Psychometric Correlates of AI Interaction Styles

Overview

"AI interaction style" is not a single construct. In the current literature it is distributed across at least five partially overlapping outcome families: general attitudes toward AI, advice-taking and reliance, anthropomorphism and felt social connection, problematic or companion-style use, and performance in AI-assisted tasks. That matters because a trait that predicts AI acceptance in a questionnaire study is not thereby shown to predict prompt length, politeness, verification behavior, or emotional dependence on a general-purpose assistant. The field already supports a modest psychometric thesis—stable individual differences do shape how people relate to AI—but the strongest evidence is on coarse-grained tendencies, not prompt microstructure. (PMC)

At a high level, the most defensible claims are these. Agreeableness is associated with more positive AI attitudes and, in one substantial advice-taking study, greater use of AI advice, but there is no robust direct evidence that agreeable users write longer or more verbally softened prompts. Openness to experience predicts AI trust and acceptance more consistently than it predicts calibrated trust. Attachment anxiety matters most when AI is used in a relational frame—companionship, counseling, or emotionally supportive interaction—but the direction of the effect depends heavily on context. Analytical-thinking constructs such as Need for Cognition, the Cognitive Reflection Test, and Actively Open-Minded Thinking are more informative about selective versus indiscriminate reliance than about raw enthusiasm for AI. Conscientiousness appears more tightly coupled to disciplined or ethical AI use than to simple "resistance" to AI suggestions. (PMC)

This entry pairs naturally with Big Five Personality Traits, HEXACO Personality Model, Human-AI Reliance, and Trust Calibration in Human-AI Systems.

Evidence at a glance

Construct	Closest measured AI behavior	What the evidence currently supports	Representative sources
Agreeableness	AI attitudes; advice-taking; small early-adopter task studies	More positive AI attitudes and more use of AI advice; not robustly linked to prompt verbosity itself	(PMC)
Openness to experience	AI trust; AI acceptance; attitudes; advice-taking	Higher trust and acceptance in broad AI surveys, but one advice-use study found less AI advice use; evidence for trust calibration is thin	(Springer Link)
Attachment anxiety	General AI trust; counseling adoption; problematic CAI use; academic overreliance	Context-sensitive: lower generalized trust in AI in one line of work, but higher adoption or more problematic/overreliant use in relational contexts	(ScienceDirect)
Anthropomorphism tendency	Social connection to chatbots; problematic use pathways	Stronger felt connection after chatbot interaction and stronger vulnerability to relational/problematic use pathways	(PMC)
CRT / NFC / AOT	AI-assisted problem solving; overreliance; appropriate reliance	Better performance and more selective use are associated with reflective/effortful cognition; AOT looks more promising than Big Five for overreliance	(Springer Link)
Conscientiousness (and related traits)	Misconduct; structured AI-assisted tasks; advice-taking	Better predictor of disciplined or ethical use than of blanket skepticism; no clear effect on AI advice use in the advice-taking study	(Nature)

The table is already a warning about the state of the field. Most rows summarize proxies for "interaction style," not direct behavioral measurement of how people prompt and converse with assistants. A digital-counseling vignette, a dermatology classifier, a student misuse inventory, a 6-minute GPT-3.5 chat, and platform-scale voice-mode telemetry are not interchangeable settings. (PMC)

What the field actually measures

A useful meta-result is that AI interaction traces do carry psychometric signal. Peters, Cerf, and Matz report that GPT-4 could infer Big Five traits from free-form user interactions with moderate accuracy, but the quality of the inference depended strongly on conversational context: mean correlation was .443 in personality-eliciting conversations, .218 in a more naturalistic setting, and only .117 when the bot behaved like a default helpful assistant. That does not show that personality determines any single prompt feature, but it does show that conversational behavior with AI is psychometrically informative rather than random residue. It also implies that assistant-side defaults can suppress or reveal trait expression. (arXiv)

That point matters because the literature often jumps from broad attitudes to claims about user style. In reality, "attitudes," "advice use," "trust," "appropriate reliance," "problematic use," and "social connection" are distinct dependent variables with different causal structures. A trait that predicts willingness to try AI may have a very different relation to whether a person checks citations, rejects a wrong recommendation, or returns to the system for emotional comfort. The field will remain conceptually muddy until those outcomes are separated rather than bundled under one umbrella label. (PMC)

Big Five and HEXACO: the baseline picture

On broad attitudes toward AI, the cleanest Big Five signal so far is agreeableness. In Stein et al.'s Study 3 (N=298, MTurk US), agreeableness was the only Big Five trait that significantly predicted more positive attitudes toward AI in the main regression; openness was only marginal in the final model, while conscientiousness, extraversion, and neuroticism were not significant predictors. Stein et al. themselves framed the broader literature as thin and fragmented, with much prior work tied to narrow AI applications or embodied robots rather than AI as a general concept. (PMC)

The HEXACO literature pushes in a somewhat different direction. In Liang et al.'s Scientific Reports study, HEXACO Extraversion and Openness to Experience were significant positive predictors of favorable attitudes toward generative AI, while Honesty–Humility, Agreeableness, and Conscientiousness were significant negative predictors of generative-AI-related academic misconduct. Importantly, once personality traits were in the model, attitudes toward GAI added little explanatory power for misconduct. That is a strong hint that personality may matter more for how people use AI ethically or unethically than for whether they simply say they like it. (Nature)

A small early-adopter study by Kovbasiuk et al. gives a more behavioral, but also much weaker, signal. Their pilot involved 62 participants in AI-assisted versus non-AI-assisted tasks and found trait-by-chatbot interactions rather than simple main effects: agreeableness and conscientiousness changed how much the chatbot helped on specific tasks and how strongly participants intended to use the technology afterward. This is interesting because it points to task-contingent moderation rather than stable "AI personality types," but the study is small and should not be over-generalized. (Emerald Publishing)

The broad takeaway is that personality correlates with AI behavior are real, but they are not large or uniform enough to justify the folk taxonomy that online discourse often implies. Algorithm Aversion and Human-AI Reliance are likely better organizing concepts than "the personality of AI users" taken in the abstract. (PMC)

Prompting verbosity and agreeableness

Here the literature is weakest. I did not find robust work directly regressing real prompt length, verbosity, hedging, politeness markers, or clarification-turn frequency against agreeableness in LLM assistant use. That gap should be stated plainly. The current literature can support weaker claims about AI uptake and cooperative use, not a clean claim that agreeable users write longer or softer prompts.

What we do have is adjacent evidence. Stein et al. found agreeableness to be the clearest positive Big Five predictor of general AI attitudes. In Asbach, Graf-Vlachy, and Fügener's cross-sectional study with 595 participants, agreeableness and neuroticism were associated with increased use of AI advice, while extraversion and conscientiousness were not, and openness was associated with decreased use of AI advice. That pattern is consistent with agreeable users being more willing to incorporate machine input, but it says nothing directly about verbosity. (PMC)

Kovbasiuk et al.'s early-adopter pilot adds nuance. In their persona-building task, lower and moderate agreeableness benefited more from chatbot assistance than very high agreeableness; but for future intention to use the technology, highly agreeable participants in the AI-assisted condition looked more favorable than highly agreeable participants in the non-AI condition. That is an awkward but informative pattern: agreeableness may increase willingness to continue working with AI even when it does not maximize task improvement on every prompt-sensitive task. (Emerald Publishing)

The reverse-direction evidence from Peters et al. is also relevant. If GPT-4 can infer personality from free-form interactions, some combination of user wording, disclosure, pacing, topic choice, and repair behavior must be carrying stable trait information. But Peters et al. do not decompose that signal into prompt length or politeness, and the signal drops sharply in default-assistant interaction. So the safe conclusion is narrow: agreeableness appears related to cooperative uptake of AI more than to any demonstrated prompt-level verbosity pattern. The microbehavioral claim remains open. (arXiv)

This is an obvious place for a future entry on Prompt Engineering to connect with psychometrics. Right now, the bridge is mostly missing.

Trust calibration and openness to experience

The first conceptual correction is that trust calibration is not the same thing as AI trust. Calibrated trust means matching reliance to the model's actual competence—accepting correct help and rejecting incorrect help. Much of the openness literature measures a looser construct: favorable attitudes, baseline trust, or willingness to adopt AI. That distinction is explicit in the clinical reliance literature, which separates trust, reliance, and appropriate reliance. (PMC)

On the broad trust-and-acceptance side, the evidence for openness is decent. In Jiang et al.'s survey of 716 university students, openness to experience was positively associated with AI trust and AI acceptance, and AI trust significantly mediated the relation between openness and AI acceptance; AI anxiety alone did not significantly mediate the relation. This is a useful result, but it supports an "openness → trust/acceptance" pathway, not an "openness → calibrated reliance" pathway. (Springer Link)

HEXACO evidence points the same way on attitudes. Liang et al. found openness and extraversion to be positive predictors of favorable GAI attitudes. But the Asbach advice-taking study complicates the story: openness there was associated with decreased use of AI advice. That is not necessarily a contradiction. It could mean that openness predicts curiosity about AI as a technology while also predicting more exploratory or less deferential behavior in concrete advice-taking tasks. The key point is that "openness to AI" and "openness to accepting AI advice" are empirically different outcomes. (Nature)

The most direct trust-calibration evidence I found comes from Küper et al.'s study of 223 dermatologists using AI-enabled clinical decision support. They explicitly distinguish trust, reliance, relative AI reliance (RAIR), and relative self-reliance (RSR). Participants relied more on correct than incorrect advice, which is a minimal sign of calibration, but trust in the system still predicted greater reliance, more RAIR, and less RSR. The study did not identify openness as the operative trait; instead, propensity to trust technology and medical experience did more explanatory work. This is exactly why the calibration literature should not be collapsed into generic personality-attitude work. (PMC)

A more general human–machine-trust study also suggests that personality effects on trust are modality-sensitive. In that work, extraversion predicted static trust across PC and VR modalities, while emotional stability predicted dynamic trust in the PC condition but not in VR. Again, that is evidence that personality matters for trust formation, but not a clean route from openness to well-calibrated reliance on contemporary assistants. (PMC)

So the defensible summary is this: openness to experience predicts trust and acceptance of AI more reliably than it predicts good calibration of trust. Current evidence does not justify the stronger claim. For that topic, Trust Calibration in Human-AI Systems is the more appropriate anchor concept.

Anthropomorphization and attachment anxiety

This is the most conceptually rich part of the literature, and also the easiest place to overstate the findings. Anthropomorphism means attributing human-like qualities—mind, intention, care, emotion—to nonhuman systems. Reviews of chatbot design note that anthropomorphic cues in appearance and language are now ordinary, not exceptional, and that users routinely anthropomorphize chatbots even outside deliberately companion-oriented products. (PMC)

Attachment anxiety enters the picture in a surprisingly context-dependent way. In Gillath et al.'s multi-study work on generalized trust in AI, attachment anxiety predicted less trust in AI, experimentally increasing attachment anxiety reduced trust, and increasing attachment security raised trust. On its face, that looks like a simple "anxious attachment → mistrust of AI" story. (ScienceDirect)

But that is not the whole story once AI is framed as a relational or supportive agent. Wu, Liew, and Dorahy surveyed 239 American adults about conversational AI for mental-health support; participants had not previously used CAI counseling. Dispositional trust in AI strongly predicted adoption intention, and attachment anxiety—contrary to the authors' initial expectation—was positively associated with the intention to adopt CAI counseling after controlling for age and gender. The authors explicitly note that the trust–attachment relation did not behave here as it had in generalized-AI research, suggesting that attachment effects are domain-specific and may depend on whether the interaction activates relational needs. (PMC)

Heng and Zhang push this much further into dependency risk. In their 2025 study of 54 Chinese adults who used conversational AI, attachment anxiety directly predicted problematic use of conversational AI and also indirectly predicted it through emotional attachment to the AI; anthropomorphic tendency moderated the relationship. That is one of the most direct demonstrations currently available of an attachment-related pathway into problematic CAI use. (Dove Medical Press)

A separate academic-use study by Athar found a related pattern in a different domain. Using a newly validated Academic AI Usage Inventory, Athar identified constructive, overreliant, and irresponsible AI-use profiles among 171 university students. Anxious attachment was associated with overreliant and irresponsible AI-usage profiles. This matters because it shows attachment patterns are not confined to explicitly companion-style or counseling contexts; they can also show up in ordinary generative-AI use when the relevant behavior is heavy dependence or weak verification. (ScienceDirect)

Anthropomorphism itself also behaves like a meaningful individual-difference variable. In Folk et al.'s two experiments (N=1274 total, one preregistered), people higher in anthropomorphism of technology felt more socially connected after interacting with a chatbot than after journaling, and the chatbot condition benefited most once anthropomorphism crossed a threshold. That is a strong indication that social payoff from AI companionship is not uniform across users; it depends materially on whether the user treats technology as mind-like in the first place. (PMC)

Related evidence from conversational-AI pathology points the same way. Hu, Mao, and Kim's 516-user study found that social anxiety was positively associated with problematic use of conversational AI, and that mind perception intensified the positive association between social anxiety and problematic use. This is not an attachment-anxiety paper, but it reinforces the broader principle that anthropomorphic mind attribution amplifies relational vulnerability to CAI. (ScienceDirect)

Industry evidence is useful here because it is more behavioral. OpenAI's 2025 affective-use study combined privacy-preserving analysis of more than 4 million ChatGPT conversations with a survey of more than 4,000 users and controlled testing. OpenAI reports that emotional engagement is rare overall and concentrated in a small heavy-use subgroup; people with stronger tendencies toward attachment in relationships and those who viewed ChatGPT as a friend were more likely to experience negative effects, with prolonged daily use associated with worse outcomes. The study also warns against overgeneralizing from average effects because the affective-use subgroup is small and atypical. (OpenAI CDN)

The right synthesis is therefore not "attachment anxiety makes people anthropomorphize AI" or "attachment anxiety reduces trust in AI." Both are too simple. The stronger claim is that attachment insecurity increases the importance of AI's relational affordances, and the behavioral result—mistrust, adoption, companionship, or dependence—depends on whether the AI is framed as a tool, a collaborator, a counselor, or a companion. This section belongs directly beside Anthropomorphism in AI and AI Companionship. (ScienceDirect)

Prompt engineering skill and analytical thinking

This is another area where the evidence is thinner than the discourse. There is plenty of writing about how to engineer prompts well, but surprisingly little rigorous work that treats prompt engineering skill as an individual-differences variable and asks whether it is predicted by standard psychometric measures. Most studies instead treat prompting as an intervention, a training problem, or a task design variable rather than as a measurable human capability. The current empirical bridge from analytical style to prompt skill is therefore indirect. (Springer Link)

The best current evidence comes from AI-assisted problem solving. In Moșoi et al.'s study of students solving economic problems with optional AI use, higher Cognitive Reflection Test scores, higher Need for Cognition, and greater AI use were each associated with better problem-solving performance. At the same time, the interaction between AI use and cognitive reflection significantly dampened the positive CRT–performance relationship, and the discussion interprets high-CRT/high-NFC students as tending toward more autonomous problem solving rather than easy delegation to AI. That is not a direct prompt-engineering measure, but it does support a model in which reflective thinkers use AI more selectively and gain less from uncritical offloading. (Springer Link)

Swaroop et al. make the selective-reliance story sharper. In their work on personalizing AI assistance based on overreliance, the authors argue that overreliance is difficult to predict from Big Five traits and Need for Cognition alone; they suggest that Actively Open-Minded Thinking (AOT)—defined there as willingness to consider different opinions—may be a better target. In experiment 2, overreliers had significantly lower AOT and higher agreeableness, while broader Big Five/NFC effects were weak or inconsistent, leading the authors to emphasize that trait-based prediction of overreliance is difficult and that overreliance may be task- and strategy-dependent rather than a stable personal trait. (Krzysztof Gajos)

Küper et al.'s dermatologist study shows the other side of the coin: Need for Cognition did not significantly affect trust or reliance in their simplified classification task. The authors explicitly suggest that the task may have been too cognitively thin—participants received only the AI's final classification and no explanation—so high-NFC users were not given much to engage with. That is an important methodological lesson: analytical traits become visible when the interface permits critical engagement, not when the model is reduced to a black-box answer source. (PMC)

So the strongest current claim is modest. Analytical-thinking variables predict something like epistemic discipline—the tendency to inspect, verify, and use AI conditionally—more than they predict generic enthusiasm for AI. Whether that cashes out as "prompt engineering skill" depends on the task. For open-ended research, coding, or writing assistants, the likely relevant capabilities are framing, decomposition, error checking, and iterative repair. The literature is not yet measuring those cleanly enough. This section belongs with Need for Cognition, Cognitive Reflection Test, Actively Open-Minded Thinking, and AI Literacy. (Springer Link)

Resistance to AI suggestions and conscientiousness

Resistance to AI suggestions is one of the most slippery constructs in this area because resistance can be either healthy skepticism or counterproductive underreliance. The clinical-reliance literature is correct to insist on that distinction: the relevant target is appropriate reliance, not maximal compliance and not maximal defiance. (PMC)

For conscientiousness specifically, the most direct advice-taking evidence is underwhelming. In Asbach et al.'s 595-participant cross-sectional study, conscientiousness did not appear related to the use of AI advice, while agreeableness and neuroticism predicted more AI advice use and openness predicted less. That directly weakens the popular intuition that conscientious people simply resist AI more. At least in this advice-taking setting, they did not. (IDEAS/RePEc)

Behavioral calibration work suggests that expertise and trust propensity may matter more than conscientiousness anyway. In Küper et al.'s dermatologist study, participants usually stuck with their own decisions: mean relative AI reliance was only 10.04%, whereas relative self-reliance was 85.6%. Trust predicted more reliance and more RAIR, while medical experience predicted more self-reliance. This is not evidence about conscientiousness per se, but it shows that "resistance to AI suggestions" is often driven by domain expertise and trust structure rather than generic Big Five factors. (PMC)

Where conscientiousness-like traits do show up is in disciplined or ethical use. Liang et al. found HEXACO Conscientiousness to be a significant negative predictor of GAI misconduct. Athar's AAUI study found anankastia positively linked to constructive AI use and negatively to irresponsible use. Kovbasiuk et al.'s small pilot found that moderate and high conscientiousness improved product-name task quality in the AI-assisted condition. These are not resistance effects; they are better interpreted as signals of structured, norm-sensitive, or goal-directed use. (Nature)

The cleanest present summary is therefore: conscientiousness is more clearly related to ethical and disciplined integration of AI than to simple refusal of AI advice. The literature does not yet support a strong "conscientious users resist AI suggestions" claim. It supports a weaker, but more interesting, claim that conscientiousness-like traits help determine whether AI is used carefully rather than sloppily. That is closer to Responsible AI Use than to a personality caricature of the skeptical user. (IDEAS/RePEc)

Methodological challenges

Proxy outcomes dominate. A major share of the literature measures attitudes, intentions, or self-reported trust rather than real conversational behavior. Stein et al. measure broad AI attitudes; Jiang et al. measure openness, trust, anxiety, and acceptance; Wu et al. measure counseling adoption intention; Athar measures usage profiles in academia. Those are all meaningful, but they are not direct transcript-level measures of prompting style. (PMC)

Self-report bias is everywhere. Many studies are cross-sectional questionnaire designs, often with the standard battery of Likert-style scales. Jiang et al. explicitly discuss common-method bias and use Harman's single-factor test; Wu et al. note that trust-in-AI measures do not cleanly separate user trust potential from perceived system trustworthiness. None of this invalidates the findings, but it means the field is still heavily dependent on what people say about AI rather than what they do with it. (Springer Link)

Outcome heterogeneity is not a nuisance; it is the central problem. "Trust," "acceptance," "advice use," "appropriate reliance," "social connection," "problematic use," and "misconduct" are different constructs. Küper et al. are especially valuable here because they insist on separating trust, reliance, RAIR, and RSR. Without that conceptual precision, claims like "openness predicts trust" or "agreeableness predicts AI use" get overgeneralized far beyond what the underlying dependent variable warrants. (PMC)

Novelty effects are baked into many designs. Some studies are literally about early adopters. Others present users with hypothetical or first-time counseling scenarios. Folk et al.'s chatbot interaction lasted six minutes; Wu et al.'s participants had not previously used CAI counseling; Kovbasiuk et al. position their work at the early stage of AI adoption. Those designs can reveal first-impression structure, but they do not establish durable, long-run interaction styles. (Emerald Publishing)

Rapid tool evolution makes replication unusually hard. The studies surveyed here span GPT-3.5-based warm chatbots, voice-mode interactions, digital-counseling vignettes, academic-use inventories, and clinical decision-support classifiers. OpenAI's own affective-use work emphasizes that even within one product family, text and voice usage can differ and effects are not uniform. A psychometric correlate measured on one assistant configuration may not transfer cleanly to another because the assistant-side behavior has changed. (PMC)

Some "interaction styles" may be strategic rather than trait-like. Swaroop et al. explicitly argue that overreliance may not be a stable trait, but may depend on the task, the form of AI assistance, and how willing the participant is to engage on a given day. That is a serious challenge to simplistic personality explanations. It suggests that some recurrent user differences are better modeled as learned strategies or situational states than as Big Five expressions. (Krzysztof Gajos)

Measurement itself is inconsistent. Wu et al. call out a lack of uniformity in how trust in AI is defined and measured, and Zhu et al.'s trust-in-machines study notes the limitation of using a single-item trust measure. Meanwhile, many studies use broad domain scores—Big Five or HEXACO dimensions—rather than facets, which makes it hard to distinguish, for example, cautious dutifulness from achievement striving inside "conscientiousness." The current trait instruments may simply be too coarse for prompt-level questions. (PMC)

What a stronger research program would look like

The field needs opt-in transcript-level behavioral data tied to psychometrics. The obvious missing design is a longitudinal study in which users complete validated trait batteries and then consent to analysis of real prompt logs, repair turns, acceptance/rejection behavior, citation checking, model-switching, and tool use over time. Peters et al. already show that free-form interactions carry personality signal; the next step is to identify which behavioral features actually mediate that signal. (arXiv)

It also needs multi-context replication. A trait effect that appears in static text prompting may disappear in voice mode, in persistent-memory agents, or in high-stakes settings where users have strong priors and professional expertise. OpenAI's affective-use study and the PC/VR trust literature both make the same point from different angles: modality changes the interaction. A future psychometric science of AI use has to treat modality as a first-class variable, not a footnote. (OpenAI)

Finally, the field needs to stop treating user traits as the whole story. Peters et al. show that assistant defaults can dampen personality signal; Folk et al. show that a warm, follow-up-heavy GPT-3.5 chatbot can induce social connection; OpenAI's own data show that affective outcomes are concentrated in a small subgroup. The right unit of analysis is probably the human–model dyad: user traits, assistant persona, interface modality, and task structure jointly produce the observed interaction style. (arXiv)

Bottom line

There is already enough evidence to reject the view that interaction style with AI assistants is just random UI noise. Personality, attachment style, and cognitive style do shape how people trust AI, when they heed it, whether they anthropomorphize it, and how vulnerable they are to overreliance or emotional attachment. But the literature is still too indirect to support confident claims about transcript-level prompt style—especially verbosity—and too outcome-fragmented to support clean trait stereotypes. The strongest current claims concern acceptance, reliance, anthropomorphic connection, and misuse; the weakest concern micro-level prompting behavior and prompt engineering skill as psychometric traits. (PMC)

Related entries that would make natural companions here are Prompt Engineering, AI Literacy, Trust Calibration in Human-AI Systems, Human-AI Reliance, Algorithm Aversion, Anthropomorphism in AI, AI Companionship, Need for Cognition, Cognitive Reflection Test, Actively Open-Minded Thinking, Big Five Personality Traits, HEXACO Personality Model, and Sycophancy in Language Models.