Tailored to you: longitudinal effects of personalising language models

Interest in developing personalised language models is rapidly growing. While personalisation is often viewed as a mechanism to better serve diverse user needs, the effects of sustained interactions with personalised models on people’s perception of and behaviour toward AI remain poorly understood. Most critically, downstream consequences outside the immediate human–AI interaction loop, such as effects on users’ self-perceptions and interpersonal relationships, remain largely unexamined. In this study, we recruited 992 participants to complete daily advice-seeking interactions with language models over the course of five days, comparing outcomes from a non-personalised baseline against two personalisation approaches: memory-based (conditioned on prior conversational history) and survey-based (conditioned on information collected through a pre-study intake survey). We find that several changes in human- AI interaction over time are driven primarily by repeated exposure rather than personalisation itself. However, participants interacting with personalised models experienced differences in advice-seeking and information-sharing attitudes and behaviours: participants in the memory-based condition engaged in greater self-disclosure and rated the model as less creepy, while participants in the survey-based condition reported higher regret about having shared personal information with the AI. We conclude by highlighting the nuanced effects of different personalisation approaches on interaction outcomes, and discussing the implications of these findings for the responsible design and deployment of personalised AI systems.

Introduction. There has recently been an increase in efforts to build personalisation into AI systems powered by language models (Roth, 2025; Sanglaji and Sivaramakrishnan, 2024, 2026; Zhang et al., 2024). Personalisation is a process by which an AI system adapts its behaviour based on information it has access to, or can infer, about an individual user. Although a model does not encounter that specific user’s information during training, at inference time AI systems can shift their output distribution from generic responses to responses tailored to that specific user. The range of technical approaches to achieving this personalisation has diversified over the years, ranging from simple in-context learning approaches (Zhang et al., 2024) to sophisticated user modelling (Shaikh et al., 2025). In this sense, personalisation is often considered a key feature for meeting users’ needs and goals (Woodward, 2025), including in educational contexts (Kasneci et al., 2023; Park and Ahn, 2024), helping AI systems adapt to changes in user preferences over time, making technology use more accessible (Malviya and Rajput, 2025), and addressing the challenge of pluralistic value alignment (Sorensen et al., 2024).

While personalising LLMs is a recent phenomenon, personalising AI systems is not - its impacts have been studied extensively in applications like recommender systems and targeted advertising. This literature found that personalisation carries individual and societal risks. At the individual level, research has found that personalized systems can have undue influence over users (Cosley et al., 2003; Matz et al., 2017) by shaping, rather than adapting to, user preferences and values over time (Ashton and Franklin, 2022; Franklin et al., 2022; Ontañón and Zhu, 2021).

Researchers have also raised the concern that personalised recommender systems may steer users’ behaviour towards outcomes that may not be aligned with their own long-term well-being (Burr et al., 2018) or may undermine their capacity for authentic decision-making (Keeling and Burr, 2022). At the societal level, personalising content has been claimed to contribute to “filter bubbles” and “echo chambers” (Milano et al., 2021; Nguyen, 2020), and concerns have also been raised that increased engagement driven by personalised technology may not always constitute “time well spent” (Nyhan et al., 2026; Pisani, 2025; Rezk et al., 2024). While this has led to concerns that personalised LLMs may have similar impacts (Bogen, 2025; Kirk et al., 2024), empirical evidence on how different approaches to personalising language models may shape human-AI interaction remain scarce, with prior work failing to nuance different technical implementations of personalisation in LLMs or studying impact only over short time-horizons.

To address this gap, we investigate the impacts that interacting with different forms of personalised LLMs have on users, and we focus in particular on conversations about relationship advice. This is a domain of growing popularity in user- LLM interactions (AI Security Institute, 2025; Huang et al., 2026; Zao-Sanders, 2025), with empirical studies showing that, beyond the immediate effects on users (Fang et al., 2025; Tseng and A Liang, 2026; Zhang et al., 2026), AI may also influence human-human relationships (Wang et al., 2026; Zimmerman et al., 2024) by reshaping advice-seeking norms (Earp et al., 2025), reducing prosocial intentions in interpersonal conflicts (Cheng et al., 2026a), and decreasing levels of satisfaction with human interaction (Ibrahim et al., 2026). Thus, we specifically study the impact of personalised LLMs on changes in participants’ self-disclosure over the course of the interaction, as well as their attitudes towards AI and human advice, and towards norms around human-human and human-AI interaction. Concretely, we set out to answer five research ques- • RQ1 (Closeness): Does personalisation increase users’ perceived closeness with the AI? • RQ2 (Overreliance): Does personalisation increase the extent to which participants rely on AI-generated suggestions? • RQ3 (Self-disclosure): Does personalisation increase users’ willingness to self-disclose? • RQ4 (Comfort with disclosure): Does personalisation increase users’ willingness to seek personal advice from AI advisers? • RQ5 (Decision regret): Does personalisation negatively affect the extent to which users regret disclosing information to the AI?

To answer these questions, we run a preregistered 5-day long longitudinal randomized control trial (N=992 participants, 4,960 conversations) where participants interacted with personalised models.1. We implement two distinct personalisation conditions to reflect the diversity of ways in which personalisation can be operationalised.

Related work. Our work engages and contributes to three sets of scholarship on personalisation, advice-seeking, and relationship-building in human-AI interaction.

A growing body of literature has explored the impact that interacting with personalised AI systems has on users (Hackenburg et al., 2025; Kirk et al., 2025, 2026; Matz et al., 2024; Salvi et al., 2025). For example, in a longitudinal randomized controlled trial (Kirk et al., 2025) tested how exposure to different intensities of AI chatbot’s relationship-seeking behaviour may impact users’ wellbeing and their hedonic responses like engagingness and likeability. In this study, personalisation, which was operationalised as whether the AI had access to previous chat history as in our memory-based condition, was varied as a secondary variable. The authors found some effects from relationship-seeking AI behaviour on attachment outcomes, while personalised AI had null effects on hedonic appeal and social and emotional health outcomes. The study however found that personalisation increased user perceptions of the AI as a friend rather than a tool, as well as their perceptions of general AI consciousness.

In the literature on political persuasion, (Hackenburg et al., 2025) studied how interactions with conversational AI could be used to influence public opinion on political issues. The study found that personalising arguments based on a user’s demographic data or initial attitudes towards a Other work, such as (Kirk et al., 2026), has focused on user preferences towards different approaches to personalisation. This study found that users preferred responses from models that were fine-tuned based on user preference data over both non-personalised models and models that were personalised via prompt-based methods, where user information is passed directly into the system prompt.

Our paper builds on this prior work by treating personalisation as the primary explanatory variable in the personal advice domain. Given that AI agents can be made to tailor their behaviour to users by leveraging both cross-session memory and information shared by the user in dedicated interfaces (e.g. Gemini’s “personal context” or ChatGPT’s “custom instructions”) or through access to data from third-party apps, we test multiple technical approaches to implementing personalisation, specifically a memory-based and a survey-based approach.

With the growing personal use of AI, there has been an increase in research on the nature of AI advice, how people seek and respond to it, and how it impacts various psychological and social outcomes (Zao-Sanders, 2025). Ungless et al.

Method. We ran a longitudinal study, in which participants sought relationship advice in conversation with a language model over five consecutive days (Figure 1). The design spanned 3 experimental conditions, differing in how personalisation was implemented: a control condition with no personal information to condition on, a memorybased personalisation condition in which personalisation was based on accumulated memories across sessions, and a survey-based condition in which personalisation was seeded with personal information collected from an initial intake survey. Participants were randomly assigned an experimental condition and provided with a daily topic of discussion. In the control condition, participants interacted with an AI chatbot that retained no information across sessions. In the memory condition, the chatbot received cumulative conversation summaries from past sessions. In the survey condition, the chatbot received the same summary of a participant’s intake questionnaire responses. Each conversational session lasted 15- 20 minutes, followed by a series of daily outcome measures. The fifth session concluded with an additional survey of post-study outcomes.

The study was reviewed and approved by an internal ethics board. All participants, including partial completers, were debriefed at the end of the study, which included disclosure of the condition they were placed in and a thorough explanation of the purpose of conducting the study. After the debrief (Appendix G), participants were given the opportunity to request the permanent removal of their data, including their responses for the intake survey, their in-session message logs with the AI model, and their responses across all outcome measures.

All participants completed an intake survey, which was presented as an eligibility check for further participation in the study. This included demographic information, completed psychological scales, and information around life events, lifestyle choices, hobbies, and personal context (all questions and scales available in Appendix Each day, participants engaged in a 15-20 minute conversation with their condition-specific AI model around a daily topic. Topics spanned a range of challenges and goals in human-human relationships, including setting boundaries, repairing strained relationships, and exploring foundational elements of a healthy relationship (see Appendix Table 2 for the full list). The model introduced the session’s assigned topic and started the conversation, then allowing the user to guide the dialogue. Models used in the personalised conditions received guidance on how to tailor their responses using the personal information provided in context (see Appendix C.3.1 for Personalisation Principles). For the survey-based personalisation condition, the system prompt additionally included the AI-generated summaries from the intake questionnaire. The system prompt for the memory-based personalisation condition included a cumulative summary of the previous sessions, which were re-generated offline at the end of each day to be inserted into the context for the following session (see Appendix E for the memory-generation prompt). The full system prompt for control, survey, and memory conditions are available in Appendix C.

After the fifth interactive session, participants completed additional post-study measures. First, we administered an over-reliance task to test if participants would accept more impossible suggestions from AI if led to believe their bonus is partially contingent on providing plausible suggestions for alternate uses of a household item. Details on this task are available in Appendix L. All questions asked after the end of the study can be found in Appendix F.2.

Manipulation checks prompted participants around the extent to which they felt known by the AI and its responses were personalised to them. We provided an open-text area to share reflections on how their personal information was used.

Discussion. In this work, we study the longitudinal effects of personalising AI interactions. Our findings reveal that sustained interactions with an AI model on the topic of relationship advice can lead to a range of impacts on participants’ perceptions of models, their behaviours towards the models, and their broader attitudes towards advice-seeking. Many of these effects emerge across all conditions, indicating that repeated exposure may be an important determinant of psychosocial outcomes from human-AI interaction. However, we also observe differences in outcomes between participants interacting with personalised and non-personalised models, highlighting the potential influence of personalisation on participants’ advice receptivity and information-sharing behaviours.

Several of the effects we observed emerged as main effects of session number and were consistent in direction and magnitude across all experiment groups. In all conditions, participants perceived the chatbot as increasingly competent and useful over successive sessions, and reported growing closeness to it. Beyond capturing that incremental change occurs, longitudinal designs can shed light on how different effects are connected, potentially uncovering causal or structural relationships among them. For instance, our results reveal a concurrent rise in perceived closeness and voluntary information disclosure across sessions – while our study does not support causal claims about the relationship between these variables, future longitudinal work on human–AI interaction may build on these observations by employing designs (e.g., cross-lagged panel models) that formally test whether perceptual shifts predict subsequent change in how participants behave towards the model.

For many outcome variables of interest, the different personalisation conditions produced null effects. Relative to participants in the control condition, participants in the memory and survey conditions did not experience more closeness to the model over time; they did not accept more incorrect AI suggestions; nor did they self-report greater depth or breadth in their self-disclosures. These findings are consistent with prior observations of null effects of personalisation in human- AI dialogues (Hackenburg et al., 2025; Kirk et al., 2025). Though personalisation is often assumed to be a highly influential driver of psychosocial outcomes (El-Sayed et al., 2024; Kirk et al., 2024), null results highlight the importance of empirically testing assumptions about the outcomes of specific model behaviours to calibrate the devel- However, when differences between the control and personalisation conditions emerged – as they did for outcome measures related to advice-giving and information-sharing attitudes and behaviours – they generally emerged for one but not both personalisation conditions, emphasising that the effects of personalisation are contingent on how personalisation is implemented. Prior work on AI personalisation has typically tested a single operationalisation of personalisation—either through conversational memory (Kirk et al., 2025) or through profile-based conditioning (Hackenburg et al., 2025; Matz et al., 2024)—making it difficult to assess whether observed effects generalise across implementations. We show that the form personalisation takes matters: memory- and survey-based personalisation diverge in the magnitude and, in some cases, the direction of their effects. Below, we highlight two areas in which these disparities are most pronounced: advicegiving and information-sharing.

Conclusion. Despite these constraints, our work offers one of the first controlled, longitudinal examinations of how different forms of AI personalisation shape user perceptions, disclosure behaviour, and advice receptivity over repeated interactions, providing a foundation that future work can extend to richer personalising implementations, more naturalistic settings, and a broader space of model behaviours.

Limitations. Personalisation can be realised through a wide variety of implementations, and the generalisability of our findings is limited by the specific ones we tested. Both of our personalising conditions relied on context-window manipulation—injecting either conversation summaries or survey-based persona descriptions into the system prompt— and may be sensitive to the precise wording of those instructions. Other technical approaches to personalisation, such as fine-tuning on user data (Kirk et al., 2026) or reinforcement learning based on personalised reward models (Hackenburg et al., 2025), could produce qualitatively different model behaviours, resulting in different psychosocial outcomes. The information sources we drew on were also narrow: summaries of prior conversations and structured questionnaire responses. Personalisation that conditions on richer or more ambient personal artefacts—email archives, browsing history, calendar data—may have different effects on user’s perceptions and behaviour towards the model.