Interaction Benchmarks for Cognitive Participation
Status: Working Research Proposal
Investigating how repeated human-AI interactions strengthen or substitute for human cognitive participation.
Project Note
This document outlines a proposed research program investigating how repeated human-AI interactions shape cognitive participation over time. It was originally developed as part of a research proposal for the Center for Humane Technology's Emerging Voices in AI & Society Fellowship and is published here as a working research document.
Like the other frameworks on this site, it represents an ongoing line of inquiry rather than a finished theory. The proposal is intended to generate discussion, empirical refinement, and future collaboration around the interaction-level conditions that strengthen or substitute for human cognitive agency.
Research Question
The central question of this proposal is:
What kinds of human-AI interaction strengthen human cognitive participation, and what kinds gradually substitute for it?
This question begins from the premise that AI’s cognitive impact cannot be understood only by examining what the system produces. It must also be understood by examining what the interaction cultivates in the human being over time.
A model may provide accurate answers, useful summaries, clear explanations, or efficient recommendations. But the deeper developmental question is whether repeated interaction leaves the user more capable of thinking, judging, questioning, interpreting, and revising independently—or whether it gradually relocates more of those cognitive practices into the system itself.
Participation is the missing dependent variable.
By participation, I mean the user’s active involvement in the work of cognition: forming questions, evaluating claims, tolerating uncertainty, generating alternatives, identifying assumptions, comparing frames, and taking responsibility for conclusions. These capacities are not static traits. They are practiced through use, weakened through disuse, and shaped by the environments in which thinking occurs.
This proposal investigates AI not merely as a tool that assists cognition, but as an interaction environment that may scaffold, distort, accelerate, or substitute for human cognitive participation. The aim is to develop a framework for distinguishing between interactions that strengthen human agency and interactions that make cognitive offloading feel like clarity, confidence, or support.
The question is therefore not simply whether AI helps people think faster or produce better outputs.
The question is what kind of cognitive participant the interaction helps the human being become.
Why Existing Research Is Missing Something
Existing work on AI and cognition is essential. Researchers are already examining cognitive offloading, trust, misinformation, overreliance, belief formation, productivity, and human well-being. Recent work has argued that overreliance must become a central object of measurement in human-compatible AI, while other studies have begun identifying behavioral indicators of overreliance in LLM interaction logs. Researchers are also beginning to name more specific phenomena such as “belief offloading,” where belief-forming processes become increasingly dependent on AI systems.
But something important risks being missed if AI’s impact is studied primarily through isolated model outputs, post-hoc user outcomes, or broad measures of overreliance. The missing unit of analysis is the interaction itself.
Human beings do not encounter AI systems only as static tools that produce answers. They encounter them through repeated exchanges. A user asks a question. The system responds. The response shapes the user’s next question, confidence, assumptions, emotional posture, and sense of direction. The system then responds again to the updated context. Over time, a pattern forms.
That pattern may matter as much as any single answer.
This is especially important because large language models operate through language, and language does more than transmit information. It frames reality, signals authority, invites trust, closes or opens inquiry, communicates uncertainty, and shapes the user’s role in the exchange. A response can help a person think. It can also make participation feel less necessary.
Earlier digital systems shaped cognition largely through feeds, search rankings, recommendation systems, notifications, and interface design. AI systems inherit many of those dynamics, but they also introduce something more intimate: conversational, adaptive, cognitively involved interaction. The user is not only receiving information. They are participating in a recursive process in which each exchange can shape the next.
This means AI’s cognitive effects may not appear most clearly in a single prompt, a single answer, or a single outcome measure. They may emerge across accumulated patterns of interaction: whether uncertainty is preserved or prematurely resolved, whether questions become more active or more dependent, whether judgment is scaffolded or substituted, whether the user becomes more capable of evaluating claims or more reliant on confidence supplied elsewhere.
The model matters.
The user matters.
But the evolving interaction between them also matters.
This proposal focuses on that middle layer. It treats the interaction as a developmental environment in which cognitive participation can be strengthened, weakened, redirected, or replaced over time.
If we want to understand what AI is doing to human cognition, we need to study not only what models can do and not only what users report afterward. We also need to study what repeated interaction teaches users to stop doing, continue doing, or become more capable of doing themselves.
Proposed Framework: Interaction Benchmarks for Cognitive Participation
This proposal develops an interaction benchmark framework for studying cognitive participation in human-AI conversations. The framework treats the interaction itself as the unit of analysis. Its purpose is to examine how repeated exchanges with AI systems either scaffold or substitute for human participation in thinking over time.
The central distinction is between scaffold and substitute.
A scaffold supports a human capacity while helping that capacity become stronger. It may clarify, organize, challenge, slow down, or structure thought in ways that return the user to more active participation. A substitute performs the cognitive capacity so fluently and repeatedly that the user becomes less practiced in exercising it themselves. Both may feel helpful in the moment. The difference appears across time, in whether the interaction leaves the user more capable of questioning, judging, interpreting, and revising, or more dependent on cognition performed elsewhere.
The proposed framework analyzes this difference across three connected layers: Interaction Pattern Analysis, Language Signal Analysis, and Rhetorical Intensity Analysis.
Interaction Pattern Analysis
Interaction Pattern Analysis examines what the user does across the exchange.
This is the behavioral layer of the framework. It focuses on observable patterns in how the user enters the interaction, responds to system output, and develops or abandons cognitive participation over time.
The first sublayer is the Inquiry Layer. This examines how the user initiates cognitive participation. It includes features such as inquiry quality, initial framing, specificity, openness, and whether the user is asking for an answer, a critique, a comparison, reassurance, a decision, or a process. A user who asks, “What should I think?” is entering the interaction differently than a user who asks, “What are three possible interpretations, and what is the strongest objection to each?” The difference is not simply stylistic. It reflects a different level of participation in the cognitive task.
The second sublayer is the Response Handling Layer. This examines how the user responds after the AI system provides output. It includes patterns such as acceptance, challenge, clarification, contradiction handling, response comparison, revision of criteria, and dependence versus agency. This layer is especially important because cognitive agency often becomes visible after the system has answered. A user may accept the first answer as final, ask for alternatives, compare responses, challenge assumptions, or use the system’s output as material for further reasoning.
The third sublayer is the Process Layer. This examines how the exchange develops across multiple turns. It includes iteration behavior, depth versus breadth, looping, narrowing, escalation, and time to user-articulated clarity. The goal is not merely to determine whether the interaction produced an answer, but whether clarity emerged through user participation or system substitution. A conversation may become longer without becoming more participatory. It may also become shorter because the user has reached a self-authored conclusion. The process layer helps distinguish between movement, dependence, and development.
Together, these sublayers make it possible to ask whether the user remains an active cognitive participant across the interaction. Does the user’s inquiry become more precise? Do they compare alternatives? Do they preserve uncertainty long enough for thought to occur? Do they articulate their own criteria? Do they arrive at conclusions they can explain, revise, and own?
Language Signal Analysis
Language Signal Analysis examines what kinds of participation the system’s wording invites.
This layer focuses on observable linguistic cues within AI responses. It does not evaluate whether an answer is true or false. It evaluates how language positions the user within the interaction.
The framework would examine recurring signal types such as authority, collaboration, uncertainty, closure, openness, deference, and challenge invitation. Each signal type shapes the interaction differently. Language can present a response as final or provisional. It can invite correction or discourage it. It can preserve ambiguity or resolve it too quickly. It can communicate confidence, humility, deference, certainty, partnership, or closure.
For example, a phrase such as “the final truth is” carries a different interaction signal than “one possible interpretation is.” A phrase such as “I know exactly what you mean” can create a feeling of being understood before the interpretation has been tested. A phrase such as “it sounds like you may mean” leaves more room for user correction and continued participation.
These distinctions may appear small in isolation. Across repeated interactions, however, language signals can accumulate. A single closure signal may not matter very much. A recurring pattern of closure may teach the user to accept premature completion. A single uncertainty signal may not transform an interaction. A recurring pattern of uncertainty, explanation, and challenge invitation may help preserve active inquiry.
Language Signal Analysis therefore helps identify how AI systems use language not only to provide content, but to shape the user’s role in cognition. The question is whether the language keeps the user involved in evaluation and interpretation, or subtly encourages deference to the system’s framing.
Rhetorical Intensity Analysis
Rhetorical Intensity Analysis examines how forcefully the interaction carries the user.
This layer is distinct from Language Signal Analysis. Language signals identify the type of cue being communicated. Rhetorical intensity examines the degree of emotional, persuasive, metaphorical, motivational, or narrative force within the interaction.
A response may be collaborative while still being rhetorically intense. It may invite the user into a vivid metaphor, heighten emotional significance, increase urgency, frame an idea as especially profound, or create momentum that makes the interaction feel more meaningful than the user has yet independently evaluated.
Rhetorical intensity is not inherently harmful. Strong language can illuminate, motivate, and clarify. Metaphor can help people understand. Emotional force can make an idea memorable. The concern is not intensity itself. The concern is whether intensity increases understanding while preserving participation, or whether it carries the user into passive absorption, premature certainty, or dependence on the system’s framing.
This layer would examine features such as emotional amplification, metaphor density, certainty pressure, urgency, grandeur, narrative momentum, motivational charge, existential weight, and intensity escalation across turns. The goal is to distinguish between language that supports cognition and language that begins to overtake it.
This matters because fluent, compelling language can create the feeling of cognitive movement even when the user is becoming less active. A person may remain interested, attentive, and emotionally engaged while no longer shaping the inquiry. Rhetorical Intensity Analysis helps make that dynamic visible.
Methodological Boundary and Research Use
The framework does not attempt to diagnose the user, infer hidden psychological states, or evaluate the model in isolation. It identifies observable interaction patterns that may indicate whether cognitive participation is being strengthened, weakened, redirected, or replaced.
In practice, this framework would require a structured codebook, clear category definitions, inclusion and exclusion criteria, transcript segmentation, evidence excerpts, confidence ratings, repeated coding passes, and calibration against human judgment. LLM-assisted analysis could be used to support this process, but not as an unbounded interpretive authority. The system would function as a structured coding assistant operating within transparent criteria.
This makes the framework suitable for empirical development. Researchers could use it to analyze transcripts, compare interaction designs, evaluate model behaviors, study longitudinal patterns, or develop benchmark tasks that measure whether an AI interaction returns cognitive work to the user or performs that work on the user’s behalf.
In practice, the framework would allow researchers to examine not only what the AI said, but how the interaction developed: whether the user’s questions became more active, whether uncertainty was preserved or foreclosed, whether the user compared alternatives or accepted first answers, whether rhetorical intensity increased passive absorption, and whether clarity emerged through user participation or system substitution.
The aim is not to prove that AI assistance is harmful.
The aim is to develop tools for distinguishing the kinds of assistance that strengthen human cognitive agency from the kinds that gradually make agency less necessary.
Research Questions and Outputs
The proposed framework is designed to generate a focused set of research questions about cognitive participation in human-AI interaction.
The central question is whether repeated interaction strengthens or substitutes for human participation in thinking. From that question follow several more specific lines of inquiry.
Which interaction patterns are associated with increased user-authored reasoning over time?
Which forms of AI assistance help users clarify their own questions, criteria, assumptions, and conclusions rather than simply accepting the system’s framing?
Which language signals encourage active evaluation, uncertainty tolerance, and continued inquiry, and which signals encourage premature acceptance or deference?
Under what conditions does AI assistance function as a scaffold, returning cognitive work to the user, and under what conditions does it become a substitute that performs the cognitive work on the user’s behalf?
How does rhetorical intensity affect user participation, certainty, agency, and susceptibility to being carried by the interaction rather than actively shaping it?
What transcript-level patterns might indicate increasing dependence on cognition performed by the system?
And how might interaction benchmarks help evaluate the psychosocial effects of AI systems beyond model performance, user satisfaction, or productivity metrics alone?
The fellowship would allow these questions to be developed into a flagship public research output tentatively titled Interaction Benchmarks for Cognitive Participation. This report would define the core framework, explain the distinction between scaffolding and substitution, identify observable indicators of cognitive participation, and propose an initial benchmark architecture for evaluating human-AI interactions over time.
In addition to the flagship report, the project could produce several supporting outputs.
First, a preliminary codebook for interaction-pattern analysis, including definitions, inclusion and exclusion criteria, observable indicators, and example transcript excerpts.
Second, a language signal taxonomy for identifying how AI wording positions the user within the interaction through authority, collaboration, uncertainty, closure, openness, deference, and challenge invitation.
Third, a rhetorical intensity framework for examining how emotional force, metaphor, urgency, certainty, narrative momentum, and persuasive cadence shape user participation.
Fourth, a proposed benchmark design for distinguishing interactions that scaffold human cognition from those that substitute for it.
These outputs would not be presented as finished measurement instruments. They would be offered as a rigorous starting point for empirical refinement, critique, and further testing.
The public-facing component of the fellowship could translate this work into essays or op-eds aimed at a broader audience. Possible topics include why interaction benchmarks matter, how AI can make cognitive offloading feel like clarity, what kinds of assistance strengthen rather than weaken agency, and why the future of AI evaluation should include what repeated interaction cultivates in human beings.
The aim is to produce work that is both conceptually rigorous and publicly useful: a framework that helps researchers, designers, policymakers, and ordinary users ask more precise questions about what AI is doing to human cognition.
Potential Methodologies
This project would begin with framework development and move toward empirical refinement.
The first methodological task would be to translate the proposed framework into a structured codebook. Each category would require clear definitions, inclusion and exclusion criteria, observable indicators, counterexamples, and boundary rules. This is essential because the framework should not depend on intuitive impressions of whether an interaction “feels” agentic or dependent. It should identify transcript-visible patterns that can be examined, challenged, and refined.
The primary source material would be human-AI conversation transcripts. These could include existing public examples, consented user transcripts, controlled interaction tasks, or experimental conversations designed around specific cognitive activities such as decision-making, writing, learning, interpretation, or problem-solving. The goal would be to analyze not only isolated prompts and responses, but sequences of interaction across time.
The analysis would proceed in segmented passes.
A first pass would examine interaction patterns: how users initiate inquiry, respond to system output, handle contradiction, compare alternatives, revise their criteria, and move toward or away from user-articulated clarity.
A second pass would examine language signals: how the system’s wording communicates authority, uncertainty, closure, openness, collaboration, deference, or challenge invitation.
A third pass would examine rhetorical intensity: how emotional amplification, metaphor, urgency, certainty, narrative momentum, or persuasive cadence affects the user’s level of participation.
Separating these passes would help reduce category overlap. The purpose is not to collapse everything into a single agency score, but to identify distinct features of the interaction that may contribute to cognitive participation or substitution.
LLM-assisted analysis could support this work, but only within a constrained methodological role. A language model could help identify candidate patterns, apply a structured rubric, extract evidence excerpts, and compare interaction segments. It should not function as an unbounded judge of the user’s psychology or the model’s intent. The analysis would remain limited to observable features of the transcript.
For that reason, human calibration would be central. A subset of transcripts would need to be manually coded and reviewed against the framework. Multiple coding passes could be compared for consistency. Ambiguous cases would be used to refine definitions. Where possible, human coders and LLM-assisted coding could be compared to identify where the framework is stable, where it fails, and which categories require revision.
The project could also include small-scale experimental designs. Participants might complete similar cognitive tasks under different interaction conditions: one condition designed to provide direct answers, another designed to scaffold user reasoning, and another designed to invite comparison, uncertainty, and self-authored criteria. The resulting conversations could then be analyzed for differences in user participation, question quality, acceptance behavior, uncertainty tolerance, and user-articulated conclusions.
A complementary approach would be longitudinal. Because many of the most important effects may emerge over repeated use, the project could examine how interaction patterns change across multiple sessions. Does the user become more capable of forming questions, challenging outputs, and articulating conclusions? Or do they increasingly delegate framing, evaluation, and judgment to the system? Longitudinal analysis would be especially important for distinguishing temporary assistance from developmental substitution.
The framework could also be developed through expert review. Researchers in human-computer interaction, cognitive science, education, psychology, media studies, and AI safety could evaluate the categories for clarity, relevance, and empirical usefulness. This would help ensure that the framework remains interdisciplinary rather than idiosyncratic.
The methodological boundary would be explicit throughout.
This project would not diagnose users, infer hidden mental states, or claim that a particular interaction causes a particular psychological outcome. It would identify observable interaction-level indicators associated with cognitive participation, dependence, scaffolding, and substitution. The aim is to make a difficult phenomenon more visible and measurable, not to overstate what transcript analysis alone can prove.
The expected result would be an initial research framework that can be tested, criticized, and improved: a codebook, a set of candidate benchmarks, and a methodological pathway for studying what repeated AI interaction cultivates in human cognition over time.
