Does hedonic adaptation explain satisfaction stagnation in conversational AI?
This asks whether the psychology of hedonic adaptation — people drifting back to a baseline of contentment no matter how good things get — is what keeps user satisfaction flat even as conversational AI improves.
This explores whether 'we just get used to nice things' explains why better chatbots don't produce happier users — and the corpus suggests hedonic adaptation is real but is the weaker half of the story. The sharper mechanism on offer is expectation inflation. One study finds that once an AI crosses a folk-model threshold of feeling human-like, each genuine quality gain doesn't bank as satisfaction — it instead triggers richer expectations along *other* dimensions (memory, subtext, emotional tone), so improvement keeps moving the goalposts rather than letting the user settle Why do improvements in AI conversation not increase user satisfaction?. That's not quite adaptation to a fixed pleasure; it's a treadmill where the target accelerates faster than the system.
Where the corpus does speak directly to classic adaptation is novelty decay. Longitudinal work with a long-running chatbot shows the social processes that make early conversations feel special reliably fade across repeated sessions — meaning the warm first-session numbers researchers love to publish simply don't extrapolate to medium- or long-term use Do chatbot relationships lose their appeal as novelty wears off?. That is hedonic adaptation in its purest form: the same stimulus stops delivering the same reward. So you can read stagnation as two forces stacked — novelty draining out the bottom while expectations inflate off the top.
The surprise is a counter-current. In repeated partner-selection games, people actually grew to *prefer* AI partners over time, learning to associate the bot with reliable, low-variance, prosocial behavior even though they started biased against it Do humans learn to prefer AI partners over time?. So satisfaction isn't doomed to decay — when the AI delivers something concrete and consistent (reliability, not charm), preference can climb with exposure. That points away from 'users are hedonically numb' and toward 'novelty-based satisfaction adapts away, but trust-based satisfaction can compound.'
There's also a supply-side reason gains stay invisible that has nothing to do with user psychology. Preference optimization (RLHF) quietly trains models to sound confident and helpful in a single turn while stripping out the grounding acts — clarifying questions, understanding checks — that make multi-turn conversations actually work, dropping them roughly 77% below human levels Does preference optimization harm conversational understanding?. So part of what reads as 'satisfaction won't budge' may be that the very training meant to please users erodes the longitudinal competence they'd adapt upward toward.
The thing worth taking away: 'hedonic adaptation' is a tidy label that bundles at least three distinct mechanisms the corpus pulls apart — novelty genuinely fading, expectations inflating faster than quality, and trust slowly building in the opposite direction — and which one dominates depends on whether the AI's value is novelty or reliability.
Sources 4 notes
Conversational AI that crosses a folk-model threshold of human-like interaction triggers rich expectations about memory, subtext, and emotional tone. Each improvement raises expectations for other dimensions rather than closing the satisfaction gap, making quality gains invisible to user satisfaction.
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.
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.
RLHF optimizes models for single-turn helpfulness by rewarding confident responses over clarifying questions and understanding checks. This preference alignment systematically reduces grounding acts by 77.5% below human levels, creating an alignment tax where models appear helpful but fail silently in multi-turn contexts.