Search

Measuring Psychological Change From Language, Without Asking How People Feel

Receptiviti Labs

1 hour ago

2 min read

Psychological changes can be measured from what people say without asking them how they are feeling. Several years ago, Receptiviti tested this approach at scale by tracking psychological changes in the language of thousands of healthcare workers over time.

During the COVID-19 pandemic, Receptiviti analyzed Reddit comments from approximately 6,500 self-identified doctors and nurses. They weren’t recruited to complete surveys or asked about their mental wellbeing. Instead, their naturally occurring language from their Reddit posts was analyzed to measure changes in their cognitive load, analytical thinking, anger, fear, anxiety, stress, and empathy. Their comments were aggregated over time, making it possible to track psychological changes as they developed.

Cognitive load measured from the naturally occurring language of doctors and nurses over time

The same underlying approach has significant applications for human-AI interaction safety. If we want to understand how interacting with AI affects people, surveys and self-reports provide one source of evidence. The language people produce during those interactions provides another that enables psychological change to be measured continuously over time.

Measuring Psychological Change From Language Over Time

Among doctors, cognitive load rose after the pandemic began, peaked in 2021, and remained elevated through 2023. Nurses showed a smaller increase that subsequently declined.

These changes were measured from language people were already producing, without asking them to report whether they were becoming more cognitively overloaded, less analytical, or more anxious.

The complete ⁠Healthcare Workers Mental Wellbeing Index tracks cognitive load, analytical thinking, anger, fear, anxiety, stress, and empathy through October 2023.

The data were observational and population-level, so the measured changes cannot be attributed solely to COVID-19. Other factors may also have contributed to the changes observed over the study period.

A New Measurement Channel for AI Evaluation

AI conversations create a new opportunity to apply the same underlying approach at the interaction level. As users generate language across conversations, psychological measures could provide an additional user-side signal alongside surveys, model metrics, and task performance.

Because the measurement method is independent of the AI model, it can provide a consistent measurement framework across models and versions. In an evaluation harness, these measures could complement traditional model and task metrics with a user-side signal, helping teams evaluate both how the system performs and how the person interacting with it changes. Applying this reliably to individual AI users requires accounting for factors such as the amount of language available, task and topic, and validating that measured changes correspond to meaningful changes in the user.

As AI evaluation increasingly considers the effects systems have on the people using them, and this method provides another source of evidence for measuring the human side of AI interactions.

Measuring Psychological Change From Language, Without Asking How People Feel

Psychological change can be measured from naturally occurring language without relying solely on self-report. Receptiviti demonstrated this at scale by tracking psychological measures in the language of thousands of healthcare workers over time.

Two different user trajectories from the same AI system, one is beneficial and the other is detrimental.

AI Safety Evaluations Should Include User Trajectories

Current AI safety evaluations are increasingly capable of assessing model behavior across multi-turn conversations, but they provide much less visibility into what those interactions are doing to users. This article argues that evaluating model behavior alone is insufficient when psychological harms develop gradually over time, and proposes that user trajectories should become a complementary part of AI safety evaluation.

SIM-VAIL and What AI Safety Evaluations Don’t Measure

SIM-VAIL is an important advance in AI safety evaluation, identifying chatbot behaviors associated with psychological risk. This analysis examines what the framework measures, what it does not measure, and why evaluating AI behavior differs from measuring psychological effects on users.