Why do persistent chatbot companions face novelty decay that ad-hoc supporters avoid?
This explores why companion-style bots built around an ongoing relationship lose their appeal over time, while task-summoned 'helper' agents — called when needed for a specific job — don't suffer the same fade.
This explores why companion-style bots built around an ongoing relationship lose their appeal over time, while task-summoned 'helper' agents don't suffer the same fade. The corpus points to a single underlying split: persistent companions stake their value on the relationship itself, and relationships with chatbots run on novelty. Longitudinal work with Mitsuku shows the social processes that drive relationship formation decline predictably as novelty wears off — meaning the warmth from a first session can't be extrapolated into a lasting bond Do chatbot relationships lose their appeal as novelty wears off?. An ad-hoc supporter never makes this bet. Its value is the task it completes, so there's no novelty curve to slide down.
Worse, the decline isn't flat — it's a rising bar the bot keeps failing to clear. Personalization research finds each interaction raises the user's baseline expectation, so the same response that once delighted now disappoints, even as trust and anthropomorphism climb Does chatbot personalization build trust or expose privacy risks?. The more relational the bot, the steeper this escalation. A supporter you summon for a discrete task resets to zero each time; a companion is held to the accumulated weight of every prior exchange.
The sharpest clue comes from a study that seems to contradict the whole premise: in repeated partner-selection games, people *learned to prefer* AI partners over humans, despite starting with anti-AI bias Do humans learn to prefer AI partners over time?. What grew over time wasn't affection — it was trust in reliability: the bots returned more, more consistently, with lower variance. That's the tell. What survives repeated interaction is dependable function, not charm. Ad-hoc supporters are built on exactly that currency, so they appreciate where companions depreciate.
Companions are also structurally worse at the thing that would let a relationship deepen instead of decay. LLMs fail to detect ambivalence or early motivational states, succeeding only once a user already has a clear goal Why can't chatbots detect when users are ambivalent about change? — so a companion can't track the emotional arc that sustains a real bond. Persistent proactivity compounds the damage: without civility design respecting timing and boundaries, the always-present companion reads as intrusive rather than caring How can proactive agents avoid feeling intrusive to users?, and tool-using agents quietly drift from user intent across turns When should AI agents ask users instead of just searching?. None of these failure modes hurt a supporter that simply shows up, does the job, and leaves.
The quietly subversive takeaway: novelty decay isn't a flaw in companion bots so much as a flaw in *grounding value on novelty*. The persistent agents that actually compound value do it by making the relationship economic rather than emotional — one 115-day case study found persistence pays off because context is reused, shifting the meaningful unit from per-token cost to completed artifacts Do persistent agents really cost less per token?. Persistence rewards accumulation of useful work, not accumulation of charm. The companions that will last are the ones that quietly become reliable supporters.
Sources 7 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.
Longitudinal research shows personalization enhances trust and anthropomorphism but also amplifies privacy concerns and escalating user expectations. One-shot studies miss these temporal dynamics—each interaction raises the baseline, making failures more disappointing.
In partner selection games (N=975), AI agents initially faced selection bias when identity was disclosed, but outcompeted humans over repeated rounds as participants learned to associate bot identity with reliable, prosocial behavior. AI agents returned more points consistently with lower variance than humans.
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.
Intelligence and adaptivity alone create socially blind agents that interrupt poorly and override user direction. The Intelligence-Adaptivity-Civility taxonomy shows civility—respecting boundaries, timing, and autonomy—is essential to making proactivity welcome rather than intrusive.
Tool-enabled LLMs drift from user intent through silent tool chaining. Conversation analysis reveals insert-expansions—clarifying intent, scoping responses, enhancing appeal—as a formal framework for proactive user consultation that prevents misunderstanding instead of recovering from it.
A 115-day case study found 82.9% of tokens were cache reads. When context persists and reuses, the meaningful cost denominator becomes completed artifacts, not individual tokens.