How Generative AI is Reshaping Recommendation Systems

Recommendation systems have integrated into various digital products we use daily. When we open YouTube, Bilibili, Douyin, Xiaohongshu, or other video and audio apps, the content on the homepage is curated. On shopping platforms like Amazon and Taobao, the products we see and their display order are usually not random. Our behaviors—searching, clicking, lingering, saving, and purchasing—continuously become new signals, helping recommendation systems predict what we might be interested in next.

The true complexity of recommendation systems lies not just in determining “what a user might like,” but also in how recommendation results change what users subsequently see, and how these new behaviors then become data for the next round of recommendations.

Over the past two decades, recommendation systems have evolved from collaborative filtering to deep neural networks, behavior sequences, and multimodal models, becoming increasingly adept at inferring what users might do next from complex behaviors. While these technologies have continuously advanced, a fundamental premise has long remained unchanged: systems primarily face an existing pool of movies, products, news, videos, or creator accounts, and judge which objects should be presented to specific users.

Generative AI is changing this mechanism in three ways: users can directly state their needs in natural language, AI assistants can integrate cross-application context within authorized limits, and systems can participate in organizing or even generating the final results beyond just retrieval and ranking. This changes not only the form of recommendation results but also how systems understand users and where that information can be obtained.

This article will distinguish which of these changes have entered products and which mainly remain in the research and prototype stages, and discuss how they impact user choice, content production, and information verification.

I. How Recommendation Systems Infer Preferences from Behavior #

Early recommendation systems solved a relatively intuitive problem. If two people have liked many of the same movies in the past, and one of them recently liked a new movie, the other might also be interested in it. Collaborative Filtering finds this relationship from the shared choices of a large number of users and items.

This method relies on accumulated interactions. New users have no historical records, and new items haven’t been clicked by anyone yet; this is the classic Cold Start problem in recommendation systems. As the number of content items and users on platforms continuously grows, relying solely on user-item similarity becomes increasingly difficult to describe complex and ever-changing interests.

With the advent of deep learning in recommendation systems, models began to automatically learn representations of users and items. What a user has watched or bought, what category a product belongs to, what scenes a video contains, what a title expresses, as well as the current time, device, and usage scenario, can all become inputs to the model. Models typically convert this information into embeddings, which are sets of numbers in a high-dimensional space; objects with similar interests or semantic relationships are usually closer in this space.

Modern large-scale recommendation systems also break down the process into multiple stages. Taking the Recall and Ranking architecture as an example, the recall stage first quickly filters hundreds or thousands of candidate objects from a vast content library, and then the ranking stage performs more refined scoring by combining user behavior, content attributes, current scenarios, and business constraints. YouTube’s 2016 deep recommendation system paper is an important representative of this type of industrial architecture [1]. The recall stage typically employs lighter models to quickly screen candidates, such as two-tower models that encode users and content separately; after the number of candidates is reduced, the ranking stage uses more complex models to integrate more features.

In recent years, recommendation models have begun to simultaneously process two types of information that were previously difficult to leverage: the sequence in which behaviors occur, and what the content itself contains. Sequential Recommendation no longer merely treats “what a user has watched” as a set but analyzes how behaviors are sequentially related. Someone who watches Japanese travel videos continuously in the evening and then searches for hotels might be planning a trip; if the same few clicks are spread over two years, their meaning is completely different. The Transformer’s self-attention mechanism can directly model the relationships between different positions in a behavior sequence, making it an important technical route in sequential recommendation [2].

Multimodal Recommendation allows systems to directly utilize information from text, images, videos, and audio. For products that have just launched and have few click records, visual and textual features can help the system determine their relationship with existing content and user interests [3].

These technologies expand the information available to models; however, user preferences are still primarily inferred from behavior. Actions such as clicks, dwell time, completion rate, searches, revisits, and purchases, which do not require active user ratings, are usually referred to as Implicit Feedback [4]. Likes, saves, and ratings are clearer forms of feedback, but the vast amount of data platforms acquire still stems from traces left by users during normal use.

However, behavior cannot be directly equated with preference. A person lingering on a video might be due to liking it, confusion, anger, or simply not swiping away in time. Models usually cannot directly know the reasons behind behavior and can only find statistical patterns from large amounts of historical data.

More importantly, the recommendation process also forms a Feedback Loop. Users can only react to content they have already seen, and the system uses these reactions as a basis for the next round of recommendations. When a certain type of content receives more exposure, it is more likely to get clicks and dwell time; these behaviors may then be interpreted as stronger interest, continuing to influence subsequent exposure. The resulting Exposure Bias means that content not shown rarely receives behavioral signals, making it harder to participate in subsequent recommendation competition [5].

Therefore, the “preference” observed by the system is not a fixed object independent of the recommendation process, waiting for the model to discover. User choices influence recommendations, and recommendations in turn determine what users have the opportunity to choose next. The data ultimately obtained by the platform is not a direct record of user preferences, but a result of the combined action of user behavior and platform exposure.

II. How Large Models Change the Input, Context, and Output of Recommendation Systems #

With the addition of large models, a new type of signal has been added to this loop: users can directly state their current needs in natural language.

When Large Language Models (LLM) are integrated into recommendation systems, users no longer need to repeatedly click, skip, or re-search to adjust recommendation results. When looking for a movie, a user can directly ask:

I want to watch a slow-paced movie, but not too depressing. No war themes, preferably under two hours.

If the results are not suitable, the user can continue to add:

That last search was for someone else, don’t count it as my long-term interest.

Natural language thus becomes a completely new preference signal. A 2023 study found that in near cold-start scenarios where user historical data is scarce, large language models that recommend solely based on natural language preference descriptions can achieve performance comparable to collaborative filtering methods based on item history [6]. This result stems from specific datasets and experimental conditions and does not imply that large models are universally superior to mature recommendation systems. Its significance lies in that users do not have to wait for the platform to accumulate a large number of clicks before they can express relatively complex needs.

Natural language provides richer information, but its reliability is not necessarily inherently superior to behavioral data. People’s stated preferences and actual choices are not always consistent. A more realistic system might use both user-expressed preferences and actual behaviors, rather than letting one completely replace the other.

From a system implementation perspective, large models are not integrated into recommendation systems in a single fixed way. In existing architectures, a Foundation Model can first understand product descriptions, reviews, and images, converting this unstructured content into relatively unified semantic representations, while recall and ranking are still performed by traditional recommendation models.

Large models can also serve as an interaction layer, first understanding the problem the user currently wishes to solve, and then letting the traditional system handle large-scale candidate sets. Large models excel at understanding complex language, while traditional retrieval systems are better suited for processing millions or even billions of candidate objects.

If authorized by the user and connected to external services, large models may also change the state of user information scattered across various platforms. Traditional recommendation systems primarily use data mastered by the platform itself; AI assistants, however, can connect multiple applications through data interfaces, placing relevant information within the same context. For example, travel videos, calendar itineraries, booking information in emails, and past consumption preferences can collectively help the system determine what the user currently needs. This does not mean that applications will automatically interoperate, nor will it necessarily form a permanently unified user profile. A more realistic scenario might be that an AI assistant temporarily integrates relevant information around specific tasks, within authorized limits; the content it can access remains restricted by data interfaces, identity associations, and privacy rules [7][8].

Taking another step forward, systems can directly participate in generating or rewriting the content users ultimately see. Here, we need to distinguish between two easily confused “generation” methods. In some research, models generate item identifiers or recommendation sequences, where the final objects still come from an existing candidate pool; when the system further generates new text, images, music, videos, or interactive characters, personalization is no longer just about ranking candidate objects, but begins to directly change the content users ultimately see. Existing LLM recommendation research has formed different paradigms such as discriminative recommendation and generative recommendation [9], and works like CALRec are also trying to generate recommendation results based on user behavior sequences [10], but this is not equivalent to directly generating new media content.

Systems can generate different content for different users through various mechanisms. System prompts can integrate user preferences, role requirements, or long-term settings into the model’s context. Although model parameters remain unchanged, different contexts will alter the response conditions.

Retrieval-Augmented Generation (RAG) can also retrieve information relevant to the current problem from user’s past conversations, preference records, or other materials, and then add these materials to the model’s context. At this point, the model faces not just “what is the problem,” but also “what relevant information has this user provided in the past.”

Continual Fine-Tuning can also update model parameters based on new data; however, continuous training for individual users still faces challenges such as cost, data isolation, and Catastrophic Forgetting, so it cannot be simply assumed to become a standard solution for all consumer-grade products.

Researchers are exploring how to use foundation models to enhance user and item representations, generate recommendation sequences, and further build systems capable of active interaction and planning [11]. Currently, using large models to understand content and user needs has relatively clear implementation methods; directly generating complete content for individual users is still rapidly developing and has not yet become a prevalent form of recommendation products.

III. More Control Also Means More Detailed User Profiles #

Conversational recommendations allow users to directly state temporary goals, exclusion criteria, and reasons for dissatisfaction, without having to rely solely on repeatedly clicking “not interested,” unfollowing, or waiting for the system to slowly adjust. This interaction method is particularly valuable for users with limited historical data, niche needs, and complex decisions.

At the same time, conversations can generate information that was difficult to obtain from clicks in the past. A single click might only tell the platform “what the user chose,” while a conversation might reveal “why they don’t like it,” “what are the limitations,” or “what they are worried about.” Users thus gain a more direct way to correct, and platforms can acquire more detailed personalized signals than mere clicks and dwell time.

These signals can improve recommendations, and also potentially enable systems to more precisely influence user choices. However, a system’s ability to explain recommendations does not automatically equate to higher transparency in the recommendation process. Recommendation explanations can help users understand why a certain product or movie might be suitable for them, but they can also be used to persuade users to make specific choices.

Balog et al., in a study involving 129 participants, found that under these experimental conditions, natural language recommendation explanations with different biases significantly influenced user choices, sometimes even leading participants to select lower-quality recommended items [12]. This study cannot be directly generalized to all generative AI products, but it reveals a mechanism-level problem: when the same system is responsible for both selecting objects and explaining “why it’s suitable for you,” the line between explanation and persuasion is not always easily discernible.

Furthermore, if products maintain long-term memory, more detailed feedback will lead to even more sensitive user profiles. A person searching for content related to a certain illness, job, or emotion on a given day does not mean this information should permanently become part of their long-term profile. For products that save long-term memory, a more practical challenge is: can users view, modify, and delete this information, and how should the system distinguish between a temporary need and a long-term preference?

IV. How Personalized Generation Affects Relationships and Information Verification #

Traditional recommendation systems can already strengthen relationships between users and creators or virtual characters. If a person continuously watches a certain creator’s content, and the system observes longer dwell times, higher completion rates, and more frequent revisits, it might continuously increase the exposure of related content.

Communication studies have long used the term Parasocial Relationship to describe the one-sided emotional connections people form with actors, hosts, celebrities, or fictional characters [13]. Users gradually become familiar with the other party, but the other party usually does not know of this specific user’s existence; recommendation systems can increase exposure to relevant content, thereby strengthening this sense of familiarity.

Generative AI brings continuous responses that traditional content recommendations cannot provide. If a digital character can recall past conversations, remember people and things the user mentioned, and continue to communicate based on that, then from a user experience perspective, this relationship is no longer entirely “I know it, it doesn’t know me.” The underlying mechanism may still only be historical data retrieval, context invocation, and conditional generation, not directly equivalent to human memory and understanding, but the user’s perceived response indeed exists.

Current research has not yet directly explained what long-term psychological effects generative recommendations will produce. A preregistered experiment by Stein et al. found no significant differences in the overall intensity of parasocial interactions elicited by real and virtual influencers; participants’ perceived similarity to the character and the psychological “human-likeness” of the character both influenced this interaction [14]. Another four-week randomized controlled trial involving 981 participants found no significant impact of different experimental groups on major psychosocial indicators; however, cross-group observations showed that participants who used chatbots autonomously for longer periods also reported higher levels of loneliness and emotional dependence [15].

These two studies focus on different questions. The former indicates that people can also have parasocial interactions with virtual characters, while the latter observes the correlation between usage duration and outcomes such as loneliness and emotional dependence. Existing evidence is not yet sufficient to prove that all personalized interactions will lead to the same consequences.

Meanwhile, the impact of personalized generation extends from human-system interactions to human-to-human interactions. When different users begin receiving different versions of content, the way people verify information will also change.

While traditional recommendations can lead to user A seeing article A and user B seeing article B, articles A and B are usually still pre-existing public objects; others can, in principle, access the same article to verify what A actually read. Personalized generation complicates the issue by another layer: user profiles and historical information alter retrieval materials and generation conditions, which in turn affect phrasing, evidence selection, explanation order, and even causal frameworks. Different users may receive not only differentiated information sources but also different versions of answers.

This does not mean that generative AI will immediately make everyone live in completely different information environments. News reports, legal texts, academic papers, products, movies, and a large number of public cultural works still retain relatively fixed public versions, and factors such as copyright, production costs, brand value, and social communication will continue to support the existence of this content. A more realistic change is that public content itself remains relatively stable, but its summaries, explanations, combinations, and presentation methods are becoming increasingly personalized: the same report can yield different summaries, the same group of products can be organized into different recommendation schemes, and the same public question can receive different answers based on user background.

This change can enhance convenience but also increases the cost of verifying information. In the past, when encountering disagreements in opinion, one could ask:

Which report are you looking at?

If more and more people’s information entry points shift to personalized generative systems, people might also need to ask:

Under what conditions did your system generate that answer at the time?

This is also an engineering problem that system auditing urgently needs to address. Traditional recommendation systems primarily need to record what content was recommended to which users; if real-time generative systems need to support post-hoc auditing, they will also need to record the context used for generation, the materials retrieved, the personalized conditions applied, and the final generated version. If platforms store all this information long-term, it will lead to new data and privacy issues; if they do not store it at all, it will increase the difficulty of post-hoc auditing. When platforms both decide what is displayed and participate in generating the content users ultimately see, recommendations, content moderation, and data governance can no longer be discussed in complete isolation.

V. Large Models Are More Likely to Be Integrated into Recommendation Systems Than Replace Them #

These changes do not mean that traditional recommendation infrastructure will disappear. Real-world systems still need to handle vast candidate sets and consider inventory, pricing, copyright, safety, advertising, latency, and other business constraints. Large language models excel at understanding natural language and organizing results, but this does not inherently mean they are suitable for all recall and real-time ranking tasks.

Therefore, a more realistic architecture is: large language models combine current user input and authorized cross-application context to understand needs, traditional recommendation models continue to be responsible for candidate object retrieval and ranking, and generative models then handle explanation, combination, or reorganization of results.

Recommendation results may also expand from single objects to comprehensive solutions composed of multiple objects, such as travel plans, learning paths, shopping combinations, or content summaries generated from existing materials. The system then not only decides which objects to select but also participates in arranging the order, explaining the rationale, and organizing the final expression.

This also means that evaluating such systems can hardly continue to focus solely on click-through rates, watch time, and conversion rates. Models that are better at persuading users might increase clicks but may not help users make better decisions; high personalization might enhance short-term satisfaction but could also affect content diversity and creators’ opportunities for exposure.

Recommendation systems inherently involve multi-party relationships including users, content providers, advertisers, and platforms [17]. When platforms further gain generative capabilities, the evaluation of systems also needs to consider choice quality, information sources, content diversity, users’ ability to correct their profiles, and the traceability and audibility of generated results.

Conclusion #

Generative AI has not eliminated the existing recall, ranking, and feedback mechanisms of recommendation systems. Instead, it has added natural language interaction and content generation capabilities on top of these mechanisms. Users can thus state their needs more directly, and systems can organize more complete results; at the same time, the basis for recommendations, the explanation methods, and the final content may be jointly determined by the same set of systems.

Judging whether such systems are superior cannot solely focus on whether they more accurately predict clicks. It also requires examining what information they use to understand users, whether users can correct this information, whether generated results are traceable, and whether personalization genuinely improves user choice, rather than merely enhancing persuasiveness.

In the past, people were more concerned about recommendation systems influencing “what we see.” The new change brought by generative AI is that it begins to participate in shaping “what the content we see is like.”

This might be the most noteworthy shift from “guessing what you like” to “creating for you.”

References #

[1] COVINGTON P, ADAMS J, SARGIN E. Deep neural networks for YouTube recommendations[C]//Proceedings of the 10th ACM Conference on Recommender Systems. 2016: 191-198. DOI: 10.1145/2959100.2959190.

[2] PETROV A V, MACDONALD C. Transformers for sequential recommendation[C]//Advances in Information Retrieval. Cham: Springer, 2024.

[3] LIU Q, HU J, XIAO Y, et al. Multimodal recommender systems: A survey[Preprint]. 2023. DOI: 10.48550/arXiv.2302.03883.

[4] HU Y, KOREN Y, VOLINSKY C. Collaborative filtering for implicit feedback datasets[C]//2008 Eighth IEEE International Conference on Data Mining. 2008: 263-272. DOI: 10.1109/ICDM.2008.22.

[5] LIANG D, CHARLIN L, MCINERNEY J, BLEI D M. Modeling user exposure in recommendation[C]//Proceedings of the 25th International Conference on World Wide Web. 2016: 951-961. DOI: 10.1145/2872427.2883090.

[6] SANNER S, BALOG K, RADLINSKI F, WEDIN B, DIXON L. Large language models are competitive near cold-start recommenders for language- and item-based preferences[C]//Proceedings of the 17th ACM Conference on Recommender Systems. 2023. DOI: 10.1145/3604915.3608845.

[7] GOOGLE. Connect your Google apps to personalize your Gemini experience[EB/OL]. [2026-08-19]. https://support.google.com/gemini/answer/16598406.

[8] HARDT D. The OAuth 2.0 Authorization Framework[S]. RFC 6749. Internet Engineering Task Force, 2012. DOI: 10.17487/RFC6749.

[9] WU L, ZHENG Z, QIU Z, et al. A survey on large language models for recommendation[Preprint]. 2023. DOI: 10.48550/arXiv.2305.19860.

[10] LI Y, ZHAI X, HAMMAD M, et al. CALRec: Contrastive alignment of generative LLMs for sequential recommendation[C]//Proceedings of the 18th ACM Conference on Recommender Systems. 2024: 422-432. DOI: 10.1145/3640457.3688121.

[11] HUANG C, YU T, XIE K, et al. Foundation models for recommender systems: A survey and new perspectives[Preprint]. 2024. DOI: 10.48550/arXiv.2402.11143.

[12] BALOG K, RADLINSKI F, PETROV A. Measuring the impact of explanation bias: A study of natural language justifications for recommender systems[C]//Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems. 2023. DOI: 10.1145/3544549.3585748.

[13] HORTON D, WOHL R R. Mass communication and para-social interaction[J]. Psychiatry, 1956, 19(3): 215-229. DOI: 10.1080/00332747.1956.11023049.

[14] STEIN J P, BREVES P L, ANDERS N. Parasocial interactions with real and virtual influencers: The role of perceived similarity and human-likeness[J]. New Media & Society, 2024, 26(6): 3433-3453. DOI: 10.1177/14614448221102900.

[15] FANG C M, LIU A R, DANRY V, et al. How AI and human behaviors shape psychosocial effects of chatbot use: A longitudinal randomized controlled study[Preprint]. 2025. DOI: 10.48550/arXiv.2503.17473.

[16] PARISER E. The Filter Bubble: What the Internet Is Hiding from You[M]. New York: Penguin Press, 2011.

[17] BOUTILIER C, MLADENOV M, TENNENHOLTZ G. Modeling recommender ecosystems: Research challenges at the intersection of mechanism design, reinforcement learning and generative models[Preprint]. 2023. DOI: 10.48550/arXiv.2309.06375.