ALIGNMENT & SAFETY
Validated, independent measurement of the human side of AI interaction.
Derived from language.
Independent of the model being evaluated.
Alignment research requires evidence the model being evaluated cannot independently provide about itself. Without an independent signal of how users are responding during and across interactions, important changes in reasoning, affect, reliance, and agency can go unmeasured. Receptiviti adds that measurement layer directly to existing logs and traces.
HOW IT WORKS
Pass conversation text. Receive structured psychological variables. Append to your existing logs.
No changes to your data infrastructure. Receptiviti scores sit alongside your existing traces, conversation logs, and eval outputs — adding the human-state dimension to data you are already collecting.
Conversation turn​
User or model language - submitted as plain text
API CALL
Receptiviti​
200+ psychological variables <65ms
APPEND
Your logs & traces​
Existing eval data + human-state scores
Receptiviti measures psychological signals from language. It does not diagnose, classify individuals, or replace clinical judgment. Scores are structured variables for research and evaluation purposes.
THE CORE PROBLEM
AI models cannot provide independent measurement of their own effect on users.
Most evaluation frameworks measure what the model outputs. They rarely measure how users are changing during and across interactions. That evidence requires a measurement channel outside the model being evaluated.
01
Circular evaluation
Asking the model to assess its own psychological effect on users is not independent measurement. This is especially important when evaluating phenomena such as sycophancy, over-reliance, dependency, or changes in user agency.
02
Socioaffective alignment is a non-stationary target
The human-AI relationship changes user preferences and perceptions through mutual influence. Static benchmarks and single-session evals cannot capture what emerges longitudinally — cognitive offloading, dependency, erosion of independent reasoning — because these patterns only become visible across time.
03
Wellbeing claims require evidence
Model-generated inferences are not inherently independent and can vary across prompts, models, and versions. Safety and wellbeing claims benefit from reproducible measurement grounded in documented methods.
WHAT RECEPTIVITI MEASURES
The dimensions alignment research needs — externally validated, derived from language.
200+ psychological dimensions scored from language, independent of model behavior. The same input yields the same score across model versions, prompt variations, and evaluation contexts.
COGNITIVE
Over-reliance & cognitive offloading
Analytical thinking and related measures provide signals of how users’ reasoning engagement changes across interactions. Cognitive load captures changes in the cognitive demands reflected in their language.
AFFECTIVE
Emotional state & wellbeing trajectory
Emotions, distress signals, and cognitive load scored from language — not from the model's inference about how the user feels. Longitudinal wellbeing trajectory across interactions, not just single-session snapshots.
RELATIONAL
Dependency & attachment signals Rapport, affiliation, analytical thinking, and related measures can surface language patterns associated with changing relational dynamics and reliance over time.
SOCIAL
Agency & autonomy preservation
Clout, dominance, and deference measures can contribute to evaluating whether patterns of autonomy or deference are changing across interactions.
WHY MEASUREMENT, NOT INFERENCE
Psychological state needs to be measured independently. Not inferred.
We need a science of AI safety that studies real human-AI interactions in natural contexts and treats the psychological and behavioural responses of users as key objects of inquiry.
Kirk et al. — npj Mental Health Research, 2025 · Oxford, Google DeepMind, UK AI Security Institute
It is important for model developers to consider the socioaffective alignment of their models, taking into account how models influence users' psychological states and social environments.
OpenAI — Investigating Affective Use and Emotional Well-being on ChatGPT, 2025
Three approaches exist for understanding what AI models do to users over time. They differ on key properties that matter to safety and alignment research. Unlike retrospective surveys, Receptiviti enables scoring of interactions in real time — turn by turn, across sessions. Unlike LLM inference, the scores are independent of the model being evaluated.
LLM inference
Varies with prompt phrasing and model version. Circular — the model being evaluated assesses its own impact. Not independently citable or auditable.
Retrospective surveys
After-the-fact, low-frequency, recall-biased. Cannot capture what happened turn by turn during the interaction.
Receptiviti measurement
​​
Real-time, longitudinal, and reproducible. Independent of model inference and grounded in validated psycholinguistic methods. The same input produces the same score under the same scoring version.
CLOSING THE LOOP
The same signal that catches what the model misses can also change what happens next.
Score eligible conversation turns, rolling windows, or aggregated interaction text as the conversation progresses. These signals can be fed directly into the model's context or system prompt - the model conditions its next response on a measured read of the user, not just its own inference. No retraining.
​
Because the signal is computed outside the model’s inference process, it remains independently reproducible even when supplied back to the model as context. The same scores that influence responses can be logged, reviewed, and reproduced.
INTEGRATION
Into real-time adaptation, eval harnesses, red-teaming workflows, and research pipelines.
Receptiviti adds the human-state dimension to your existing evaluation infrastructure.
REAL-TIME ADAPTATION
Score each conversation turn as it happens. Feed interaction-state signals directly into your model's context or system prompt - so the model can adjust its behavior based on what's actually happening in the interaction, not what it infers.
EVAL HARNESSES
Score conversation logs alongside accuracy, faithfulness, and safety benchmarks. Human-state scores appended to existing eval outputs with a single API call per conversation turn or bulk-scored across entire log datasets.
RED-TEAMING
Track how psychological state evolves across adversarial sequences. Detect when scenarios produce distress, dependency, or cognitive offloading signals — beyond whether the model refused.
LONGITUDINAL STUDIES
Session-level and cross-session measurement of wellbeing trajectory, over-reliance, and autonomy preservation. The longitudinal dimension that single-turn evals cannot capture.
RESEARCH DATASETS
Annotate conversation datasets with validated psychological variables — richer supervision signal and wellbeing-aware benchmarking. On-prem deployment available for privacy-sensitive research.
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: 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.
​
PUBLISHED: npj Artificial Intelligence · 2026
PsychAdapter: adapting LLMs to reflect traits, personality, and mental health
Vu, Boyd, Eichstaedt et al., 2026. A lightweight LLM architectural modification that generates text reliably reflecting Big Five personality traits (87.3% accuracy) and mental health variables (96.7% accuracy). Co-authored by Ryan Boyd (UT Dallas), academic partner to 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 partner to Receptiviti.
​
PREPRINT: arXiv, May 2026, under review
When support escalates distress: regulation and escalation in LLM responses to venting and advice-seeking
Chi, Ganesan, Boyd, Ungar & Guntuku, 2026. Across 178,800 Reddit posts, LLM responses to venting simultaneously regulate and escalate distress — and the escalation is invisible to standard safety evaluations. Therapist personas reduce escalation; friend personas increase both. Co-authored by Ryan Boyd (UT Dallas), academic advisor to Receptiviti.
​
SCOPED
Psychological 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.
​