search providers are being sued, the greater the interest for them to preempt these costly and time-consuming events; therefore Subsection 5 deals with licensing agreements. Finally, Subsection 6 provides a possible solution: agentic AIs, almost like a deus ex machina.
1. Copyright Infringement
To see whether there is copyright infringement, there should be substantial similarity and access.Footnote 74 Substantial similarity can be assessed at different abstraction levels (chapters, sections, paragraphs, sentences, and words, letters). Changing each word with a synonym, does not necessarily immunize one from copyright infringement. For example, if one is copying details of a plot, it could still be copyright infringement.Footnote 75 And even though ideas are non-protectable under copyright, if they are ever more refined, there will come a moment when these ideas become expressions.Footnote 76 In 1985, the Supreme Court in Harper & Row v. Nation Enterprises held that The Nation’s use of the paraphrased material still violated copyright because the essence of the copyrighted work—the “heart” of the expression—was taken.Footnote 77
To paraphrase original human-created content does not necessarily indemnify a RAG from copyright infringement.Footnote 78
Dow Jones & Co. argued because of the grand scale reproduction and/or derivative content paraphrased or not, “[t]he illegality of this massive copyright violation at the input stage does not depend on whether the particular outputs of Perplexity’s so-called ‘answer engine’ are sufficiently similar in each instance to the copyrighted works of Plaintiffs as to constitute identical reproductions of those works.”Footnote 79
Generative search companies argue that fair use should be used as an affirmative defense against copyright infringement. Let us compare this with an analogous scenario similar to Google Overviews. In 1977, the plaintiff in Wainwright Securities v Wall Street Transcript, was providing reports on the stock market and the defendant used some opinions on certain sectors and stocks from the Wainwright report.Footnote 80 The question was whether the Wall Street Transcript’s abstracts were fair use.Footnote 81 The fair use test consists of four elements that a court should consider;Footnote 82 the elements are not exhaustive and need not be applied mechanistically.Footnote 83
Judge Lasker of the Southern District of New York (SDNY) found that Wall Street Transcript’s abstracts did not constitute fair use because: (1) the takings were “substantial in quality, and absolutely, if not relatively, substantial in quantity.” So, the third fair use factor favors the plaintiff; (2) publication of the abstracts probably reduced the value of the research reports; this refers to the fourth fair use factor, which favors the plaintiff as well.Footnote 84 The District Court for the SDNY decision was affirmed by the Second Circuit. The argument made by Wall Street Transcript was that the reports consisted of financial news events: facts. One can argue that one needs to make a distinction between news events and how these events were expressed. Nevertheless, since Salinger v. Colting in 2010, the Second Circuit has decided to abrogate the case and held that transformativeness should play a bigger role than derivativeness in such a case.Footnote 85
To see whether Google Overviews is fair use, let us also compare it with earlier Google services.
With Google News one can at least argue it is fair use if it only uses newspaper headings.Footnote 86 Because users will still have to click on the websites of the newspapers, they will be exposed to the advertisement—or in case of content behind a paywall, to be prompted to subscribe. The Google News case (AFP v. Google, Inc.) was settled between the parties.Footnote 87 In contrast, the Second Circuit decided that Google’s digitization of books, including orphan works and out-of-print works and display of “snippets” in the Google Books project, was fair use.Footnote 88 It held them to be transformative: One cannot enjoy snippets of a book in the same way as reading an e-book. However, the generative search result can be enjoyed as the text on the website itself, and is arguably therefore not fair use.
2. Plagiarism and Semi-Plagiarism
Some generative search results do not provide sources. This lack of attribution leads to plagiarism,Footnote 89 which Posner describes as academic fraud.Footnote 90 Fraud is a tort, and often a crime, but plagiarism is neither.Footnote 91 In case of most plagiarism two parties are harmed: the original author of the content who was deprived of his right of attribution, and the readers that were deceived about the source of the content. Posner points out that this only matters when the readers actually care about the deception.Footnote 92
Copyright only protects expressions of ideas, but plagiarism can also protect the right of the author to be recognized as the author of ideas, concepts or the reporter of facts.
Other generative search results, such as those of Google AI Overviews and Perplexity AI, do provide sources, but it is not always clear to which parts of the result they refer. This is what this author calls “semi-plagiarism.” In the case of Google AI Overviews, sources are provided more as references that are placed at the end of a text than as footnotes that refer to particular segments within the text. In the case of Perplexity AI, the generated answer is presented followed by the sources, presented as tiny buttons, which have a clickable number corresponding to a particular segment. It seems clear that the generative search provider does not want to encourage the user to actually click on these links, otherwise it would have made the text itself clickable. Users make use of this constellation, seemingly oblivious or indifferent to the implications to the sourced websites.
To avoid copyright issues, in January 2024, Aravind Srinivas, CEO of Perplexity AI, has made his intention clear to attribute every part of a generated answer.Footnote 93 However, Randall Lane, Chief Content Officer and Editor of Forbes, wrote about two journalists from Forbes that had been reporting on a developing story on former Google CEO Eric Schmidt’s secretive drone project. On June 6, they reported on ongoing testing of the drones in Silicon Valley. The very next day, Perplexity AI used its generated search to repackage the story and sent it to its subscribers via a mobile push notification. In addition, it created an AI-generated podcast based on the Forbes journalists’ work, giving credits to Forbes only via small “F” buttons.Footnote 94
3. Unfair Competition
International News Service v. Associated Press (1918) (INS v AP)Footnote 95 can shed light on the question whether generative search is fair competition. During WWI, the allied forces did not want to provide INS with news because its owner, William Randolph Hearst, favored Germany. Therefore, INS hired people to read East Coast newspapers that included AP news articles and subsequently to telegraph these to INS so that these parts could be published by its West Coast newspapers. Justice Oliver Wendell Holmes, dissenting, asserted that there was no property in news, because the news is merely facts,Footnote 96 and that there was no passing off (misrepresentation to deceive readers), as would be necessary to reach a finding of unfair competition.Footnote 97 Even though the INS might have given attribution to AP, the Supreme Court held that there was still misappropriation of “hot news,” and qualified it as unfair competition.Footnote 98 The court held that the investments in news-gathering need to be protected. This is a sweat of the brow argument, probably inspired by the realization that in news-gathering a certain degree of sharing among journalists should be allowed, but not to an excessive degree. Too much free-riding would make investigative journalism, which is labor intensive indeed, no longer worthwhile.Footnote 99 Therefore, the plagiarism or semi-plagiarism concomitant with generative search can be qualified as deceptive and a fortiori as a misrepresentation.
4. Unjust Enrichment and Tort of Negligence
Google Search has been the dominant intermediary between internet users and websites, almost worldwide.Footnote 100 For a decade, over 90 percent of internet users made use of Google Search.Footnote 101 Website holders, too, were dependent on Google Search for the vast majority of their traffic referrals.Footnote 102
Millions of websites holders have placed Google Display Ads on their websites to earn revenues.Footnote 103 These website holders are thus not only dependent on Google for traffic referrals but are also in contractual relation with Google.
Google gradually implemented generative search: on May 10, 2023, Google tested Google Search Generation Experience (SGE) via Google Labs. Individual Google Search users could activate SGE if they choose to do so.Footnote 104 A year later, Google launched SGE’s successor called AI Overviews in the U.S.Footnote 105 Three months later, Google announced that AI Overviews was rolled out to six more countries: Brazil, India, Indonesia, Japan, Mexico, and the UK.Footnote 106 On October 28, Google provided AI Overviews to over 100 countries.Footnote 107
Google’s implementation of generative search foreseeably harmed dependent companies by reducing internet users’ visits to their websites. Because “Google is doing the searching for you,”Footnote 108 users are less likely to click through to external sites, undermining website holders’ revenue streams, particularly those reliant on advertisements (mainly through Google) or paywalls.Footnote 109 This situation may give rise to legal claims based on restitution, unjust enrichment, and the tort of negligence, particularly in the context of a contractual relationship. Under the principle of unjust enrichment, “a person who has been unjustly enriched at the expense of another is required to make restitution.”Footnote 110 It can be argued that a search engine provider, such as Google, which implements generative search and sends web crawlers to index websites for its RAG systems instead of directing traffic to those sites, should have anticipated a substantial decline in traffic. Given Google’s dominant market share of approximately 90 percent, one could argue that it failed to exercise reasonable care in informing websites about the potential impact of these changes, thereby depriving them of crucial information and causing foreseeable harm.
5. Licensing Deals
Generative search deprives website holders of the revenue that can be generated from internet users visiting their websites: the fee for subscriptions to have access to their website, or the pay per clicks on advertisements on their website. However, generative search providers can compensate website holders for this deprivation and indemnify them against liability for copyright infringement. A current leader of AI services, OpenAI, has stated that it is committed to “a thriving ecosystem of publishers and creators.”Footnote 111 Since being sued by authorsFootnote 112 and publishers in 2023,Footnote 113 OpenAI has started to conclude licensing agreements with a slew of publishers.Footnote 114 The Copyright Clearance Center announced to include some AI rights in its Annual Copyright License for corporations.Footnote 115 Large publishing houses can negotiate remuneration in exchange for their information, leveraging their scale and resources to secure favorable agreements. In contrast, small- and medium-sized enterprises (SMEs) that operate websites often lack the bargaining power and infrastructure to achieve similar outcomes. This disparity puts SMEs at a significant disadvantage, as they struggle to monetize their content effectively or protect it from unauthorized use by generative search systems. Consequently, this dynamic accelerates market concentration, consolidating power and revenue among dominant players while further marginalizing smaller content creators. Such an imbalance not only stifles competition but also undermines the diversity and innovation that SMEs bring to the digital ecosystem. To address this, regulatory interventions may be necessary to ensure equitable opportunities for all content providers.
6. Agentic AI
Fenwick, Jurcys, and Loikkanen posited AI agents as a possible remedy against the traffic and revenue loss of websites:Footnote 116 Google’s AI agent could gather information from websites to generate search results and remunerate AI agents of these websites for this information.
While RAG is reactive, responding to input prompts with contextually relevant information retrieved from external sources, agentic AI is capable of autonomous decision-making and acting towards a goal, and typically can learn from experience.Footnote 117 RAG focuses on factual grounding, AI Agents provide planning capabilities and adaptability within complex environments.Footnote 118 Both systems can be merged into Agentic RAG, which combines RAG’s knowledge capabilities with AI agents’ decision-making skills.Footnote 119 Until this agentic-centric solution is implemented, generative search providers might expect lawsuits from authors, publishers and website holders.
D. Summary Generative Search
The internet’s shift from a decentralized, content-rich ecosystem to centralized generative platforms marks a pivotal change. While RAG-based search enhances user experience, it undermines website sustainability by reducing traffic. Website owners can adopt technological and legal strategies, but ongoing disputes, like Dow Jones & Co. v. Perplexity AI,Footnote 120 highlight unresolved issues such as copyright infringement and unfair competition. Licensing agreements and agentic AI offer potential solutions, though they remain underdeveloped.
Generative search contributes to the decline in website traffic, just as shadow banning can reduce visibility of individuals’ posts on social media. Both the impact of generative search on traffic and the rules governing shadow banning remain opaque and lack transparency.
“It exists” is an eternalist view; “It does not exist” is an annihilationist idea. Therefore the wise one should not have recourse to either existence or nonexistence (Madhyamaka).Footnote 121
Perhaps not the wise, but opportunistic platforms use shadow banning, operating in the algorithmic twilight zone between presenting search results and blocking them. This is how they can avoid the risk of being sued, and can optimize the copyright holders’ interests to directly or in a time-phased way de-rank or reduce traffic to unauthorized but possibly legal content into relative oblivion. This midway manipulation of traffic to suspected content is opaque to the uploader of content, and the general user, lacks any redress mechanism, and possibly chills the freedom to share transformed content that includes copyrighted works. Oftentimes, the platform has license conditions which would immunize it for de-ranking or reducing traffic, without the need to provide reasons, which is arguably incompatible with fundamental rights.
E. Shining a Light on Shadow Banning
To avoid liability for copyright infringement, platforms tend to err on the side of over-enforcement, often suppressing lawful but unauthorized user-generated content. This exposes them to potential legal claims from users whose expression is curtailed. Platforms increasingly deploy algorithmic enforcement mechanisms that operate with minimal transparency. This section examines how such systems implement stealth measures—including algorithmic dissuasion, de-ranking, and copyright strikes—to moderate content in ways that often escape public scrutiny and procedural challenge. It examines the legal and ethical challenges of these methods, particularly in balancing copyright enforcement with user rights, and fair use or exceptions and limitations to copyright. Focusing on China’s Regulation on Algorithmic Recommendations (RAR) and the EU’s Artificial Intelligence Act (AIA), this section highlights the impact of algorithm-driven moderation on content visibility, innovation, and free expression.
I. Algorithmic Dissuasion
How to make algorithmic dissuasion transparent and assailable with regard to uploaded works that allegedly infringe copyright, but might be fair use or fall within an exception or limitation? This section uses China’s Provisions on the Management of Algorithmic Recommendations in Internet Information Services, hereinafter Regulation on Algorithmic Recommendations (RAR),Footnote 122 as a case study. RAR is arguably the most far-reaching attempt to regulate aspects of recommendation algorithms.
1. Reining in Recommendation Algorithms
China’s copyright system—which is a hybrid of exceptions and limitations, the three-step test, and fair use principles—provides an example of a system within which unauthorized works uploaded on a platform are not always copyright infringements. China’s third amendment to the Copyright Law (2020) provides 13 limitations,Footnote 123 each subject to the three-step test of Article 24(1). The three-step test, based on Article 13 Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPs), which is an extension of Article 9.2 Berne Convention for the Protection of Literary and Artistic Works,Footnote 124 includes three open norms: “Members shall confine limitations or exceptions to exclusive rights to [1] certain special cases [2] which do not conflict with a normal exploitation of the work and [3] do not unreasonably prejudice the legitimate interests of the right holder.”Footnote 125
Article 21(1) number 13 of the Copyright Law of the PRC states: “Other circumstances provided for by laws and administrative regulations.” However, since 2004 at least some Chinese courtsFootnote 126 have been willing to apply fair use principles,Footnote 127 despite the lack of fair use in the Copyright Law of 2010Footnote 128 (and 2020). In 2011, the Supreme People’s Court issued an opinionFootnote 129 where in the “special case” of uses that promote innovation and business development, a more flexible limitations regime is admissible, if it comports with four fair-use factors,Footnote 130 inspired by Section 107 of the US Copyright Law.Footnote 131
Instead of blocking or deleting content that is not straightforward, counterfeit, or pirated, a platform can instruct its algorithm to allow uploading content but label it as suspect. Subsequently, the algorithm of the platform will put the search or feed results directly or after a certain period or in certain geographical areas as nearly unfindable by squeezing the traffic to the content.
2. Unauthorized but Legal User-Generated Content
This article is investigating the legitimacy and chilling effect of this practice in light of RARFootnote 132 and AIA,Footnote 133 and takes user-generated content that was created unauthorizedly but possibly legal if it did not infringe copyright as a case study. The results of this case study are to a large extent also applicable to content that includes one or more unauthorized but possibly legal trademark—but that is outside the scope of this Article.
Article 2 RAR provides examples of recommendation algorithm technologies that can generate and synthesize individualized pushing of content; refine the sequence of content (feed) and search filtering.Footnote 134
The recommendation algorithms of platforms such as Google, YouTube, TikTok/Douyin, et cetera, extract a fraction of content and rank it, based on the interactions of the user on that platform and possibly on several other sites as well. The algorithms learn to read the content preferences and usage patterns, combined with the location and time of users. The resolution of the user profiles is becoming ever sharper. Paraphrasing Yanis Varoufakis, we are training the algorithm so that the algorithm can train us.Footnote 135
Floridi has pointed out five ethical principles for algorithms; “beneficence, non-maleficence, autonomy, justice, and explicability.”Footnote 136 Neuwirth demonstrates that these general ethical standards are not met with concrete safeguards against different forms of manipulation, for instance regarding subliminal AI systems.Footnote 137
In the same vein, this Article asserts that there are neither concrete safeguards yet for exceptions and limitations in regard to the automatic content recognition algorithms, nor for de-ranking or the reduction of traffic.
Next to the twilight zone, as mentioned above caused by uncertain algorithmic moderation, one can argue that there is a second twilight zone: the automatic content recognition algorithm. If it labels content as unauthorized, it does not know whether it is legal and thus non-infringing. Automatic content recognition can filter copyright infringements at the moment of uploading or thereafter. These filtering algorithms, such as YouTube Content ID, are supposed to account for copyright exceptions and limitations in civil law jurisdictions and for fair use in common law countries—also recognized to some extent in China—but in practice they often fail to do so.Footnote 138 In the case of copyrighted content, the automatic content recognition algorithms of platforms in at least the U.S. and EU have transformed from under-enforcers with false negatives to over-enforcers with false positives,Footnote 139 to steer clear of the chance of being sued by copyright holders. In the case of false negatives, online service providers were unable or unwilling to take down unauthorized copyrighted works. The Viacom Intern., Inc. v. YouTube, Inc. case, heard by the U.S. Second Circuit Court of Appeals, arguably demonstrates willful blindness.Footnote 140 The founders of YouTube were taking their chances, after a notice-and-takedown request by the copyright holders, by failing to remove uploaded, copyright-infringing works which were increasing traffic to the site.Footnote 141
Since the advent of technologies that more accurately filter unauthorized copyrighted works, both overenforcement and underperformance have occurred depending on the size of the online service provider. Bar-Ziv and Elkin-Koren point out that copyright holders normally focus their enforcement strategies on global intermediaries instead of local platforms.Footnote 142 This can lead to overenforcement, for the former, and underenforcement, for the latter, in the same time period. In the case of false positives, the online service providers were unable or unwilling after a notice-and-takedown request to make sure that the works were indeed infringing copyright by assessing whether the works were in the public domain or subject to limitations of fair use. The Ninth Circuit case Lenz v. Universal Music Corp. exemplifies this overprotection, where YouTube took down a homemade video of toddlers running around a kitchen because Prince’s “Let’s Go Crazy” was barely perceptible in the background.Footnote 143
In addition to the deployment of filter algorithms to enforce against copyright infringements, search engines and platforms started to use algorithms to provide users with personalized recommendations. In 2003, Amazon made clear it used a recommendation algorithm for its e-commerce site.Footnote 144 The recommendation algorithm, owned by China’s ByteDance Ltd., has garnered significant attention not only for potential national security risks and its susceptibility to foreign adversaries,Footnote 145 but also for its extraordinary effectiveness, which has been criticized for fostering user addiction.Footnote 146
II. De-Ranking or Reduction of Traffic
Most internet users will focus on the first page of search results, and few users dive any deeper. With feeds of video or music-sharing sites this often happens as well, although many users scroll a bit more. If one scrolls down, new feed results will load, again and again. Just as each search result, every feed will be fed by a recommendation algorithm, based on the user’s historic profile that is updated with every click. The direct visibility of a search result using Google or Baidu is imperative for the chances that a user will find it. A fortiori, at least since 2012, Alphabet’s Google has been penalizing sites that infringe copyright by delisting them from their index or downranking them and making them practically unfindable,Footnote 147 a practice that continued in 2023Footnote 148 and 2024.Footnote 149 The recommendation algorithms of Alphabet’s YouTube and X Corp’s X work in a similar way. These intentional reductions in visibility are comparable to shadow banning in content moderation.Footnote 150 The algorithm that dictates the degree of visibility of content on YouTube is determined by factors such as the click-through rates and average view duration, but also by whether works are unauthorized and whether there are copyright claims or strikes against the uploader.
III. Copyright Claim and Copyright Strike
YouTube Content ID lets the copyright holder claim the content. Instead of submitting a notice and take-down request, he or she can track the video’s performance, block it on a country-by-country basis, or choose to monetize the video by placing ads on it. During the “tracking performance” phase, “[t]he claim may affect [the views of the uploader’s content] as the [copyright] owner may restrict the video from appearing on certain websites, devices, or even various countries.”Footnote 151
YouTube does something like a soft ban. Your video’s [sic] lose rankings on search terms or sometimes even do not show up on search results. There is no proven way to sort this out. You could try changing the title and headlines of the videos and see if it starts appearing on search results again or just wait it out and see if the views increase after two to three weeks.Footnote 152
A copyright strike is when YouTube receives a complete and valid legal takedown request from the copyright holder,Footnote 153 then takes down the video to comply with copyright law. After three strikes, the account and any associated channels are subject to termination, and all videos that were uploaded will be removed.
After the first strike, the uploader needs to go through an online copyright course. He or she can resolve the copyright strike by waiting for it to expire after 90 days, contacting the owner and convincing him or her to withdraw the takedown request, or submitting a counternotification opposing the claim that there is copyright infringement. Regarding the lattermost, for example, the uploader can claim that the work is in the public domain,Footnote 154 that the uploader is the copyright holder of the work, that the copyright holder has granted permission to use the work, that an exception or limitation applies, or that the use can be qualified as fair use/fair dealing.
After the first and second strike, YouTube may reduce the visibility of the channel and its videos in search results and reduce recommendations, which means fewer views and subscribers. If the uploader receives three strikes, his or her account is subject to termination to prevent repeat copyright infringement.
F. Algorithmic Dissuasion: De-Ranking Possible Copyright-Infringing Content into Oblivion
This section explores the interactions between algorithmic dissuasion and IP regulation across three distinct contexts: the interests of stakeholders, China’s regulatory framework, and the EU’s approach to algorithmic governance. Subsection I examines the key actors—copyright holders, platforms, and users—who are affected by algorithmic dissuasion, highlighting the strategic motivations and challenges surrounding the suppression or promotion of suspect content. Subsection II focuses on China’s RAR, emphasizing its balance of transparency, user rights, and enforcement while addressing IP concerns. Subsection III evaluates the EU’s AIA, assessing its risk-based regulatory framework and its implications for algorithmic content recognition in relation to fundamental rights and copyright protections. Together, these sections provide a comprehensive analysis of the legal and practical tensions arising from algorithmic decision-making in the context of IP.
I. Instigators, Perpetrators, and Victims of Algorithmic Dissuasion
To cast light on the method of fading out suspect content, this section will look at the effects on the respective protagonists of algorithmic dissuasion: copyright holders, platforms, and users.
There can be different reasons for the interest of copyright holders in reducing but not completely blocking unauthorized content, whether illegal or legal. In the interest of promotion of, for example, a movie, its copyright holder might prefer to allow at least for a certain period that a movie, or parts of that movie, can be shared by YouTube channels that cater to early adopters, whether these videos are infringing copyright or not. Platforms such as YouTube can recommend such videos to those YouTube users that meet for example the profile of early adopters. However, once the movie debuts in cinemas in a certain region, the copyright holder might want the traffic reduced to all unauthorized videos, including those that are legal.
Another reason is that copyright holders might not see it as conducive to alienate users that are also potential consumers by blocking their content, which could backfire on social media and become a public relations debacle.
A platform might be inclined to reduce the traffic to content that uses unauthorizedly copyrighted works. The platform’s automatic content recognition algorithm cannot yet apply the three-step test and concepts such as parody and transformativeness, so it cannot detect unauthorized but legal content. Platforms have only a nominal safe harbor and are still being sued by copyright holders. Even if the copyright holders can find the direct infringer and he or she is in the same jurisdiction, he or she might not have the financial means to pursue a lawsuit, while the platform has deep pockets.Footnote 155 By reducing traffic to unauthorized content, platforms reduce the chance of being sued.
With impunity, opportunistic platforms seem to be able to reduce the traffic to users that unauthorizedly uploaded a copyrighted work, no matter whether this was legal or not. Because of the opaqueness of this process, users can hardly detect that this is happening. And if they suspect it, they cannot prove it. Information about such events is also known as dark data.Footnote 156 A decrease in the number of visitors to a certain uploaded content on a platform can be due to several reasons related to the relevance ranking algorithm of the platform—which has a number of components, all trade secrets. There is obviously no redress mechanism for something that cannot be detected with certainty. In addition, there is a contractual relationship between the platform and the uploader, where the first limits its liability and the latter de facto agrees with any curatorial interventions,Footnote 157 but which arguably is not compatible with fundamental rights, such as the right of freedom of expression.
Limitation of liability
1. ERRORS, MISTAKES, OR INACCURACIES ON THE SERVICE;
4. ANY INTERRUPTION OR CESSATION OF THE SERVICE;
7. THE REMOVAL OR UNAVAILABILITY OF ANY CONTENT.Footnote 158
Since June 2022, 500 hours per minute were uploaded to YouTube.Footnote 159 Because the manual assessment of copyright infringement and exceptions or limitations of copyright is not scalable, platforms increasingly rely on automatic content recognition filters, such as YouTube’s Content ID.
According to YouTube’s Copyright Transparency Report,Footnote 160 Content ID generated 757.9 million unique claims or copyright removal requests during the first halve of 2022. In 0.5 percent of the cases, the claim was disputed. This amounts to 3,690,786 disputed cases.Footnote 161 It is likely that other users did not know how to counterclaim, or had no time or were fearful of being sued.Footnote 162 Therefore, the number of false positives is probably higher, especially because automatic content recognition filters at the moment are not able to take exceptions and limitations and the three-step test in a meaningful way into account.Footnote 163
The inner workings of machine learning, such as deep learning algorithms that are used for automatic content recognition filters, remain hidden. They are taking millions of data points as inputs and correlate specific data features to produce an output, which makes the creation of the results impossible to precisely explain and interpret. This black-box characteristic leads to a lack of algorithmic accountability.Footnote 164 This is amplified by what Cotter calls “black box gaslighting”—platforms that hide behind the inexplicability of the algorithm.Footnote 165 As mentioned above, this black-box characteristic is increased only in regard to de-ranking search results and recommendations and the decrease of traffic to the content. Moreover, there is no redress possible against these hidden measures.
II. China’s Regulation on Algorithmic Recommendations (RAR)
The Cyberspace Administration of China (CAC) promulgated the Provisions on the Management of Algorithmic Recommendations in Internet Information Services (Regulation on Algorithmic Recommendations (RAR)) on December 31, 2021, which came into effect on March 1, 2022.Footnote 166
Article 1 RAR makes clear that the provisions are not only on the basis of the Cybersecurity Law, Data Security Law, Personal Information Protection Law, and Measures on the Administration of Internet Information Services, but also other relevant laws and administrative regulations. The Copyright Law of 2020,Footnote 167 Trademark Law of 2019,Footnote 168 and Anti-Unfair Competition LawFootnote 169 are clearly within the RAR’s scope. Among other authorities, the State Administration for Market Regulation (SAMR) is also responsible for RAR,Footnote 170 just as it is for the China National Intellectual Property Administration (CNIPA).
Article 1 RAR also clarifies the functions of the provisions, which include “protect[ing] the lawful rights and interests of citizens, legal persons, and other organizations, and promot[ing] the healthy and orderly development of internet information services.” Article 6 RAR provides the same in negative form. Both provisions point to the need of recommendation algorithms to balance the rights and interests of users (citizens), platforms (legal persons), and copyright holders (legal persons and other organizations). Such a balance should help maintain a vibrant medium of expression while keeping intellectual property infringements to a minimum, and should also take copyright exceptions and limitations into account. Acting contrary to this would violate the freedom of expression as guaranteed under Article 35 Constitution of the PRC.Footnote 171
The RAR prescribes transparent algorithmsFootnote 172 and their optimization.Footnote 173 The problem is that automatic content recognition algorithms that are transparent are not able to take exceptions and limitations and the three-step test into account, and those that could, although that stage in the development of AI has not been reached yet or not yet implemented,are not transparent.
The transmission of unlawful information shall be immediately ceased once discovered, “and measures shall be employed to eliminate it or otherwise address it….”Footnote 174 The latter leaves open the possibility of de-ranking suspected content. The need for the resources and the setting up of enforcement mechanisms of copyright infringements, but also taking limitations and three-step test into account, can be derived by implication from Article 6(2) RAR.
The algorithm should not push infringing content.Footnote 175 Even though Article 17 RAR provides users the possibility to opt out of a personalized recommendation algorithm, a general recommendation algorithm that prohibits or discourages infringing content prevails. One can argue that the platform should inform the user who searches for an alleged infringing work, that it has been removed and the reasons why.Footnote 176
The platforms shall inform the users of which conduct is prohibited.Footnote 177 It would be logical that at least the uploader is informed that the traffic to some content is reduced and why, so that he or she can oppose this procedure, in accordance with RAR’s requirement that algorithms need to have convenient and effective portals for complaints and appeals.Footnote 178
The RAR is encouraging industrial organizations (platforms) to self-regulate.Footnote 179 But it seems that governmental supervision is necessary to enforce the RAR. Algorithms must be periodically checked,Footnote 180 if they are to be compatible with (or in conformity with) the RAR.
Algorithms need to set up convenient and effective portals for user appeals and public complaints or reports.Footnote 181
III. The EU’s Artificial Intelligence Act (AIA)
On August 1, 2024, the EU’s AIA came into effect.Footnote 182 It can be materially qualified as a product safety law addressed to those that use algorithms in the course of business, including platforms, to mitigate the risks for users and society at large, including copyright holders.
The proposed AIA uses a pyramid of risks, from unacceptable, to high, to low or minimal risk. Unacceptable risk, regulated by Title II, includes “AI systems used by public authorities, or on their behalf, for social scoring purposes.”Footnote 183 In contrast to the prohibition under the AIA, in China, a connection between IP infringements and social score is allowed and used to deter dishonest behavior in regard to IP.Footnote 184
High risk, regulated by Title III, is relevant to AI systems that create adverse impact on people’s safety or their fundamental rights.Footnote 185
One can argue that automatic content recognition filters should be qualified as high risk, because they may jeopardize fundamental rights such as the right of expression.Footnote 186 On the one hand, copyright in the EU has a catalogue of exceptions and limitations to copyright to promote this right of expression.Footnote 187 On the other hand, these exceptions and limitations are restrained to protect copyright holders.Footnote 188 Thus, automatic content recognition filters need to do a delicate balancing act.
So far, AI systems compare and identify uploaded works with works in a database that are allegedly eligible for copyright protection, and automatically either block those works or throttle the traffic to those works, without taking exceptions and limitations or fair use into account.Footnote 189 Despite significant numbers of false positives, automatic content recognition systems have not been qualified in the AIA as high risk. This contradicts with fundamental rights. Nevertheless, at the end of 2022, the Council of the EU called for promoting safe AI that respects fundamental rights,Footnote 190 without explicitly referring to the freedom of expressionFootnote 191 or the protection of property or intellectual property.Footnote 192
G. Summary Shadow Banning
Just like generative search, shadow banning on social media platforms is opaque and reduces traffic to an affected website. Shadow banning involves algorithmic manipulation, such as de-ranking or reducing visibility of content, often to avoid legal risks or visibly and directly enforce copyright. These actions, though intended to protect rights holders, lack transparency and can chill free expression.
H. Conclusion
The internet, once envisioned as a decentralized, democratized public sphere, is increasingly shaped by opaque practices like generative search and shadow banning, which undermine transparency, equity, and expression. Generative search diminishes website traffic by synthesizing content into direct answers, threatening the sustainability of independent content creators and small enterprises. Similarly, shadow banning on social media suppresses visibility and engagement, often through non-transparent algorithmic moderation aimed at avoiding legal risks or enforcing copyright. The public sphere, as conceptualized by Jürgen Habermas, runs the risk of being reduced to a simulation of dialogue structured by opaque commercial logic rather than democratic engagement.
Website holders are often uncoordinated, and they have been too dependent on one search provider for their traffic referrals. Generative search, driven by technological progress for consumer convenience and reduced transaction costs, may evolve into a system where only large publishing houses benefit from licensing deals, further concentrating the news market. In this regard too, antitrust law has failed to fulfill its role. Government regulations should mandate that platforms provide a simple and accessible option for websites to opt out of having their information being used by generative search systems and provide strict conditions for generative search providers to avoid copyright infringement, plagiarism, and semi-plagiarism. Similarly, individual social media users are often connected solely through the network of a dominant platform. To address this systemic power imbalance, government regulation is needed to mandate transparency and the use of explainable algorithms (XAI).Footnote 193
The EU and China both take a cautious path regarding algorithms in general and recommendation algorithms in particular, to safeguard their respective values. This is in stark contrast to the “[a]sk for forgiveness, not permission” approach, which might be conducive for technological innovation but might also lead to negative externalities. Paradoxically, this could mean that automatic content recognition filters that do take exceptions or limitations or fair use of copyright into account might be developed first in the U.S., if the pace of improvements of GPT-N and its ilk are any indication.
Legal and technological responses, such as China’s RAR and the EU’s AIA, attempt to address these challenges but remain insufficient in ensuring accountability, protecting fundamental rights, and fostering a fair ecosystem. Both practices reveal a troubling shift toward centralized control that prioritizes risk management and profit over open discourse, innovation, and the diversity of online expression. The stealthily applied algorithmic dissuasion via de-ranking and reduction in traffic of unauthorized but legal works can be qualified as extra-legal measures and is incompatible with the values of transparency and the fundamental rights of the users. In addition, one can argue that fundamental rights pre-empt licenses that are designed to immunize platforms against these measures.
To restore balance, platforms must adopt greater transparency and fairness in their algorithms, paired with robust redress mechanisms for affected users and creators. Licensing agreements, agentic AI, and clear regulatory frameworks can help mitigate harm, but only if designed to preserve the fundamental values of a democratic, decentralized internet. Without such measures, the modern “public square” risks becoming a tightly controlled and inaccessible space, further eroding its role as a platform for free expression and innovation.
Acknowledgements
The author declares none.
Competing Interests
The author declares none.
Funding Statement
No specific funding has been declared with respect to this Article.
Open access
