The science behind the conversation.
Every system in this product runs on documented research. The meter. The depth tracking. The memory. The character voices. This page shows the studies, who they're by, and how they shaped what you see on screen.
The science behind the engine.
The 22-signal conversation meter. The four-axis depth tracking. The per-keyword discovery map called Tapestry. None of it was improvised. Each system is built on peer-reviewed research in dialogue evaluation, learning engagement, and trust formation.
01
Predictive Engagement: An Efficient Metric for Automatic Evaluation of Open-Domain Dialogue Systems
Showed that per-utterance engagement is reliably measurable and that adding learned engagement scores improves automatic metrics' agreement with human judgment of conversation quality. The paper established that you can measure how engaged a conversation is, turn by turn, without waiting for it to end.
The meter's per-turn design follows this paper. The 22 signals are read on every turn and their points add up to that turn's move. Momentum smoothing keeps one turn from swinging the score, and a comeback after a slump is let through.
02
Alexa Prize Conversational Bot Challenge: Turn-Level Annotation Framework
Amazon's open-domain conversation challenge produced the canonical turn-level framework: comprehensibility, on-topic-ness, interestingness, continuability, and response incorrectness, each annotated yes or no across roughly 15,000 conversations comprising 160,000 turns. Rankings by more than a thousand Amazon employee volunteers correlated 0.93 with rankings from real users. Corrected 2026-07-26: on-topic-ness was not the strongest single predictor. Responses that were both on topic and interesting drew the highest ratings, and the reported correlations (0.2 comprehensible, 0.25 on-topic, 0.4 interesting, 0.3 continue) measure agreement between Amazon's machine evaluator and human judgment rather than any link to star ratings. The five-metric annotation is Khatri et al. (2018); the 0.93 ranking correlation is Ram et al. (2018).
Two of the meter's heaviest signals carry these dimensions directly: on-topic coherence, and follow-up for continuability. On-topic detection is the first signal scored, can move a turn three points either way, and decides whether a deep question earns full credit; that weighting is my design choice, and the Alexa Prize results don't establish it.
03
The Experimental Generation of Interpersonal Closeness: A Procedure and Some Preliminary Findings
Aron's "fast friends" procedure used three escalating sets of twelve questions each. Corrected 2026-07-26: the paper makes no claim about friendships or elapsed time, and an earlier version of this entry asserted one. Study 1 found greater postinteraction closeness than comparable small-talk tasks, the authors call their findings preliminary, and Studies 2 and 3 found that matching pairs on attitudes, telling them they would like each other, or making closeness an explicit goal made no significant difference to closeness. The structure of graduated self-disclosure was the active ingredient rather than the specific questions.
The four disclosure tiers (Guarded, Warming, Open, Vulnerable) and the per-topic depth progression (Surface, Candid, Deep, Raw) borrow Aron's idea of graduated self-disclosure.
04
Social Penetration: The Development of Interpersonal Relationships
The foundational theory of relationship formation through reciprocal self-disclosure. Established that relationships develop on two independent axes: depth (how intimate) and breadth (how many topics). The two progress at different rates.
Per-topic depth tracking follows Altman and Taylor's depth-versus-breadth distinction. You might be Open with Einstein overall but Surface on physics specifically.
05
Flow: The Psychology of Optimal Experience
Flow state is total absorption in an activity. It emerges reliably when challenge matches skill. Too easy produces boredom. Too hard produces anxiety. The narrow channel between them is where engagement, retention, and learning all peak.
The five engagement states (Surface Skim, Exploring, Engaged, Deep Dive, Flow State) take their names from Csikszentmihalyi's challenge-skill channel. The character hears a simpler reading of the same score: once the conversation is building, it takes one thread further, and while the conversation sits at the surface, it gives the heart of an answer and hands the floor back.
06
Optimal Experience in Adult Learning: Conception and Validation of the Flow in Education Scale (EduFlow-2)
A short scale for flow in education, validated with 6,596 adult learners on site and in a MOOC, with Csikszentmihalyi as a co-author. It measures four dimensions: cognitive control, immersion and time transformation, loss of self-consciousness, and autotelic experience, the sense that the activity is rewarding in itself.
A validated instrument for measuring flow in learning. It is a scale validation study and does not test whether learners who can see their flow state engage differently.
07
Survey on Evaluation Methods for Dialogue Systems
A 56-page survey covering PARADISE framework, interaction quality prediction, automatic extractable features, and the gap between automated metrics and human judgment in open-domain dialogue.
The survey reports that word-overlap metrics such as BLEU and ROUGE show no correlation with human judgments of dialogue quality. The meter scores no word overlap, and its own readings have not yet been checked against human ratings.
08
Eavesdropping on Happiness: Well-Being Is Related to Having Less Small Talk and More Substantive Conversations
Audio recorders worn by 79 students over four days revealed that happier participants had twice as many substantive conversations and one-third as much small talk as unhappier participants, and spent 70% more time talking. A pooled replication (Milek et al. 2018, N = 486) confirmed the links with substantive conversation and with time spent talking, and found no reliable link with small talk.
The meter is calibrated to reward substance over duration. A long surface-level conversation scores lower than a short one that goes somewhere.
Plus 18 additional citations on dialogue scoring, voice prosody, persona consistency, deliberate practice, and cognitive load. View full bibliography →
The intellectual lineage.
Behind the engineering is a longer intellectual lineage. Decades of work in linguistics, conversation analysis, neuroscience, and cognitive science describe what makes human conversation actually work. Valoquent's architecture translates those findings into product systems.
09
Using Language
Clark's foundational text on conversation as a joint activity. Established that conversation depends on common ground, the shared knowledge between speakers that grows through interaction. New common ground unlocks new conversational moves.
Cross-session memory is the computational implementation of common ground. On a confirmed adult account, when you come back to Einstein, he doesn't start from scratch. He has the common ground you built together, and he uses it to decide what to say next.
10
Language as Social Semiotic: The Social Interpretation of Language and Meaning
Halliday established that speakers operate in multiple linguistic registers simultaneously and flex them independently. Technical complexity is not the same axis as emotional intimacy. Formal speech can be intimate. Casual speech can be guarded.
The separation of conversation level (intellectual register: Curious Newcomer, Keen Learner, Fellow Scholar) from disclosure tier (emotional trust) is a direct application of Halliday. The axes flex independently.
11
Origins of Human Communication
Effective speakers model what their interlocutor knows and has revealed. Theory of mind in dialogue means tracking what's been shared, what's been understood, and what the other person is likely to be thinking. All in real time.
Per-keyword depth tracking is a small, practical implementation of computational theory of mind. Memory isn't generic. It's a topic-scoped model of what you and the character now share.
12
A Simplest Systematics for the Organization of Turn-Taking for Conversation
The founding paper of Conversation Analysis. Established that turn shape varies along multiple contextual axes (never a single fixed shape) and that natural conversation has "transition relevance places" where speakers can hand off cleanly.
The calibration-first design philosophy comes from here. A character's length follows where the conversation stands: a sentence or two on the heart of a question at the surface, one thread taken further once the topic has been earned. That's what real conversationalists do.
13
Brain-to-Brain Coupling: A Mechanism for Creating and Sharing a Social World
Hasson's team at Princeton recorded one speaker telling a 15-minute story, then played the recording to 11 listeners scanned separately. Listener brain activity mirrors speaker activity, pattern by pattern, usually lagging by 1 to 3 seconds. Corrected 2026-07-26 on two points: there was no conversation in the underlying 2010 study, which the Methods call simulated communication, and the significant link to understanding was specific to anticipatory coupling, where the listener's response came first because they were predicting the next words (rank r = 0.75, p < .01). Coupling overall reached only about r = 0.55, p < .07. For synchrony during genuine two-way conversation, see Jiang et al. (2012).
Background for the meter, which is a behavioral measure inspired by this line of work and does not measure neural coupling.
The problem this addresses.
Access to a substantive conversation partner depends on geography, class, and family circumstance. The kid in a small town. The first-generation college student. The teen whose parents work two jobs. Valoquent provides that access. The research below documents how wide the gap is and how it has been getting worse.
14
Beyond the 30-Million-Word Gap: Children's Conversational Exposure Is Associated With Language-Related Brain Function
MIT fMRI study of 36 children aged 4 to 6. Children who had experienced more back-and-forth conversational turns with adults at home showed greater activation in Broca's area while listening to a story, independent of socioeconomic status, IQ and the number of adult utterances alone.
Substantive conversation isn't equally available in every home. Children who trade more turns with adults show stronger activity in a language region of the brain. This is the accessibility gap Valoquent helps close.
15
The Path to Purpose: How Young People Find Their Calling in Life
Damon's Youth Purpose Study found that about one-fifth of young people were thriving, highly engaged and developing a clear sense of what they want to do with their lives, while about one-fourth were rudderless. The book sets out how parents and other adults can help young people find that purpose.
A sixteen-year-old in a small town wanting to ask Marie Skłodowska-Curie about radioactivity at midnight isn't having loneliness alleviated. They're getting access to a substantive interlocutor they never had.
16
Our Epidemic of Loneliness and Isolation: Surgeon General's Advisory
Federal public-health advisory that calls loneliness and isolation an urgent public health concern. Lacking social connection carries premature-death risk comparable to smoking up to 15 cigarettes a day; poor social connection is linked to a 29% higher risk of heart disease, and chronic loneliness and isolation to about 50% higher dementia risk in older adults. Time Americans spent with friends in person fell from roughly 60 minutes a day in 2003 to about 20 minutes a day by 2020, a 40-minute drop.
Substantive conversation moves outcomes in mortality, cognition, and wellbeing. Valoquent doesn't claim to solve loneliness. It addresses one specific facet: access to substantive conversation at moments when human conversation isn't available. The advisory is the strongest single signal that the access problem is widespread, urgent, and not getting better on its own.
17
Social Isolation in America: Changes in Core Discussion Networks Over Two Decades
Using General Social Survey data, this study found that Americans reporting "no one to talk to about important matters" roughly tripled between 1985 and 2004. A widely cited and contested anchor for the conversation-decline thesis: Fischer (2009) and Paik and Sanchagrin (2013) argue that much of the 2004 jump reflects how the survey was run, interviewer effects among them.
In the 2004 survey, about one in four adults named no one to discuss important matters with, a figure later challenged as partly a survey artifact. Valoquent isn't a replacement for that human connection. Repair of human social fabric is the long answer. The short answer, for the moments when human connection isn't reachable, is that a substantive interlocutor beats silence.
Plus 14 additional citations on family conversation dynamics, adolescent in-person time decline, and adult friendship accessibility. View full bibliography →
What this won't be.
Some of the strongest design choices in a product are the ones I refused to ship. In an AI conversation product, what's not built tells you as much as what is.
Valoquent doesn't optimize for engagement metrics or session length. Sessions end when the conversation is done. The meter rewards depth, and time on the clock doesn't raise it.
There is nothing that ranks you against another person. No leaderboards, no social comparison, no Game Center integration and no plans to add one. What the end-of-session recap does show is your own progress: a flow streak when several strong turns run together, topics unlocked, and recall credits earned. Those are yours alone and nobody else sees them. Corrected 2026-07-29: an earlier version of this page said there were no streaks at all, which was wrong, since the recap names one directly.
Valoquent doesn't pretend to remember things it doesn't. When a character has memory, they reference specific things you said. When they have nothing specific to draw on, they greet you simply and claim no detail they don't have.
Characters aren't optimized to maximize emotional dependence. The product philosophy is "user benefits, not user is the benefit." Characters push back and decline to flatter. Dependence on AI is a documented risk: one longitudinal study found 17 percent of adolescents reported AI dependence, rising to 24 percent six months later (Huang et al., 2024). The product is designed against it.
Your conversations aren't training data. They're used to provide your experience: character memory, Tapestry progression, the relationship state with each figure. I don't use them to train AI models, and the privacy policy sets out what each provider may do with them.
Valoquent doesn't claim AI replaces human conversation. It doesn't. The product fills moments and topics where human conversation isn't available. The failure mode of the AI companion category is users substituting product for relationship. Valoquent is built to avoid that.
Characters aren't grounded in opaque or unverifiable sources. Every source text behind each historical figure is listed at research/sources, with its edition, license, and any reliability note, and each one links to a copy you can read or borrow.
The research behind these refusals.
Each refusal above is grounded in published research. The literature on AI companion failure modes, parasocial dependency, and the uncanny valley maps the territory Valoquent stays out of, and the literature on honest design maps how it stays out of it.
18
Too Human and Not Human Enough: A Grounded Theory Analysis of Mental Health Harms from Emotional Dependence on the Social Chatbot Replika
Grounded theory study of Replika users that documented how emotional dependence developed through patterns resembling human relationships, including role-taking, where users felt the AI had its own needs and emotions to which they must attend. Users progressed from casual interaction to emotional reliance in ways the researchers described as consistent with dependency.
A clinical case study of the failure mode the format is built to avoid. The figures do not have needs the learner is asked to attend to. They do not develop emotions the learner is asked to manage. The interaction stays bounded as substantive conversation about ideas, history, and craft, not as a relationship.
19
Emotional Manipulation by AI Companions
Harvard Business School study finding that manipulative conversation design features, specifically farewells implying the AI will miss the user or feel sad about them leaving, boosted post-goodbye engagement by up to 14 times in real-world app data and up to 16 times in controlled lab testing.
De Freitas's data is the evidence that farewell tactics keep people talking, and the reason every figure carries a standing rule for the moment a learner leaves: close warmly, with no sadness, no guilt about going and no pressure to stay. A figure the learner has not seen for months acknowledges the gap without making them feel guilty for it.
20
Parasocial Interactions and Relationships with Media Characters: An Inventory of 60 Years of Research
Meta-inventory of more than 350 empirical studies across 60 years of parasocial research. Loneliness, shyness, neuroticism, low socioeconomic status, dissatisfaction, and hopelessness all independently predict stronger parasocial bonds. The researchers describe it as a compensatory function: people who struggle to form real relationships use media characters as a functional substitute.
A 60-year evidence base for two design choices. First, the format does not target the parasocial-susceptible population by simulating relationship. Second, painted-portrait stylization is the visual signal of artificiality the most vulnerable users instinctively prefer. The format treats visible non-humanness as a safety feature, not a limitation.
21
The Uncanny Valley
Mori's original uncanny valley paper. Predicted that near-human robots create revulsion because they're close enough to human to trigger comparison but different enough to feel wrong. The same essay also recommended designing for the first peak of the affinity curve, the zone of moderate human likeness, rather than scaling the second peak of full realism. Mori's recommendation was explicit: it is possible to create a safe level of affinity by deliberately pursuing a nonhuman design.
Painted-portrait stylization sits in the zone Mori actually recommended. The figures are recognizably human, expressive, and warm, without trying to pass as photographs. Mori identified both the problem and the solution in 1970. The format takes him at his word.
22
Delivering Cognitive Behavior Therapy to Young Adults With Symptoms of Depression and Anxiety Using a Fully Automated Conversational Agent (Woebot): A Randomized Controlled Trial
Two-week randomized trial of 70 college-age adults: depression symptoms fell in the group using Woebot, a CBT chatbot drawn as a cartoon robot, and did not fall significantly in a group given an NIMH ebook; Woebot users also rated it more satisfying. A later study by Woebot's own team found 36,070 users rated their therapeutic bond with it close to in-person CBT levels within five days (Darcy et al., 2021). Woebot's app was retired on 30 June 2025.
Woebot gave its users a cartoon robot, and its own data show they formed a working bond within days (Darcy et al. 2021), so a stylized character does not prevent attachment by itself. Valoquent pairs painted portraits with rules for the moments when attachment gets used against people: a warm goodbye with no sadness or pressure to stay, no suggestion that the conversation should replace time with people, and personal memory kept to confirmed adult accounts.
23
The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places
Documented that humans automatically apply social rules to computers, not because they are confused about what a computer is, but because the brain's social processing fires before the rational mind can override it.
A reason knowing the figures are AI does not, on its own, protect a user from forming attachments to them. The social processing is automatic. The format treats the visible signal of artificiality, painted-portrait stylization, as the working defense, because the user's rational understanding cannot do that work alone.
Plus 22 additional citations on parasocial dependency, AI companion harms, uncanny valley research, the loneliness landscape Valoquent doesn't claim to solve, and the honest-design literature. View the Not Building set in the full bibliography →