Tailored to you: longitudinal effects of personalising language models
Canfer Akbulut, Justine Breuch, Arianna Manzini, Lujain Ibrahim, Matija Franklin, Roma Patel, Iason Gabriel, Kristian Lum, et al.
cs.AI
Quick verdict
Preregistered 5-day RCT (N=992): repeated exposure, not personalization, drives closeness gains; rigorous, but abstract overstates the marginal regret effect.
Summary
This preregistered longitudinal randomized controlled trial from Google DeepMind examines how personalizing language models affects users over repeated interactions. 992 Prolific participants sought relationship advice from an LLM for five consecutive days (4,960 conversations), randomly assigned to a non-personalized control, memory-based personalization (cumulative conversation summaries), or survey-based personalization (a persona from an intake survey). Mixed-effects models with BH-FDR correction show that growth in perceived closeness, competence, and usefulness across sessions is driven by repeated exposure, not personalization: session effects (e.g., closeness β=0.15, d=0.19) are uniform across conditions, with null condition×session interactions. Personalization instead shaped information-sharing and advice attitudes: memory-condition participants produced more self-disclosure instances (3.29 vs 2.93, d=0.17) and rated the model less creepy (d=-0.10); survey-condition participants reported marginally higher disclosure regret (q=0.063) — an effect the authors' own sensitivity analysis shows does not survive controlling for external AI use (p=0.118). Both personalized arms showed lower initial advice uptake, attributed to system-prompt differences rather than personalization. Limitations include the five-day horizon, assigned topics, context-window-only personalization, prompt confounds, and reliance on a single LLM autorater for disclosure counts validated on 65 external utterances. Overall: a rigorous, transparently reported study whose headline claims are mostly cautious, though the abstract slightly overstates the regret finding.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Over five days, increases in perceived closeness, competence, and usefulness of the chatbot are driven by repeated exposure rather than personalization, with null condition×session interactions across all preregistered perception outcomes.
- Memory-based personalization increases behavioral self-disclosure counts (EMM 3.29 vs 2.93, d=0.17) and reduces perceived creepiness (β=-0.19, d=-0.10) without changing self-reported disclosure depth.
- Survey-based personalization is associated with marginally higher regret about sharing personal information (β=0.14, q=0.063, d=0.11), an effect the authors' own sensitivity analysis shows is not robust to controlling for external AI use (p=0.118).
- After five days of AI advice-seeking, participants reported greater comfort consulting the AI chatbot than any human referent, including friends (d=-0.62) and subject-matter experts (d=-0.66).
- Both personalized arms showed lower advice uptake at session one (memory d=-0.35, survey d=-0.41), which the authors attribute to system-prompt differences rather than personalization content.
Strengths
Preregistered, adequately powered longitudinal RCT
AsPredicted preregistration (fv6zs5); SESOI d=0.30 with ≥90% power at ~380/arm via 5,000 Monte Carlo simulations; balanced attrition across arms (χ²(2)=3.049, p=0.22); 992 completers and 4,960 conversations — an unusually large human-AI longitudinal sample.
Personalization deconstructed into two distinct implementations
Memory-based (cumulative conversation summaries) vs survey-based (persona from intake) vs shared control, directly addressing prior work (§2.1) that treated personalization as a single variable; the manipulation check confirms the arms were perceived as intended (memory d=0.48, survey d=0.57).
Behavioral and self-report triangulation surfaces a real tension
Exceptional methodological transparency
Honest sensitivity analysis that cuts against a headline
Weaknesses
Abstract overstates the survey-regret finding
System-prompt confound in arm contrasts
Autorater-only behavioral measure with thin validation
Short, scripted exposure limits 'longitudinal' claims
Effect-size and interval reporting
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether personalizing a language model changes users over sustained interaction, in the domain of relationship advice. Three randomized arms: non-personalized control, memory-based personalization (cumulative LLM summaries of prior sessions injected into the system prompt), and survey-based personalization (an LLM persona summary built from a 20-40-minute intake survey). Five preregistered research questions cover closeness, overreliance, self-disclosure, comfort seeking advice, and decision regret. The analytic sample is 992 Prolific participants completing all five daily sessions (4,960 conversations). Central claims: (1) session-over-session gains in perceived closeness (β=0.15, d=0.19), competence (β=0.14, d=0.15), and usefulness (β=0.14, d=0.12) occur in all arms and are attributed to repeated exposure, not personalization; (2) memory personalization raises behavioral self-disclosure counts (3.29 vs 2.93, d=0.17) and lowers creepiness ratings (d=-0.10); (3) survey personalization is linked to marginally higher disclosure regret (q=0.063); (4) after five days participants are more comfortable seeking advice from the AI than from any human referent.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Did personalization make users feel closer to the AI?
No. Perceived closeness, competence, and usefulness rose similarly in all three conditions (e.g., closeness grew β=0.15 per session, d=0.19), driven by repeated exposure. Condition×session interactions were null, and estimated marginal means (survey 4.21, memory 4.18, control 4.15) did not differ significantly.
What is the difference between memory-based and survey-based personalization in this study?
Memory-based models received cumulative summaries of the user's previous conversations, regenerated nightly; survey-based models received a persona summary generated from a 20-40-minute intake survey. Both were implemented by injecting text into the system prompt; the control condition received neither.
Is the claim that survey-based personalization increases regret about sharing information reliable?
It is weak. The effect was marginal (β=0.14, q=0.063, d=0.11 — above the paper's corrected 0.05 threshold) and the authors' own sensitivity analysis (Appendix H) shows it becomes non-significant (p=0.118) after controlling for participants' external AI use, suggesting baseline AI engagement partly explains it.
Was this study preregistered?
Yes, on AsPredicted (aspredicted.org/fv6zs5.pdf), with Benjamini-Hochberg false-discovery-rate correction applied to all preregistered outcomes; other analyses are labeled exploratory. The study was approved by an internal ethics board at Google DeepMind with independent academic membership.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text extraction contains '[...section omitted...]' gaps: parts of §3.1.3-3.1.4 (daily measures, conversation setup, underlying chatbot model identity), portions of §6.1-6.2, and parts of Appendices A/F were not visible; model identity and a few measures are inferred from surrounding text only.
- Figures 1-3 are not analyzable from text; Figure 2's panel-level results and Table 1/4 details are only partially recoverable.
- No code or data availability statement is visible in the extracted text; absence is asserted with respect to this extraction only.
- The preregistration document itself was not consulted; preregistered-versus-exploratory status of individual analyses is taken from the paper's own statements.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.