The science of what happens between humans and AI.
Validated psychological measurement of interaction state — derived from language, not model inference.
200+ dimensions — cognitive load, rapport, distress, wellbeing trajectory — each traceable to 34,000 peer-reviewed citations. Deterministic scores in under 65ms, deployable on-prem.
34,000+
Peer-reviewed research citations
200+
Validated psychological dimensions
30+
Years of foundational research
<65ms
API response in production
AI systems continuously form judgments about the people they're talking to. Those inferences inform every response, and those responses shape how users think, feel, and behave.
But the inferences themselves are invisible. Nobody can inspect what was inferred, whether it's correct, or how it's affecting the person on the other side. Horvitz and West have argued that our understanding of AI is falling behind even as AI's understanding of us deepens - and this opaque inference layer is part of why.
Making it measurable is one of the more tractable ways to close that gap.
Receptiviti Labs is building the measurement layer for the human side of AI.
THE GAP
AI can tell you what the model did, but not how it impacts the user.
AI observability stacks are comprehensive on the model side — traces, latency, accuracy, hallucination rate, output drift. The interaction itself has no instrumentation at all.
Whether the interaction is building understanding or eroding it, landing or misfiring, fostering independence or dependence — all of that is left to inference. Today it's all invisible.
Current tooling
What Receptiviti measures
WHY IT MATTERS
Influence runs both ways. What you can't measure, you can't correct.
An AI influences the state of the person it's talking to. That person's state should determine how the AI responds. Safety, alignment, personalization, and trust all depend on seeing both directions — and today both are left to inference.
OUR POSITION
Four things we think are true.
01
Human agency is the variable AI systems are least equipped to protect.
When people delegate their reasoning, judgment, and decision-making to AI, those capabilities can atrophy. Peer-reviewed evidence already documents over-reliance, reduced critical engagement, and narrowing judgment as measurable effects of sustained AI interaction. AI systems that genuinely preserve human agency may sometimes need to introduce friction to prompt users to reflect, slow their deference, and resist the pull toward dependency. Doing so requires the system to measure how its interactions affect a user's cognition over time. Measuring the dynamics of AI-user interactions is the only way to give systems the information they need to introduce the right friction at the right time.
02
The signals that matter most are the ones no one is measuring.
Every conversation with an AI system contains a layer of linguistic information that encodes the user's cognitive load, emotional state, distress, and psychological risk. These signals contain the most consequential information in the interaction, but they are invisible to every current AI evaluation pipeline. The models, however, already respond to these signals, using inference to form judgments about the user's state and adjusting their responses accordingly. But inference is often inconsistent and inaccurate, and those judgments exist only inside the model. Nobody can see what was concluded, whether it was accurate, or when it went wrong. Making user state an explicit, structured variable is what turns those hidden judgments into something that can be understood, inspected, and corrected.
03
Understanding how AI understands users is the most underexplored entry point in interpretability.
AI systems continuously infer psychological signals from the people using them. Those inferences shape every response, and those responses have a direct impact on how the user thinks, feels and behaves. This inference happens between input and output, which is why standard evals miss it entirely. Making those inferences visible, explainable, and consistent is one of the most practical paths to understanding what these systems are actually doing, how they perceive users, and why they respond the way they do.
04
True alignment and safety will require measuring what's actually happening to each user, across time.
A system optimised for the average user is misaligned with almost every actual user. We believe alignment needs to become user-specific by measuring what is happening to each user, in the current session and across their relationship with that system. That requires a different kind of measurement than exists today - measurement that can see, for example, cognitive load rising across a session, confidence eroding conversation by conversation, or dependency deepening over weeks. These changes don't appear in individual interactions; they emerge across time. A system that can't see the impact it's having on a user doesn't have the safeguards needed to keep users safe, and probably shouldn't be operating autonomously.
WHY EXPLICIT MEASUREMENT
What explicit user state measurement gives you that inference cannot.
The problem with inference
Implicit and unauditable
When a model infers psychological state, that state never becomes an observable variable. It influences behavior without being visible, citable, or challengeable.
Prompt-sensitive and unstable
LLM-based inference shifts unpredictably across model versions and phrasings — the same conversation yields different inferences on a different day.
Circular — the model grading itself
Asking the model being evaluated to assess its own impact on the user is not independent measurement. It's the model grading itself.
What Receptiviti provides
Structured, observable variables
200+ dimensions derived from language — explicitly computed, consistent across runs, independent of model behavior. Observable, auditable, citable.
Version-stable
Derived from validated psycholinguistic frameworks built over 30 years, not from prompt-dependent inference. The same input yields the same score.
34,000 citations
Every dimension traceable to peer-reviewed science. The kind of evidence that holds up in research, in product decisions, and in accountability conversations.
WHERE IT FITS
The interaction gap shows up everywhere AI meets people.
Different problems and the same underlying need: measurement of what's happening on the human side of the interaction.
EVALUATION BLIND SPOTS
Eval stacks measure the model completely — traces, drift, faithfulness, hallucination rate. Receptiviti adds the interaction-state signal — an eval dimension that scores alongside existing model quality metrics
ALIGNMENT AND SAFETY
Foundation model teams need validated, external evidence of how interactions affect users over time — cognitive load, emotional trajectory, over-reliance. The measurement layer for socioaffective alignment. Not inferred by the model being evaluated. Measured independently, derived from language.
LONG-TERM IMPACT
Products built for ongoing relationships need to know whether those relationships are helping or harming. Therapeutic alliance, dependency signals, distress escalation, wellbeing trajectory - the signals that determine whether an AI relationship is beneficial over time.
ENTERPRISE AI
Resolution rates and CSAT tell you what happened. Interaction-state measurement tells you why — whether the AI frustrated or calmed the person, built rapport or lost it, and how that trajectory unfolds across the relationship.
RESPONSIBLE AI
Impact assessments and responsible AI programs require measurement that is validated, traceable, and independent of the model being assessed. Receptiviti's scores are grounded in peer-reviewed science and structured for documentation - evidence that holds up to external review.
Built on three decades of validated psycholinguistic science.
Dr. James W. Pennebaker — Co-founder and Chief Science Officer
Regents Centennial Professor, University of Texas at Austin. Creator of LIWC — the foundational psycholinguistic framework at the core of Receptiviti's measurement science. The three decades of research that made this possible.
34,000+
Peer-reviewed research citations
ACTIVE RESEARCH
We publish. We contribute.
Receptiviti's team and academic advisors publish peer-reviewed research on the psychological dimensions of language and human behavior. Receptiviti also conducts its own research and experiments, contributing to the questions AI safety, alignment, and evaluation teams are actively working on.
PUBLISHED: PNAS Nexus, 2024
Large language models display human-like social desirability biases in personality surveys
Salecha, Ireland et al.,2024. Large language models display human-like response biases when they infer they are being evaluated, with effects up to 1.20 human SD across GPT-4, Claude 3, Llama 3, and PaLM-2. Co-authored by Molly Ireland, Receptiviti.
PUBLISHED: Perspectives on Psychological Science, 2026
Artificial intelligence and the psychology of human connection
Boyd & Markowitz, 2026. Introduces the MIRA model - a theoretical framework for when and how AI functions as a relational entity in human ecosystems. Language is the primary modality through which that relationship operates. Co-authored by Ryan Boyd (UT Dallas), academic advisor to Receptiviti.
INTERNAL STUDY
Conditioning customer-support AI on measured user state
Whether AI agents conditioned on real-time interaction state deliver better resolution outcomes and lower escalation rates.
PUBLISHED: npj Mental Health Research, 2025
Psychosocial dynamics of suicidality and nonsuicidal self-injury: a digital linguistic perspective
Entwistle, Hoemann, Nightingale & Boyd, 2025. Large-scale naturalistic study of the language dynamics surrounding suicidality and self-injury in 992 individuals with borderline personality disorder (66,786 posts). Co-authored by Ryan Boyd (UT Dallas), academic advisor to Receptiviti.
INTERNAL STUDY
Conditioning educational AI on measured user state
GPT Study Mode: conditioning on psycholinguistic interaction-state signals produced a +4% improvement in overall educational effectiveness across 25 blinded evaluations. Largest gains in reasoning & scaffolding (+6.3%) and cognitive-load management (+5.9%).
SCOPED
Interaction state as a missing dimension in AI evaluation
The case for human-state signals as a first-class eval criterion alongside accuracy, helpfulness, and harmlessness.
Field Notes.
