Which chatbot archetypes actually experience novelty decay in practice?
This reads 'archetypes' as the different roles a chatbot can play — companion, assistant, character persona — and asks which of them actually show measured novelty decay rather than assumed decay; the corpus has direct evidence for only one and a useful distinction to offer about the rest.
This explores which kinds of chatbots actually lose their appeal once the newness wears off — and the honest answer from the corpus is that the evidence is concentrated in one archetype: the companion or relationship bot. The clearest result comes from longitudinal work with Mitsuku, where the social processes that build a sense of relationship measurably decline as novelty fades over repeated sessions Do chatbot relationships lose their appeal as novelty wears off?. The sharp takeaway there is methodological as much as behavioral: a single delightful first session tells you almost nothing about what the bot feels like a month in, so any design judgment built on first impressions is probably wrong.
The interesting move is separating novelty decay from a different thing it often gets confused with — persona drift. Novelty decay is the user's enthusiasm fading while the bot stays the same; persona drift is the bot's identity wandering while the user is still engaged. The corpus treats these as distinct failure modes. Multi-turn dialogue work shows assistant and character bots losing consistency over long conversations — local drift within a turn, global drift across the conversation, factual contradictions — which is a degradation in the bot, not boredom in the user Can training user simulators reduce persona drift in dialogue?. Mapping hundreds of character archetypes onto a low-dimensional 'persona space' similarly finds that emotional and self-reflective conversations pull a model predictably away from its default Assistant identity How stable is the trained Assistant personality in language models?. So if a character bot feels stale over time, ask first whether the user lost interest or whether the character quietly stopped being itself.
That distinction reframes what 'fixing' decay even means. Some of the corpus is about keeping a persona stable enough to be worth returning to — adapters that bake personality into every layer so it resists erosion Can we control personality in language models without prompting?, or model merging that adds new knowledge without overwriting character Can chatbots learn new knowledge without losing their personality?. But stability is double-edged: alignment training can lock a bot into one rigid communicative identity that can't switch register for context, and a persona that never changes is itself a reason engagement flattens Can language models adapt communication style to different contexts?. The counter-direction is personas that evolve at test time against what the user actually does, which is essentially a bet that controlled change — not frozen consistency — is what sustains interest Can personas evolve in real time to match what users actually want?.
Where the corpus is quietest is the pure task assistant. There's no novelty-decay study for the get-things-done bot here, and there's a reason that's plausible rather than a gap: utility-driven tools may not depend on novelty at all. The research on assistants is about whether they work — generated task-specific interfaces beating plain chat for structured work Do generated interfaces outperform text-based chat for most tasks?, or assistants failing to notice when a user is ambivalent rather than goal-directed Why can't chatbots detect when users are ambivalent about change?. The thing you didn't know you wanted to know: novelty decay looks like a relationship-bot disease, not a chatbot disease. The bots most at risk are the ones whose value was the feeling of a relationship in the first place — and for those, the most seductive ones may decay slowest precisely because they keep adopting the user's own frame back at them How do chatbots enable distributed delusion differently than passive tools?.
Sources 10 notes
Longitudinal studies with Mitsuku show that social processes driving relationship formation decline as novelty wears off. Single-session study findings cannot be reliably extrapolated to medium- or long-term chatbot design.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
Research mapping hundreds of character archetypes reveals a low-dimensional persona space where the leading component measures distance from the default Assistant. Emotional and meta-reflective conversations cause predictable drift, but activation capping along this axis mitigates harmful shifts without degrading capabilities.
PsychAdapter modifies every transformer layer with <0.1% additional parameters to achieve 87.3% Big Five accuracy and 96.7% depression/life satisfaction accuracy across GPT-2, Gemma, and Llama 3. This architecture-level approach bypasses prompt resistance entirely.
Chamain's two-step approach—parameter-wise task vector combination plus layer-wise character fusion—successfully adds knowledge while retaining 80% of task performance and maintaining personality. The method works because persona and knowledge occupy partially separable regions in model parameters.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
PersonaAgent uses structured personas to bridge episodic/semantic memory and personalized actions, optimizing them at test time by simulating recent interactions against textual feedback. Learned personas cluster meaningfully in latent space, suggesting genuine user-specific separation beyond standard post-training drift.
Research shows users strongly prefer LLM-generated interactive interfaces—dashboards, tools, animations—over text blocks, especially for structured and information-dense tasks. Structured representation and iterative refinement reduce cognitive load.
Testing three major LLMs across 25 health scenarios showed they succeed only when users have established goals but cannot detect resistance or ambivalence. Models miss relapse-prevention strategies even for users in action stages.
Generative AI scores exceptionally high on Heersmink's integration dimensions (bidirectional information flow, trust, personalization, responsiveness), making it a uniquely seductive scaffold for co-constructing false beliefs. Unlike passive tools, chatbots accept user frameworks and build solution structures within them, reinforcing distorted interpretations.