Behind the screen: laypersons' mental models about ChatGPT

ChatGPT's conversational interface invites anthropomorphic interpretations that conflict with its underlying probabilistic architecture. Prior research has identified folk theories and conceptions of generative AI among technical laypersons, but has largely examined these mental concepts in isolation or at the group level. Less is known about how conceptualizations of ChatGPT as a deterministic, probabilistic, anthropomorphic or dynamic AI system coexist, interact, and contradict one another within technical laypersons. This study examined (RQ1) recurring qualitative mental model configurations of ChatGPT among participants and (RQ2) whether technical AI terminology was supported by mechanism-specific understanding or coexisted with misconceptions concerning the same or another functional mechanism. We conducted semi-structured interviews with technical laypersons (N = 21) to ask them about their understanding of how ChatGPT works. Statements were deductively coded into four mental concepts (deterministic, probabilistic, anthropomorphic, and dynamic). Mental models were constructed for each participant. Statements containing technical terminology were additionally classified according to whether the term was contradicted by the participant's explanation of the same mechanism (“Chauffeur Knowledge”) or co-occurred with a misconception in another mechanism. Three recurring mental model configurations emerged: (i) Predominantly Probabilistic Thinking Patterns (n = 8); (ii) Anthropomorphic and Dynamic Thinking Patterns (n = 4), characterized by limited probabilistic reasoning; and (iii) Predominantly Deterministic Retrieval-based Thinking Patterns (n = 9), characterized mainly by deterministic conceptions alongside limited probabilistic awareness. All participants showed at least one misconception about ChatGPT and among 10 participants, who used technical terminology in their explanations, seven showed chauffeur knowledge, seven demonstrated correct or partially correct understanding of one mechanism alongside a misconception about another, and four participants met both criteria. Participants' mental models of ChatGPT are fragmented and often internally contradictory. Technical terminology may mask misunderstanding of the mechanism it describes or coexist with correct understanding in one domain and misconceptions in another. These findings challenge AI literacy assessments that rely predominantly on knowledge of technical terminology. Identifying recurring qualitative configurations provides a basis for developing measures that capture the complexity of mental models and may improve predictions of over-reliance on ChatGPT output.

“Last time I asked ChatGPT a question about quantum mechanics, it replied super-fast. So, it must have a very deep knowledge about physics, right?” The human tendency to navigate unfamiliar environments by projecting human-like traits onto non-human agents is a fundamental cognitive heuristic (Waytz et al., 2014). When laypeople interact with ChatGPT, this anthropomorphizing tendency is amplified by the system's conversational architecture (D'elia et al., 2026). As illustrated by the fictional quote above, the rapid, coherent response to a complex query about quantum mechanics may give the user the impression that ChatGPT possesses deep, expert-level understanding. In reality, however, the system merely predicts statistically probable next words based on patterns in its training data, lacking any genuine comprehension of the subject matter (Ray, 2023). This cognitive mismatch between the tool's interface and its underlying architecture—which operates as an opaque, probabilistic black box inexplicable even to experts (Bauer et al., 2023)—creates a significant gap between how users intuitively interpret AI behavior and how the system actually functions (Li et al., 2026).

ChatGPT is built on a large language model (LLM) and designed for human dialogue. However, because it relies on patterns without independent criteria of truth to verify information (Shanahan, 2024), the system frequently generates false information, termed “hallucinations” (Borji, 2023). While this limitation is typically well-understood by domain experts, ChatGPT is widely used among the general public, with more than 800 million weekly active users reported in 2025 (Singh, 2026). Understanding this limitation requires more than familiarity with terms such as “large language model,” “probability,” or “hallucination” Users may reproduce technically appropriate vocabulary without understanding why ChatGPT produces plausible but false information or how such outputs should be evaluated. Nevertheless, technical terminology is one of the most common indicators of AI knowledge or literacy (Ng et al., 2021; Jin et al., 2025). It therefore remains unclear whether the use of technical terminology reflects a functional understanding of ChatGPT that enables users to recognize its limitations and respond appropriately to erroneous outputs.

Identifying a lack of functional understanding, however, is particularly important because flawed conceptions of how ChatGPT works pose several risks for users, including misuse, overreliance on AI systems, and privacy risks (Weidinger et al., 2022; Bauer et al., 2023). A recent review has shown that the overreliance on and overestimation of AI-chatbots' capabilities is systematic among technical laypersons (Schneller et al., 2026). Horowitz and Kahn (2024) even draw a connection to the Dunning-Kruger effect, suggesting that individuals with limited background knowledge are less able to recognize the probabilistic and error-prone nature of LLMs. As a result, they tend to perceive AI systems as infallible authorities, and consequently have a drastically higher error rate when engaging with them. Unlike with other new technologies, in the context of AI systems, the group of technical laypersons also includes young adults and digital natives who, despite their daily use of digital technologies, struggle to understand the complex and constantly evolving AI systems.

To make sense of the world, individuals construct informal cognitive frameworks, or mental models (Johnson-Laird, 1995). When interacting with unfamiliar environments and novel technologies, mental models provide guidance and help individuals organize experiences (Plaks et al., 2009). The concept of mental models traces back to Kenneth Craik, who argued that people construct “small-scale models” of the real world based on their perceptions (Craik, 1943). These internal representations are driven by implicit processes guided by certain framework conditions, namely one's goal, existing knowledge and beliefs on the subject (Johnson-Laird, 1995, 2001). Developed through instruction, behavioral observation or socio-cultural knowledge exchange, including media exposure, mental models influence perceptions and help individuals to predict events, draw conclusions, and devise action plans (Plaks et al., 2009). Rather than serving as precise scientific truths, mental models are functional working representations used to simulate outcomes, reduce cognitive load, and guide action (Hemforth and Konieczny, 2006; Rezwana and Maher, 2025).

In practical use, mental models are particularly important when generative AI systems make errors. Whilst hallucinations may depend on the wording of an input, they can also persist due to model-inherent properties relating to training data, architecture or inference behavior (Anh-Hoang et al., 2025). Users' mental models can therefore influence whether an erroneous response is attributed to their own input, to retrievable but incorrect information, or to an inherent limitation of probabilistic generation. The extent to which a user is able to categorize and evaluate such an error then influences whether they will rephrase their input, verify the output externally, or continue to rely on the response.

While cognitive science has long studied the internal representations of traditional technical systems, and research on mental models of AI has grown (Villareale et al., 2022; Horstmann et al., 2023; Dain, 2025; Rezwana and Maher, 2025), empirical efforts to capture the explanatory logic of laypeople regarding generative AI-chatbots remain limited. Current state-of-the-art research reveals that users heavily rely on anthropomorphic attributions—the tendency to project human-like traits, such as consciousness, intent, or emotion, onto the system. Colombatto and Fleming found that only one-third of participants rejected the possibility that ChatGPT could experience consciousness, with a majority attributing some degree of phenomenal consciousness to the tool, and this tendency correlated positively with the frequency of use (2024). Another survey further highlighted pervasive uncertainty, with 30%−60% of participants often responding with “don't know” or “perhaps” when asked questions regarding the mechanisms of automated text generation tools (Lermann Henestrosa and Kimmerle, 2024). To overcome uncertainty, users often employ “metaphorical thinking” by associating ChatGPT either with familiar technological entities such as “Google” or “encyclopedias” (An et al., 2024), or with human-like attributions (Li et al., 2026), termed anthropomorphizing.

The recently published work by Li et al. (2026) represents a significant advance by utilizing focus group discussions to map AI users' “folk theories” across structural dimensions, including working mechanisms, perceived traits, and user expectations. Specifically, the authors identified two distinct folk theories of how generative AI systems “acquire knowledge” (Li et al., 2026). The first conceptualizes AI as a static, pre-existing repository of information that receives updates only occasionally. The second views learning as an active, continuous process wherein the system dynamically adapts to real-time prompts to personalize individual outputs while simultaneously building a “collective intelligence” from global user inputs. Regarding the working mechanisms, participants, who were described as being more literate, explained that the AI-chatbot is “just predicting the best output they think you want,” and uses “temperature algorithms” to provide varying output (Li et al., 2026).

Across prior findings, four recurring concepts can be identified in individuals' descriptions of ChatGPT: anthropomorphic, deterministic, and probabilistic perspectives, which primarily concern response generation, and a dynamic perspective (as opposed to a static system), which concerns how the system evolves and changes. However, we identified two main research gaps within mental models of AI-chatbots among participants.

First, phenomena such as anthropomorphic tendencies have been researched quite extensively, but mostly only in relation to trust and user behavior, rather than in relation to which other mental concepts or schemas they occur alongside and how they shape a person's mental model as a whole. Mental models are composed of several components and must therefore be recognized as complex and potentially fragmented configurations of different concepts and schemas held by individuals (Jones et al., 2014). It is therefore necessary to determine individuals' various concepts, the extent to which they prevail, and the combinations of other concepts with which they co-occur. This can be achieved by eliciting people's explanations and verbalized ideas, in order to make mental models more apparent (Jones et al., 2014). For instance, a user who attributes human characteristics, such as thinking, to a chatbot may also tend to view it as a dynamic system, or conversely, use the distinction between dynamism and stability to differentiate between living organisms (i.e., dynamic beings) and a non-living machine. Insights into these co-occurring concepts within a mental model are necessary to be able to predict decisions and actions of users.

Second, it remains unclear whether the use of technically appropriate AI terminology indicates a functional understanding of generative AI. Prior research has often inferred technical competence from isolated statements, such as references to probabilistic word prediction (Jin et al., 2025; Li et al., 2026). However, a person may use such terminology while simultaneously assuming that ChatGPT searches a database of pre-formulated answers, learns directly from every interaction, or can independently recognize whether an answer is true. These contradictions become visible only when multiple statements from the same individual are examined together.

One manifestation is “chauffeur knowledge,” historically inspired by Max Planck's chauffeur's ability to reproduce scientific lectures without understanding them (Dobelli, 2010). While we refer to the concept of chauffeur knowledge only when a person uses a technical term but cannot correctly explain the mechanism to which that same term refers, interviewed participants may also hold a fragmented understanding, where they use LLM-specific jargon for one mechanism, but have a misconception about another fundamental LLM functionality. In the context of widespread public exposure to AI terminology (Abdel-Hafez and Yousef, 2025), a person reproducing such vocabulary may create an appearance of AI competence that is not matched by functional understanding. This challenges approaches to AI literacy that use familiarity with technical terms as a proxy for users' ability to interpret system behavior and limitations (Jin et al., 2025). Demonstrated in an educational intervention by Sivenas, students who were deliberately exposed to hallucinated outputs subsequently reported restricting their use of ChatGPT to topics about which they already possessed sufficient knowledge to evaluate its responses (2025). At the same time, most students initially interpreted a general warning that ChatGPT could provide incorrect responses mainly as referring to errors in mathematical calculations (Sivenas, 2025). This finding suggests that the awareness of a warning alone or technical term does not necessarily entail an adequate understanding of the scope and practical implications of generative AI failures.

While the aforementioned studies provide important insights, our current understanding of users' theories on AI-chatbots reflects group-level patterns across topics and only manages to capture the variety of perspectives on isolated aspects between individuals, instead of exploring the variety of perspectives within a single person. While focus group discussions are effective for capturing collective narratives, they are optimized for consensus-building and therefore highly susceptible to social desirability bias (Bergen and Labonté, 2020). The group dynamic tends to smooth over internal cognitive mismatches within or between individuals, incompletely formed assumptions and doubts. Naive or intuitive ideas may be suppressed to avoid perceived ignorance once a more plausible explanation emerges within the group. Additionally, those with higher apparent or objective competence or expertise may take up more speaking time and influence the positions of others (Nikoleizig et al., 2021).

To address these limitations, our study focuses on the individual level, acknowledging that individuals may simultaneously hold multiple partially incompatible representations that activate depending on the context. By operationalizing mental models as potentially fragmented structures (DeVito, 2021), we investigate the following research questions:

RQ1: To what extent are the mental models about ChatGPT's functionality distinct across technical laypersons and how do deterministic, probabilistic, anthropomorphic and dynamic concepts manifest within these mental models?

RQ2: How do misconceptions of ChatGPT's functional mechanisms manifest across different mental models, and in what ways is technical terminology supported or contradicted within participants' functional explanations, through same-mechanism mismatches (chauffeur knowledge) or through correct understanding of one mechanism alongside misconceptions about another (fragmented understanding)?

Through semi-structured individual interviews, we examine mental models as participant-specific configurations rather than inferring understanding from isolated concepts or statements. First, we analyze how deterministic, probabilistic, anthropomorphic, and dynamic concepts coexist within individuals' explanations of ChatGPT. Second, we examine mechanism-specific misconceptions, terminology-explanation mismatches, and fragmented understanding within individuals' mental models. This approach allows us to reveal difficult-to-grasp processes that often remain undetected in quantitative surveys and focus groups but are important for designing educational interventions against overreliance and misuse.

Participants

Participants in this study were eligible to participate if they were at least 18 years old and met our criteria for being technical laypersons. The latter criteria were operationalized by including individuals who had no formal education in computer science, no professional IT experience, and no prior coursework in artificial intelligence, but who nonetheless had a basic level of contextual awareness, either as active users of ChatGPT or at least knowing that it is an AI-chatbot. Participants were recruited via convenience sampling, with invitations distributed through university email lists, posts on social media platforms (such as LinkedIn and WhatsApp), and personal invitations within the researchers' immediate networks. During data collection, the interviewer monitored the recurrence and diversity of the explanations provided, and later interviews predominantly reproduced explanatory patterns encountered in earlier interviews. Recruitment concluded after 21 interviews, taking into account this increasing informational redundancy, the breadth and depth of the interview protocol, and the resources available for conducting and analyzing the interviews. As systematic coding was performed after completion of data collection, formal saturation was not used as a prospective stopping criterion. The interviews were conducted in person between November 27 and December 19, 2024. At that time, the ChatGPT product could provide access to different underlying models and system-level features depending on account tier, usage limits, settings, rollout status, and the date of interaction. Free-tier users could access GPT-4o subject to usage limits and GPT-4o mini, whereas paid users could have access to additional models and higher usage limits. Participants may also have had prior experience with earlier versions of ChatGPT, including GPT-3.5. However, participants' account tiers, the specific models they had used, and their exposure to or use of search and memory were not systematically assessed. To investigate our central research question regarding specific misconceptions of technical laypersons about ChatGPT, we conducted qualitative interviews using a semi-structured 1-on-1 format. This approach allowed participants to articulate their own understanding without being constrained by predefined categories. Most interviews were conducted in German; only two were conducted in English. Prior to all interviews, participants were asked to fill out a short questionnaire regarding their demographics, their self-rated prior knowledge about and attitude toward AI as well as their usage frequency of AI-chatbots. The complete questionnaire is available at OSF1. As an incentive for participating in the study, participants were offered either compensation of €20 or, as an alternative for psychology students, a participation credit equivalent to 60 mins.

A total of 21 participants (6 male, 15 female) participated in the study, with ages ranging from 20 to 57 years (M = 27.6, SD = 9.8). More than half of participants (57%) were students and nearly half of all participants worked or studied in the field of social services (see Table 1 for a demographic overview). Self-reported prior knowledge of AI was predominantly moderate to low, with 10 participants reporting intermediate knowledge and a further 10 indicating low or very low knowledge levels, while one participant reported a rather high knowledge level. Daily AI-chatbot usage was only reported by one participant, with most participants instead engaging with the technology on a weekly (n = 8) or monthly (n = 7) basis and four participants using it less than once a month. One had never used an AI-chatbot. Their attitude toward AI was mostly described as “neutral” (n = 10) or rather positive (n = 8), two reported a very positive attitude, and one expressed a rather negative attitude.

The primary field of education or profession, self-assessed knowledge, and attitude toward AI and frequency of AI-chatbot use are summarized in Table 1.

Materials

As the interviews were conducted using a semi-structured approach with open-ended questions, an interview guide was developed by the researchers themselves. For its creation, we used both theoretical considerations and empirical findings regarding AI (Antonenko and Abramowitz, 2023; Kim et al., 2023) and, more specifically, LLMs (Amaratunga, 2023; Lermann Henestrosa and Kimmerle, 2024).

In addition, we integrated questions derived from the Blog, Help Center, and FAQ pages provided at the OpenAI website2, particularly those about ChatGPT's functionality, data sources, and processing. The interview guide was further reviewed by colleagues from other psychology research groups at our university to ensure a clear and comprehensive design and to reduce the impact of initial questions on participants' answers to subsequent questions as much as possible.

The interview always started with initial open-ended questions about the chatbot's core architecture, focusing on how it works (e.g., “ How would you describe how a system like ChatGPT works ?,” “ How do you think ChatGPT answers your questions? ”) and what it is composed of (“ Do you have any idea how a system like ChatGPT is developed?,” “ What do you think is needed for the development of ChatGPT?,” “ Do you have any idea where the information that ChatGPT outputs as answers comes from ?”). Follow-up questions were adapted based on participants' responses to explore specific concepts or misconceptions in greater depth (e.g., by adding “ What makes you think so? ”). Additionally, to assess whether participants actually understood the technical terms, if they used any, or whether they were just repeating popular buzzwords, follow-up questions involving secondary, counterfactual, and boundary-testing questions were posed. For example, if a participant explained that ChatGPT predicts text using probabilistic patterns, we challenged that claim with concrete or contradictory analogies, by asking whether the system relies on pre-stored answers like an encyclopedia, or if it has the functionality to recognize its own mistakes and self-correct.

With this approach, we targeted a bottom-up questioning strategy to elicit participants' authentic mental models rather than simply rate knowledge superficially. The complete interview guide is openly available in English and German on OSF (see text footnote 1).

Data analysis

Data analysis comprised three primary phases, followed by a focused recoding of technical-vocabulary statements and associated functional explanations. In the first phase, two researchers independently reviewed the full interview transcripts of all participants to extract all statements containing beliefs, assumptions, or ideas regarding the functional mechanisms of ChatGPT. This phase served as a data reduction step to establish the unit of analysis. The researchers aggregated relevant text segments into a separate file organized by participant. Then, they compared the extracted statements, back-checked them against the whole transcript and discussed misalignments until reaching agreement on which statements contained a hypothesis, idea, description, or explanation of how ChatGPT works. Where the meaning of a sentence or phrase was unclear from the quotation alone, context was added in brackets.

Following statement extraction, the second phase focused on the formal coding of these statements, which were in most cases full sentences or parts of longer sentences. For this, a third researcher joined the team to conduct a deductive qualitative content analysis (Mayring, 2015) of the extracted statements. The underlying coding scheme comprised four categories—the Deterministic concept (DET), Probabilistic concept (PROB), Anthropomorphic concept (ANTH), and Dynamic concept (DYN)—which we had previously identified as recurring themes in prior work (Bewersdorff et al., 2023; Brauner et al., 2023; Horstmann et al., 2023; Li et al., 2026) and subsequently refined and defined conceptually. To operationalize this scheme according to Mayring's protocols, the team developed a formal coding agenda specifying definitions, explicit inclusion and exclusion criteria, and prototypical anchor examples for each category (see Supplementary Table S3).

The four mental concept categories were operationalized within the agenda as follows:

(i) The Deterministic concept (DET) identified the view of ChatGPT as a static retrieval system or “library of facts,” in which coding focused on instances where participants described a 1-to-1 retrieval process or compared the system to an encyclopedia or search engine.

(ii) The Probabilistic concept (PROB) captured the understanding that the system constructs responses based on statistical likelihoods and “next-word” prediction, rather than retrieving pre-stored answers.

(iii) The Anthropomorphic concept (ANTH) was applied when participants attributed human-like traits, such as consciousness, intent, or emotion, to the system. It distinguishes between purely figurative language and a substantive belief in the AI's “understanding” or “desire” to help.

(iv) The fourth code was the Dynamic concept (DYN), which captured statements regarding the system's evolution, irrespective of whether the proposed mechanism was technically accurate. The code identifies, for instance, the belief that ChatGPT is a self-correcting or adaptive entity that fundamentally updates its global knowledge base through individual user interactions.

To evaluate the comprehensiveness of the agenda with the predefined categories, we executed a formative validation loop on the first five interview transcripts, which were coded independently by all three researchers. The definitions, coding rules, and inclusion and exclusion criteria were then refined through an iterative process to solve all discrepancies. Following this adjustment, the full set of transcripts was coded independently, with coders blinded to one another's classifications by using the structured scheme to systematically map verbalized concepts (Mayring and Fenzl, 2019). Coding was completed before the annotated documents were exchanged. Based on these contemporaneous coding records, exact percentage agreement for the assignment of the DET, PROB, ANTH, and DYN categories was 93.99%, with discrepant assignments identified in 11 of 181 assignments. These discrepancies were subsequently discussed and resolved through joint meetings until intersubjective consensus was reached among all three researchers, resulting in a reconciled dataset of 170 codes. In difficult or ambiguous cases, the coders first documented their respective classifications and justifications for their interpretations independently of one another in the annotated coding documents. During the joint meetings held to reach coding consensus, competing interpretations were discussed with reference to the full statement and the surrounding interview context, as well as to the category definitions specified in the coding agenda. Consensus was defined as agreement among all three researchers following this comparison, not by majority vote. If a statement lacked sufficiently precise information for clear assignment to a code and the coders could not reach consensus, it was excluded. In the third phase, the qualitative coding data were quantified to explore potential patterns in participants' mental models. For each participant, code frequencies for the four concepts (DET, PROB, ANTH, and DYN) were converted into relative proportions. This transformation controlled for variations in transcript length and total verbal output, representing each participant's conceptual framework as a relative percentage profile totaling 100%. In the Supplementary Table S2, we provide a table with the total number of codes per participant for transparency. These frequencies represent the concepts expressed across the complete semi-structured interview, including responses to standardized prompts and adaptive follow-up questions. Therefore, they should be interpreted as exploratory patterns and not as context-independent measurements of the mental models. Following the calculation of the relative code proportions, the participant-level profiles were initially inspected in ascending participant-number order, without group labels or visual separators. Provisional qualitative configurations were identified by comparing the relative prominence and co-occurrence of the four mental-model concepts across participants. After this visual inspection, the profiles in Figure 1 were reordered according to these qualitatively identified configurations, and numerically within each configuration, solely to improve their visual interpretability. The subsequent k -means analysis was conducted separately as a secondary exploratory analysis and was not used to construct or order Figure 1.

To complement the descriptive profile analysis, an exploratory k -means clustering procedure was conducted as an illustrative, secondary visualization to inspect potential data-driven groupings based on mental model compositions (see Supplementary material S3 for detailed methods).

Additionally, to investigate whether technical terminology was supported by corresponding functional understanding, we conducted a focused recoding of the statements and corresponding explanations of the 10 participants who used such terminology. Each technical term was linked to the participant's explanation of the same functional mechanism.

As part of this analysis, three members of the author team independently classified the relevant extracted statements as either expressing a misconception or not. Statements were classified as misconceptions when they expressed a substantively inaccurate understanding of ChatGPT's functionality or limitations. Percentage agreement among the three coders, calculated from their independent classifications before consensus was established, was 93.55%. For discrepant or difficult cases, the coders revisited the complete statement and its surrounding interview context and evaluated its technical accuracy against technical publications and official OpenAI product announcements, documentation, and support pages applicable during the data-collection period. Divergent interpretations were then discussed in joint meetings until agreement among all three coders was reached. Consensus was not determined by majority vote. Statements for which an inaccurate belief could not be established after considering the interview context and relevant sources were not classified as misconceptions. The accuracy assessment distinguished between the inference-time mechanism of the underlying language model and product-level functions of ChatGPT. For statements concerning response generation, the reference point was the probabilistic token-generation mechanism of the underlying LLM. References to web search, retrieval, memory, or personalization were not classified as misconceptions solely because they described functions beyond the base model. Web-search statements were classified as misconceptions only when participants treated live web retrieval as the default mechanism for every response or as ChatGPT's general mechanism of response generation. References to cross-conversation memory were likewise not classified as misconceptions solely on that basis because memory was available in some product configurations during the data-collection period. Claims concerning learning were classified as misconceptions only when participants described individual interactions as immediately and automatically modifying the system's globally available model knowledge. Statements that model parameters remain unchanged during an individual interaction were not classified as misconceptions, whereas categorical claims that ChatGPT is never updated were evaluated against developer-mediated model and product updates. The reconciled statement-level classifications were subsequently used in the mechanism-specific comparison of technical terms and corresponding explanations. A participant was coded as demonstrating chauffeur knowledge (1 = present, 0 = not identified) when at least one explanation was incompatible with the minimal functional meaning of the technical term referring to the same mechanism. Fragmented understanding was coded separately (1 = present, 0 = not identified) when a participant described one technical concept at least partially correctly and used technical terminology but expressed a misconception about another fundamental mechanism. The two classifications were not mutually exclusive. Ambiguous statements for which the interview context did not establish an inaccurate belief were not coded.

For descriptive reporting, a participant-level misconception indicator was coded as 1 when at least one statement had been classified as a misconception and as 0 when no such statement was identified. This indicator represented only the presence or absence of at least one misconception and not their number, severity, or coherence, and it was not included in the exploratory clustering analysis.

Visual inspection of the participant-level profiles in their original, unsorted order suggested three broad qualitative configurations based on the relative prominence and co-occurrence of the four mental-model concepts. Figure 1 presents the same proportional profiles reordered according to these visually identified configurations to facilitate comparison. Specifically, participants appeared to group around profiles characterized by differing emphases on deterministic, anthropomorphic, probabilistic, and dynamic conceptions of ChatGPT.

Configuration 1: Predominantly probabilistic thinking patterns (n = 8)

Configuration 1 comprises eight participants (P01, P06, P09, P15, P17, P18, P19, and P21). Compared to the others, this group aligns closest with the actual statistical architecture of LLMs, showing high proportions of PROB codes and low on the “DET” concept. These individuals differ from those in the other configurations in that statements describing ChatGPT as a tool operating on large datasets, identifying text patterns, and evaluating word probabilities predominated over descriptions of a static registry of answers.

The use of probabilistic vocabulary, however, does not prevent misunderstandings regarding ChatGPT's mechanisms. When asked for concrete explanations, participants in this configuration also resort to descriptions typical of search engines. Half of the group mentioned a background database (P06: “He queries his database and whatever comes back there, he basically spits it out at me”), which they suspect ChatGPT uses to verify facts, assess the accuracy of results, or protect answers from user corrections (P21: [when asked what happens when you point out an error that ChatGPT made] ”[...] that it just says the person or that input is invalid because my data says the opposite.”).

The “DYN” mental concept, including ideas about the chatbot's adaptability, is very inconsistently interpreted by participants in this configuration. Some assume global, real-time learning from individual input, such as P17, who imagines this “knowledge base” to be corrected when enough users point out a potential mistake: “ I believe if very, very many people were to write that this [an output by ChatGPT] is wrong, then it would indeed learn,” while others believe the system is completely static and unaffected by user feedback.

The contradiction between probabilistic terminology and deterministic or anthropomorphic assumptions is evident in quotes like “ I think I know that it encodes word-for-word in some way... it has something to do with word probabilities...” (P18). Two participants (P01 and P15) hold a predominantly probabilistic mental model without anthropomorphic or deterministic concepts, while others tend to slip into anthropomorphic explanations, which is reflected in statements such as P09: “ But sometimes, I mean when you really start to sort of argue with it... it lets itself be persuaded. That is, it changes its mind […]”; and P19: “No, no [there is no register of pre-stored answers or topics]. I think that it [ChatGPT] actively looks at your inquiry and then it tries to figure out what to do with that.”

Participants in configuration 1 work or study primarily in science, economics, and health. This group contains the sample's highest concentration of daily or near-daily users. Despite using more technically accurate vocabulary than other participants, their self-rated prior knowledge was mostly “rather low” or “average” and only one reported a “rather high” baseline knowledge. When asked where they got their knowledge about ChatGPT from, answers reflected self-directed usage experience (“just trying it out”), touched upon it in university lectures, and digital media platforms such as LinkedIn, YouTube and Instagram, but also through informal social discussions with colleagues, partners, or peers with specialized technical backgrounds, such as friends studying IT.

Configuration 2: Anthropomorphic and dynamic thinking patterns (n = 4)

Configuration 2 is the smallest group (P05, P07, P14, and P16) and shows the highest structural divergence by having higher proportions on “ANTH” and ”DYN” with no “PROB” codes assigned.

Participants in this configuration expressed no probabilistic concepts or vocabulary and relied instead on a combination of live internet scraping and human-like cognitive processing. Participants in this configuration believe that the system “searches” for information on the internet, “picks it out,” and then “synthesizes” it into an answer. The anthropomorphic component is particularly prevalent among the individuals in this configuration. Although they correctly associate the system with “networks,” this shifts into a direct analogy with the human brain, which “thinks” or processes information like a human being (P05: “ Google basically gives you a thick folder... [ChatGPT] basically gives you an answer that it thinks is correct.”). The lack of word-prediction mechanisms is also shown in the assumptions that ChatGPT requires a higher mental load for less frequently asked questions (P07: “ Questions that have been asked very often are probably easier for ChatGPT to answer [...] ”), or that ChatGPT distinguishes between “easy” and “difficult” topics (P16: “ […] Probably not in my usage, because I don't have such difficult topics that I would ‘teach' ChatGPT.”).

For instance, P14 explained how answers were generated as follows: “ […] it probably goes through its entire system and looks to see what information is available on the topic. And then it just responds based on what's most common—that is, how best to answer it—or whatever information is most readily available.” Here, “the most common” information reflects the concept of a “register” or “database” where ready-made answers or factual building blocks are stored, rather than understanding this as probabilistic word prediction. They also assume that ChatGPT learns directly from corrections, such as giving feedback by typing “That's wrong” as a response to an output, and that it immediately stores this knowledge for future use or even for other users. As a result, they believe that the system becomes “smarter” with every interaction, much like a human learns through experience.

The participants in configuration 2 are educated or working in social services and healthcare. Their usage is moderate to low: one participant had never used the system, while the other three interacted with it several times per week. Self-rated prior knowledge was uniformly low. Their technical understanding comes from informal word-of-mouth, social environments, or general news media.

Configuration 3: Predominantly deterministic retrieval-based thinking patterns (n = 9)

As the largest segment of the sample, configuration 3 includes nine participants (P02, P03, P04, P08, P10, P11, P12, P13, and P20). The group is defined by the highest proportion of deterministic codes and, compared to configuration 2, lower proportions of the anthropomorphic codes.

Among the individuals in this configuration, the belief in a deterministic system predominates and their mental model follows an “Input → Search in database → Output” logic. P02 and P12 follow a clear “data pool” and “register/lexicon” model, expecting that the answer already exists “ready-made” somewhere. The individuals in this configuration with the “DYN” component (P03, P04) believe that the model learns through their questions and that they become part of this “database,” which is a departure from the completely static deterministic model.

Some participants recognize that the system's outputs are not strictly deterministic, identifying an element of chance or statistical variability in how answers are generated (P11: “ I don't think it [the output for the same input] is exactly the same. I think it's partly down to chance.”). This understanding, however, remains flawed, as they believe that the system actively calculates a “truth” or “consensus” regarding content from different sources, rather than simply generating the next probable token. Less dominant than in configuration 2 (anthropomorphic-dynamic), but still present in a few participants of configuration 3, are the anthropomorphic beliefs that ChatGPT is capable of correcting errors through mental or cognitive reflection. Even though many participants in this group understand that ChatGPT “learns” and improves with feedback or input, there are misunderstandings about how this works technically, and learning is often confused with simply storing new facts in a database (P04: “ I think that basically every piece of my input [into ChatGPT] is stored as information and can then be reused by others.”). They also share the belief that information is pre-existing and static, with the AI acting as a sophisticated filter (P11: “… the topics are already there; it simply filters out whatever you're currently asking.”). They perceive the generation of answers as a patchwork process of aggregation (P20: “ I think a lot of this stuff gets copied and pasted from the internet, so the answers might be cobbled together from different articles.”).

Configuration 3 is demographically varied. Four participants work in social services, with others distributed across health, tourism, economy, media, and production. Usage frequency spans a wide range, from less than once per month to several times per week. Self-rated knowledge about AI ranged from low to average: two participants reported “very low” four “rather low” and three “average” knowledge about AI. The participants' prior knowledge about ChatGPT in this configuration stemmed mostly from friends, colleagues, and word of mouth, but also from university and “trying it out.”

Exploratory clustering analysis

As a separate secondary analysis, k -means clustering was subsequently applied to the proportional coding profiles to examine whether a data-driven partition produced a broadly comparable grouping. This analysis was exploratory and illustrative and did not determine the qualitative configurations or the ordering presented in Figure 1. The solution had a Bayesian Information Criterion (BIC), of 70.20, an R 2 of 0.579 and an average Silhouette coefficient of 0.350. Because the four proportional features constituted compositional profiles, the sample was small, and the cluster separation was modest, the resulting groupings were not interpreted as discrete latent participant types. The complete analytical settings, model-selection results, and clustering visualizations are provided in the Supplementary material S3.

Chauffeur knowledge and fragmented understanding

At least one functional misconception was identified for every participant. Consequently, the presence of misconceptions did not distinguish participants who used technical terminology from those who did not. We therefore examined how technical terminology related to participants' functional explanations. Among the 10 participants who used technical terminology, we distinguished two non-exclusive patterns. Chauffeur knowledge was identified when a technical term was accompanied by an incompatible explanation of the mechanism denoted by that term. Fragmented understanding was identified when an at least partially correct account of one mechanism coexisted with a misconception about a separate mechanism. Seven participants met the criterion for chauffeur knowledge, seven met the criterion for fragmented understanding, and four (P12, P13, P15, and P21) met both criteria (Table 2).

Same-mechanism mismatches were particularly apparent when probabilistic generation was described using a retrieval model. P15 stated that ChatGPT works based on probabilities and “knows which word typically follows the next one,” but explained it does so by “actively ‘searching' through the data it was fed.” Similarly, P18 referred to word-pair probabilities and variable outputs but explained that ChatGPT identifies these combinations by looking them up in an exhaustive list, database, or on the internet. P17 likewise connected probabilistic generation and associative networks to the claim that ChatGPT searches the internet and then returns the most likely answer, which was classified as a mismatch because live retrieval was presented as the general by-default response-generation mechanism rather than as an optional product feature. P12 similarly described responses as “statistically most likely” combinations of word forms while also portraying the same process as the retrieval of pre-formulated answers from an encyclopedia-like register. Another example of chauffeur knowledge was provided by P13, who initially stated “the answers are always the same” but, when confronted with the fact that repeated prompts can produce different outputs, attributed this to ChatGPT seeking “to provide a bit of variance.”

Fragmented understanding was evident when an appropriate explanation of one mechanism coexisted with a misconception about another. P01 described probabilistic word co-occurrence while assuming that new information supplied by an individual user would immediately be learned by ChatGPT. P09 and P19 provided partially correct accounts of pattern recognition or word prediction but also attributed human-like intention, opinions, or understanding to the system. In addition to P12 's same-mechanism mismatch concerning response generation, P12 also assumed that the system permanently learns from each verbal confirmation provided by a user, representing a separate misconception concerning model updating. P13 provided a correct account of training, stating that numerous texts had been “fed into the data set so that ChatGPT could essentially learn how to write texts,” but separately assumed that, when an error was identified, the system “recalculates its own response and compares it with new sources.” P15 's partially correct account of next-word prediction likewise coexisted with a categorical rejection of any post-deployment updating or adaptation. P21 showed an analogous combination of same- and separate-mechanism misconceptions, as detailed in Table 2.

The DYN code was descriptive rather than evaluative and captured statements portraying the system as changing over time, regardless of whether the proposed mechanism was technically accurate. Some of these statements nevertheless reflected misconceptions about adaptation through individual interactions, as illustrated by P01 and P12 above.

This study investigated how technical laypersons conceptualize the functional mechanisms of ChatGPT. Specifically, we examined the configurations of deterministic, probabilistic, anthropomorphic, and dynamic concepts in participants' mental models and distinguished whether technical vocabulary was contradicted within the mechanism it named (chauffeur knowledge) or accompanied by a misconception about another mechanism despite partially correct technical understanding (fragmented understanding). The qualitative analysis revealed substantial heterogeneity in participants' proportional coding profiles. Deterministic and probabilistic concepts were most prominent, whereas anthropomorphic and dynamic concepts occurred less frequently and, in some participants, appeared in combination. These patterns are broadly consistent with conceptual distinctions described in prior literature (Bewersdorff et al., 2023; Brauner et al., 2023; Li et al., 2026; Schneller et al., 2026). An exploratory but purely illustrative k-means cluster analysis provided a three-group summary of similarities among the observed profiles, which we included in the Supplementary material. However, given the small sample size, the compositional structure of the proportional variables, and the modest cluster separation, these groupings should not be interpreted as discrete latent participant types or as evidence of a stable population-level cluster structure.

The focused recoding qualifies our original interpretation of technical-vocabulary use. Because at least one misconception was identified for every participant, the data do not demonstrate an association between jargon use and the occurrence of misconceptions. Rather, the analysis shows two analytically distinct ways in which technical terminology can fail to indicate comprehensive functional understanding. In seven cases, the explanation of the same mechanism was incompatible with the technical term, providing direct evidence of chauffeur knowledge. In seven cases, participants correctly or partially correctly described one mechanism but held a misconception about another, indicating fragmented understanding, and four participants displayed both patterns. Thus, technical-vocabulary users should not be treated as a homogeneous category, nor do the findings imply that they are more prone to misconceptions than participants who did not use such terminology. As observed previously, the interviewed participants often use buzzwords that are widely circulated in the media, such as “probability algorithm,” “large language model,” and “pattern recognition” (Brauner et al., 2023). According to Cruz and Lombrozo (2025), jargon in explanations can reduce laypersons' sensitivity to gaps in them and lead to explanatory satisfaction despite not increasing actual understanding. The authors observed that unfamiliar technical terms in scientific explanations were treated by non-expert individuals as if they filled missing mechanistic links, even when the explanation remains circular or otherwise flawed. The cases in which misconceptions were based on the same mechanism as the one the participants described using technical jargon suggest that technical terms can serve as explanatory placeholders when one explains probabilistic response generation by explaining retrieval-based mechanisms of indexed information. Cases of participants with fragmented understanding, on the other hand, show that the use of AI-specific jargon regarding one component does not necessarily transfer to a fundamental understanding of other components, such as knowledge about model updating or the lack of fact verification or agency. These findings suggest that, when assessing competence in the field of generative AI, one should rely more on mechanism-based explanations and questions regarding distinctions rather than exclusively on terminology. Our findings extend Cruz and Lombrozo's conclusion by suggesting that jargon may not only influence the evaluation of received explanations, but also the active construction of explanations by non-experts. Technical jargon may fill perceived gaps when users construct mental models of generative AI without improving their functional understanding. A resulting sense of sufficient knowledge could make users less receptive to corrective information about generative AI outputs. Unlike technical terms from many other scientific domains, AI-related jargon has become highly visible in public discourse (Abdel-Hafez and Yousef, 2025).

Implications for AI literacy assessment

Current frameworks defining AI literacy, such as the one provided by Ng et al. (2021) decomposing AI literacy into the categories “Know & Understand” “Use & Apply” “Evaluate & Create,” and “AI Ethics,” often structure competency hierarchically. In practice, however, AI literacy tests often assess the basic “Know & Understand” through superficial and very specific multiple-choice questions (e.g., asking about specific LLM training-data sources) (Ng et al., 2021) that can be answered correctly without understanding what LLM training even entails. Higher-order assessments, such as those evaluating knowledge of “temperature settings” in LLMs (Jin et al., 2025), are also only assessing technical jargon, while presuming a baseline of understanding that our data suggest is often absent. As self-reports and objective-based assessments correlate poorly (Chiu et al., 2024; Zhang et al., 2026), current literacy tests likely fail to capture misconceptions that may lead to overreliance or misuse. A user may correctly describe probabilistic generation or training while still assuming a separate, automatic truth-verification process; such fragmented understanding would remain undetected by vocabulary-focused items.

Retrieval as the dominant metaphor

Across all identified configurations participants frequently used search-engine or encyclopedia metaphors to explain ChatGPT's response generation. Users often perceive ChatGPT and Google Search as functionally similar, despite attributing higher information quality to the former (Xu et al., 2023). This assumption was reflected in participant statements claiming the systems function similarly “behind the scenes.” OpenAI mirrors this in a recent report, where information retrieval remains a primary purpose for ChatGPT use (Chatterji et al., 2025). The retrieval metaphor is not inherently inaccurate at the level of the deployed ChatGPT product, which may incorporate web-search functionality. However, it becomes a misconception when users assume that every answer is automatically based on a live web retrieval, that complete answers are selected from a stored database, or that the retrieval itself explains the probabilistic generation of the final answer. In our coding, we therefore took into account the differences between references to potentially available search functions and statements that portrayed automatic retrieval from the internet as ChatGPT's standard and general mechanism for generating responses. For instance, users may interpret longer response times of an LLM as a sign of intensive database searching rather than compute-bound resource saturation, mostly occurring during peak hours (Recasens et al., 2025). Furthermore, believing in a retrieval-based system obscures awareness of hallucinations, which is amplified by evidence that long, default explanations increase user confidence regardless of factual accuracy (Steyvers et al., 2025).

Anthropomorphism and dynamicity as heuristic conflation

The co-occurrence of anthropomorphic and dynamic mental concepts suggests that participants struggle to conceptualize systems that evolve without human-like intent (Waytz et al., 2014). We interpret this conflation of dynamic and anthropomorphic mental concepts as a cognitive heuristic, since, historically, learning and adaptive systems have typically been encountered almost exclusively in the form of living organisms, and conversational agents were largely limited to other humans before the advent of generative AI. Participants treat the AI-chatbot as an organism that learns and adapts in real time, rather than a system bounded by discrete training cycles (OpenAI, 2026b,c). While prior work identified users' folk theories of LLM learning as “collective intelligence” (Li et al., 2026), our findings extend that by demonstrating individuals' overestimation of their personal agency, often believing that written feedback (e.g., “This is false”) causes the chatbot to delete errors from or adjust its “database” immediately. This anthropomorphizing is further evidenced by the belief that AI possesses subjective experience (Colombatto and Fleming, 2024) and the tendency to perceive AI as more rational in decision-making than humans (Vijayakumar et al., 2026). While Colombatto and Fleming (2024) found a positive correlation between usage frequency and the attribution of agency to LLMs, participants in our study, who differed in reported usage frequency, either rejected analogies to humans entirely or imagined that ChatGPT undergoes human thought processes, confirming earlier findings regarding public polarization over the similarity of problem-solving mechanisms between humans and AI (Bewersdorff et al., 2023).

Consequences for trust, reliance, and risk

Probably very few laypeople understand how a phone call works from a technical standpoint, and yet we use our smartphones every day without any problems. In contrast, AI-mediated interaction involves delegating cognitive tasks to a system that simulates intelligence without actually possessing understanding or an independent capacity for reflection and fact verification (Shanahan, 2024). If users fail to understand these limitations due to the stochastic nature of LLMs, they will be more likely to regard the plausible-sounding responses as a reliable source of facts and evaluate them less critically. Flawed mental models lead to systematic misinterpretations of system failures. When confronted with errors, laypeople often interpret them as an indication of intentional agency (Lee et al., 2025a; Rapp et al., 2025). We found that individuals with a mental model in which the deterministic concept is predominant typically attribute errors to “incorrect data” within the system, whereas those with a predominantly probabilistic model may use the term “hallucination” without fully understanding the underlying mechanism. This is relevant since hallucinations can arise from prompt-sensitive as well as model-intrinsic factors (Anh-Hoang et al., 2025). A user who treats all erroneous output as a prompting problem may repeatedly reformulate a request while continuing to rely on an intrinsically unreliable response, whereas a user who understands the system's broader limitations may be more likely to seek external verification. Further evidence indicates that students who encountered hallucinated output subsequently reported restricting their use to domains in which they possessed sufficient prior knowledge to evaluate the response (Sivenas, 2025). Furthermore, using ChatGPT for work tasks can increase user confidence while simultaneously reducing active information processing, perceived control, and sense of responsibility (Lee et al., 2025b; Kares et al., 2026).

Practical implications

The results of our study suggest that interventions aimed at addressing flawed mental models of ChatGPT's functionality, which are associated with over-trust and overreliance, must go beyond simply teaching technical vocabulary. For example, educational materials that focus on scenario-based explanations could directly confront users with a discrepancy between the chatbot's behavior and their mental model, thereby helping to dispel misconceptions (Smith et al., 1994).

User education could, for example, ask users to compare responses generated from the same input multiple times before explaining the underlying mechanisms, rather than primarily teaching definitions of technical terms such as an LLM's “temperature.” However, we argue that merely making such information available is insufficient if the tool's user interface—the context in which users are expected to internalize that information—systematically undermines these educational efforts. The homepage at chatgpt.com is designed in such a way that it immediately prompts a user to start a conversation with a default question and an additional “ask anything” hint in the input field. The note “ChatGPT Can Make Mistakes. Consider Checking Important Information” only appears in a subtle gray font after the first question has already been posed, in small print at the very bottom of the screen. The interface provides no indication that, although users can theoretically “ask anything,” some questions are unsuitable or cannot be answered correctly, reliably. However, simply expanding such disclosures may not be sufficient. Recent research suggests that disclaimers concerning ChatGPT's limitations do not substantially affect users' trust or reliance on the system (Metzger et al., 2024), regardless of how they are framed (Lermann Henestrosa and Kimmerle, 2025). In contrast, characteristics of the chatbot's communication style appear to have a stronger influence on trust judgments, with authoritative and confident responses increasing perceived credibility (Metzger et al., 2024). Consequently, design interventions should focus more on calibrating the system's communication style to accurately reflect its probabilistic and fallible nature. Responses could, for example, communicate uncertainty where appropriate, encourage independent verification of consequential information, and avoid unwarranted expressions of confidence when evidence is limited. Extending to data privacy, where OpenAI's public documentation specifies that system updates and training from user feedback do not occur automatically or constantly, and that users have the right to opt out of data collection (OpenAI, 2026a–c), our interview data shows a total absence of individuals' awareness regarding this option. This suggests that epistemic transparency should be communicated through interaction design and conversational behavior rather than delegating information to separate support pages. Additionally, drawing on an example from the financial sector, where banks and brokers are often required by regulatory guidelines to verify whether clients have sufficient knowledge and experience before they can trade certain products (Directive 2014/65/EU), a brief assessment of users' understanding of AI system limitations and risks could mitigate uncritical reliance. Future research should also investigate how different chatbot communication styles shape users' conceptions of AI systems and whether these effects contribute to the formation, reinforcement, or correction of users' mental models of ChatGPT.

Limitations and outlook

This work has several limitations. First, generalizability and statistical tests to identify differences among participants dependent on their knowledge sources, field of work or education, age, etc. cannot be derived, given the qualitative nature of this study and our small sample of 21 participants. The sample was not designed to be representative and small absolute differences are analytically fragile. The descriptive values for “chauffeur knowledge” and “fragmented understanding” are also not independent of one another, as the two codes can occur together. They should therefore not be interpreted as prevalence estimates or as evidence of a correlation between technical terminology and misconceptions. Rather than drawing statistical conclusions about relations at the population level, our goal was to identify patterns of meaning, understand reasoning structures and reconstruct conceptual models. Further research would need larger samples to identify the potential influence of demographic characteristics, knowledge sources or usage frequencies on mental models expressed by participants in this sample. The extent of usage and the variety of tasks that participants use ChatGPT for might also have an influence.

Second, potential prompt biases may exist within the interview protocol. To mitigate this risk, we initiated the interviews with entirely open-ended questions, placing explicit prompts about whether ChatGPT possesses a “register” or “consciousness” later in the session. General descriptions of how ChatGPT works emerged primarily in response to the initial open-ended questions, whereas statements concerning specific functionalities or system responses were generally elicited through explicit probes. Nevertheless, follow-up questions were adapted to participants' preceding responses, and participants may therefore not have received equivalent opportunities to produce statements in every concept category. Because spontaneous and probe-elicited statements were not separately coded, the participant-level frequencies cannot be interpreted solely as indicators of the spontaneous salience or strength of particular mental-model concepts. Consequently, the proportional profiles and exploratory clustering may partly capture differences in interviewer probing and response opportunities alongside differences in participants' mental models.

Third, our methodology assesses only how individuals verbally express their beliefs and assumptions regarding ChatGPT's functionality. Actual behavioral interactions between individuals and the system were not observed. Relying exclusively on retrospective verbalization limits the validity of the captured mental models, as stated beliefs do not always align with situated interaction behavior. Capturing the real-time adjustments users make during system failures requires different methods. Future studies may integrate user-tool interaction logs or think-aloud tasks during active ChatGPT sessions to observe how mental models function in practice.

Fourth, the interviews were conducted between November 27 and December 19, 2024, when the available ChatGPT product could differ according to account tier, settings, feature rollout, and date of access. We did not systematically assess participants' account tiers, the specific models they had used, or their exposure to web search and cross-conversation memory. Consequently, the exact product configuration informing each participant's mental model cannot be reconstructed. We addressed this issue analytically by distinguishing claims about the underlying LLM's response-generation mechanism from descriptions of product-level functionality. Nevertheless, some ambiguous statements may have reflected individual experiences with configurations that were not documented. Moreover, because ChatGPT has continued to evolve since data collection, the identified misconceptions and the practical recommendations derived from them are temporally situated and should be reassessed as product capabilities and user experiences change.

Finally, 19 of the 21 interviews were translated from German into English for analysis, since one of the three coders is not proficient in German, which introduces the risk of losing subtle linguistic nuance. To minimize this distortion, two researchers from the team systematically compared two distinct versions of the transcripts: the original German text and the automatically translated English output generated by our transcription software. Despite this comparative verification, translation boundaries may still have flattened specific metaphorical expressions used by participants.

Conclusion

We analyzed holistic, participant-specific profiles and showed that individual mental models about ChatGPT's functioning and underlying architecture were fragmented and often internally contradictory. Participants often conflated correct and incorrect assumptions about various aspects of ChatGPT, suggesting a fragmented understanding of how it works. Appropriate technical terminology can be accompanied either by an incompatible explanation of the same mechanism or by a misconception about a different mechanism despite partially correct understanding elsewhere. Furthermore, there seems to remain a persistent human tendency to attribute intent, opinions, and attitudes to ChatGPT, indicating that our cognitive architecture for social interaction is deeply ingrained and may considerably resist attempts to regard an LLM with a conversational, chat-style interface as a purely technological tool. Moving forward, information and education on conversational AI should focus on functional mechanisms, especially the differences between probabilistic generation, information retrieval, and model updating, rather than relying on terminology that can create a false sense of understanding.