“It Can’t Be That Simple”: exploring Finnish non-experts’ conceptions of AI and whether these predict emotional orientation towards AI

The difficulty of rigorously defining “AI” is reflected in the variety of expert definitions found in literature (Table 1). Bearman and Ajjawi (2024) suggest categorizing different conceptualisations and definitions of AI in technical definitions (“What AI is”), capability definitions (“What AI does”), and relational definitions (“How AI works within a social system”). Computer scientists tend to define AI by its technical functionality (LeCun et al. 2015; Russell and Norvig 2016) or capability (McCarthy et al. 2006; Rai et al. 2019), while the public's definition of AI tends to compare it with human behaviour and intelligence (Zhang 2024). When investigating conceptions of non-experts, it may be sensible to focus, besides the technical and capability aspects, also on the relational aspects (i.e., how humans interact with these technologies), as suggested by Bearman and Ajjawi (2023).

Several studies have found that people are confused and have misunderstandings on how AI works at the technical level (Bewersdorff et al. 2023). These include, for example, inability to distinguish technical terms (Antonenko and Abramowitz 2023; Teng et al. 2022; Lindner and Berges 2020), and a view that AI algorithms cannot make mistakes (Antonenko and Abramowitz 2023). Misunderstandings of AI s technical foundations are often related to the use of metaphors that oversimplify complex processes. Colburn and Shute (2008) argue that metaphors related to computer science help manage complexity through abstraction, but this abstraction can also obscure the true nature of systems. For example, Murray-Rust et al. (2022) highlight how the metaphor of “learning” in machine learning can lead to misunderstandings: while models do adapt based on new data, equating this process to human learning implies cognitive abilities like reasoning and generalisation and that “fitting” might be a better metaphor to capture the mechanical nature of the process. Brauner et al. (2023) conclude that AI is still a “black box” for many and inability to assess the risks and possibilities of AI can lead to irrational control beliefs of AI. On the other hand, Bearman and Ajjawi (2023) point out that uncertainty of the technical details of AI does not imply that AI could not be used effectively.

AI ethicists are concerned that portraying AI as human-like may give people a misleading understanding of its risks and benefits (Cave et al. 2018). The tendency to anthropomorphize non-human entities, i.e. attribute human-like characteristics or attributes to them is well documented (Bewersdorff et al. 2023; Epley et al. 2007; Blut et al. 2021). This natural tendency arises from our cognitive predisposition to interpret behaviour in terms of intentions and emotions, even when interacting with inanimate or technological systems (Boyer 1996). For example, Scott et al. (2023) showed how non-experts perceive some level of consciousness in voice assistants, chatbots, and robot vacuum cleaners. Furthermore, recent evidence suggests that people's moral evaluations of AI agents are shaped by multiple human-like features, including emotion expression, cooperation, and moral judgment (Ladak et al. 2024). In a Finnish context, studies with children have shown that conceptions of AI are often anthropomorphic, without references to the role of data in training AI (Mertala et al. 2022). Another study further explored children s misconceptions of AI, showing that Finnish 5th and 6th graders typically framed AI either as human cognition, as a human-like entity, or as a machine with pre-installed intelligence; however, since most children assessed their AI knowledge as poor, their misconceptions appear superficial rather than profound. Mertala and Fagerlund (2024).

The introduction of large language models (LLMs) (Floridi and Chiriatti 2020) and machines’ increasing ability to produce coherent language about a variety of topics has added a new flavor to the discussion on what AI is. Bender et al. (2021) used the metaphor of a “stochastic parrot” to describe a language model and warned about the dangers should these models be used at scale. Whether language models can “understand” is an ongoing debate among AI researchers, splitting the views according to a survey (Michael et al. 2023). Mitchell and colleagues have suggested that the capabilities of language models should be considered “new modes of understanding” or an “alien form of intuition” (Mitchell and Krakauer 2023), accepting a distinction from human understanding. In this, the discussion around LLMs is approaching, terminologically, the distinction made in 1980 by Searle (1980) between “weak” and “strong” AI. Weak AI acts as a technological tool under the control of humans, whereas strong AI “can literally be said to understand and have other cognitive states”(Searle 1980).

Understanding what AI is and how it works shapes how we trust, use, and relate to these tools and technologies in our daily lives. Misconceptions about AI can lead to unrealistic expectations or misplaced trust. As AI becomes more integrated into society, it is crucial that people have a clear, accurate view of what these systems can and cannot do.

As Braidotti (2019, p. 50) aptly describes, “Moments or periods of euphoria at the astonishing technological advances [...] alternate with moments or periods of anxiety in view of exceedingly high prize we [...] are paying for these transformations.” In our responses to rapid technological progress, excitement over innovation is tempered by growing concerns about its societal and ethical costs. When Hick and Ziefle (2022) explored public perceptions towards AI in qualitative interviews, two main themes emerged: a dystopian view where AI could lead to destructive outcomes or loss of human control (akin to movies like The Terminator) and an exaggerated or utopian attitude about the possibilities and abilities of AI as a tool that could transform society for the better (Hick and Ziefle 2022). This tension is reflected in the growing body of research focused on public beliefs, attitudes, and emotions towards AI and the factors that shape them (Schepman and Rodway 2023; Neudert et al. 2020; Kelley et al. 2021; Araujo et al. 2020; Stein et al. 2024).

Familiarity with AI, algorithms, or computer programming often leads to more positive views regarding AI technologies, particularly in contexts like automated decision-making (Araujo et al. 2020). Similarly, Mantello et al. (2023) found that individuals familiar with AI, and those who perceived tangible benefits, were less likely to feel threatened by it. Moreover, Li and Zheng (2024) found that social media use predicted more positive attitudes towards AI indirectly through increased perceived AI fairness and reduced perceived AI threat. On the other hand, deeply held beliefs can intensify negative emotional reactions towards AI. Kozak and Fel (2024) found that religious individuals, who may see AI as challenging the divine control or human uniqueness, are more likely to experience intense emotions, such as fear and anger. Furthermore, Stein et al. (2024) found that susceptibility to conspiracy beliefs connects to a more negative attitude. Moreover, Kelley et al. (2021) identified four key public sentiments towards AI: exciting, useful, worrying, and futuristic—with notable differences across countries: developing nations expressed more excitement, while developed nations exhibited greater concern.

Another construct that describes our affective relationship with AI is trust. People form varying levels of trust in machines (Mcknight et al. 2011), and as Jacovi et al. (2021) highlight, too much trust in AI can lead to over-reliance while too little trust can result in resistance, even in cases where AI offers benefits. Bao et al. (2022) found five public segments based on perceptions of AI s risks and benefits: the negative, who saw risks outweighing benefits; the ambivalent, who perceived both as high; the tepid, who leaned slightly towards benefits; the ambiguous, who saw moderate levels of both; and the indifferent, who perceived low levels of both risks and benefits. Pataranutaporn et al. (2023) found that framing AI systems as empathetic or caring increased perceptions of trustworthiness and led to more positive emotional responses. In contrast, the absence of trust may generate “algorithm anxiety”: Jhaver et al. (2018) interviewed Airbnb hosts who reported significant anxiety caused by a perceived lack of control and uncertainty over how algorithmic evaluation works on the platform. Trust in humans versus machines also varies by context. Mou and Xu (2017) found users more willing to disclose personal information to humans than AI. However, Ta et al. (2020) discovered that a social companion chatbot creates a “safe space” where users feel comfortable discussing any topic without fear of judgment. Similarly, Sundar and Kim (2019) found that users were more likely to trust a machine agent with their credit card information than a human agent. Zhang et al. (2023) expanded on this theme by exploring how trust is affected by the perceived identity of “AI teammates”. In their study, participants who were deceived into thinking their AI teammate was human trusted and accepted their teammate's decisions less often than those who knew their teammate was AI.

Emotional orientations towards AI matter because they strongly influence how people engage with, adopt, or resist new technologies. Prior work shows that affective reactions such as trust, fear, or excitement often have a substantial effect on their behaviour (Bao et al. 2022; Pataranutaporn et al. 2023). By examining not only what people think AI is but also how they feel about it, our study contributes to a more complete understanding of public orientations towards AI and their implications for societal adoption.

Insufficient understanding of AI deepens the digital divide (Carter et al. 2020) in terms of access, skills, and outcomes (Lutz 2019), and also limits public engagement in AI governance (Curtis et al. 2023). As AI development continues to advance, it is crucial for different societal stakeholders to understand the diverse conceptions, attitudes, beliefs, and emotions people have towards these technologies. While existing research has provided valuable insights into public perceptions and emotional orientations towards AI, more subtle understanding is needed regarding how non-experts understand and feel about AI—and whether and how specific emotional orientations follow specific conceptions. In the present study, we focus on AI conceptions, emotional orientations to AI, and the relationship of these among non-experts in Finland (Fig. 2).

Bild vergrößern

First, we investigate AI conceptions. While it is known that expert definitions of AI differ from those of non-experts (Zhang 2024; Salles et al. 2020; Krafft et al. 2020), a more detailed picture is needed. We use Bearman and Ajjawi’s (2024) model (technical, capability, relational) to analyse Finnish non-experts’ written descriptions of AI. Our first research question is:Using abductive analysis, we recognize different features in written descriptions of what AI is and how it works. By addressing this question, we aim to uncover how people make sense of AI across different dimensions. Understanding these conceptions can inform the design of AI literacy efforts and improve how to communicate about AI with the public.

Second, we examine emotional orientations towards AI. Prior research (Hick and Ziefle 2022; Neudert et al. 2020; Kelley et al. 2021) has highlighted the presence of both highly positive and highly negative emotional responses to AI. However, the picture may be more nuanced – e.g., individuals may experience positive and negative feelings simultaneously (Liu et al. 2024), or neither. To explore this complexity, we adopt a person-oriented approach, similar to Bao et al. (2022) in their analysis of perceived risks and benefits. Our second research question is:Using multiple affect-related measures and clustering, we identify distinct emotional orientation profiles. Recognizing these patterns can deepen our understanding of emotional diversity and support the development of tailored AI literacy efforts that align with different emotional orientations.

Third, we investigate the relationship between AI conceptions and emotional orientations. Previous research has shown, that increased knowledge about AI fosters trust and more positive affective relations towards AI (Araujo et al. 2020; Mantello et al. 2023; Li and Zheng 2024), while deeply held personal beliefs, lack of control, and uncertainty on how algorithms work may cause anxiety (Jhaver et al. 2018; Kozak and Fel 2024; Stein et al. 2024). Based on this, we hypothesise that individuals with conceptions more near the expert definitions of AI show more positive emotional orientation towards AI. Our third research questions is:By answering this question, we aim to better understand the cognitive affective interplay in public perceptions of AI and provide groundwork for future research in this area. Overall, the present study aims to provide increased detail in analyzing non-experts’ understanding of AI and feelings towards AI, and explore the dynamics between these. A more detailed understanding of public’s relations with AI is needed for developing interventions to improve AI literacy, for ensuring that diverse public voices are included in decisions about how AI governed, and in designing meaningful human–AI interactions.

Participants signed up for a general audience online lecture series on artificial intelligence offered by a non-profit organisation responsible for organising education for the general public. The lecture series was given by one of the authors. Before the lectures, the participants were asked to complete an online survey about their understanding of and attitudes towards AI. Lectures and survey were conducted in Swedish, one of the two official languages in Finland.

This type of study does not require ethical review per the guidelines of the relevant national board of research ethics, or the guidelines of the host university: the participants were adults making an informed decision to volunteer for the study, and the study does not involve physical intervention, especially strong stimuli, risk of causing mental harm that exceeds the limits of normal daily life, or a threat to the safety of the participants, their families or the researchers. The participants also had the opportunity to take part in the survey, thereby providing information to the lecturer, while still opting out of the study.

A total of 150 participants completed the survey, with 141 giving informed consent for their data to be used for research purposes. Most respondents (83.7%) identified as female, 14.9% as male, and 1.4 % preferred not to disclose their gender. The largest age group was 50–59 years (36.2% of respondents) followed by 40–49 years (29.8% of respondents), 30–39 years (15.6% of respondents) and over 60 years (15.6% of respondents).

As AI is a rather new topic for non-experts, we deemed that asking for participants’ education level would not provide us with much information on their AI knowledge and skills. Given the introductory nature of the lecture series mentioned in marketing and the reach of the organization offering the series, it is feasible to assume that the participants were non-experts/novices in relation to AI. To get some self-assessed information of the participants, we asked them which of the five roles on Rogers’ innovation curve (Rogers 1962; Rogers et al. 2008) best describes how they view themselves in relation to digitalisation and technological development. For each of the five roles, we also wrote a plain text explanation of how to interpret the role:The results showed that only 5.7% viewed themselves as innovators and 19.1% as early adopters. Most (53.9%) identified as early majority, 20.6% as late majority, and 0.7% as laggards.

Participants responded to an online survey prior to attending the lectures. The survey consisted of background variables and questions related to conceptions of AI and emotional orientation towards AI. Participants were asked to report their gender, age group, and self-assessed role on Rogers’ innovation curve. Open response items regarding conceptions of AI (e.g., ‘What do you think AI is? and ‘How do you think AI works?’) were adapted from the study by Mertala et al. (2022). Regarding emotional orientation, we employed several items that were answered on a Likert scale from 1 (‘Completely disagree’) to 5 (‘Completely agree’). In line with work by Kelley et al. (2021), we employed items on excitement (‘AI is exciting’), worrying (‘AI is frightening’) and personal relevance (‘AI is not for me’). Additionally, we employed one item on trust towards AI versus humans (‘I trust humans more than AI’).

To understand the non-experts’ conceptions of AI and its functionality, an abductive thematic analysis was conducted on responses to two open-ended questions (What do you think AI is? and How do you think AI works?) as described in section 2.3.1. In order to shed light on how the technical, relational and capability features were manifested in the public's responses, we illustrate the codes by providing excerpts from the open-ended answers. Most responses manifested multiple codes, either representing one or multiple definition types (technical, relational or capability) (see Table 4). Table 5 presents a selection of quotes, with the parts related to a certain code highlighted in the text. The quotes were selected to illustrate the broad diversity found in the responses, with some being mainly technical or relational in their nature, and others being more mixed.

On the level of definition types, relational features describing AI as a part of a social system were mentioned by 82% of the participants, features describing AI’s technical operations by 79% of the participants, and capability features describing what AI can do by 55% of the participants. Additionally, 37% of participants mentioned uncertainty about what AI is or how it works. There was a statistically significant positive correlation between relational and capability-based features (\(\tau _{{\text{b}}} = 0.36\),\(p<0.001\)), indicating significant co-occurrence of these categories in non-experts’ definitions. Furthermore, there was a statistically significant negative correlation between technical features and uncertainty (\(\tau _{{\text{b}}} = - 0.21\),\(p=0.01\)), indicating that non-experts’ definitions including technical features were less likely to include mentions of uncertainty. Interestingly, correlations between capability and technical features as well as between relational and technical features were near zero.

On the level of coded features, the most prevalent features were Autonomous (72% of participants), Data-driven (66%), and Anthropomorphic (53%), followed by Human-controlled (40%), Rule-based ( 31%), Self-learning (20%), and Probabilistic (9%). To investigate co-occurrence of the features, pairwise correlations using Kendall’s \(\tau _{{\text{b}}}\) were employed. Anthropomorphic and Autonomous showed a very high correlation (\(\tau _{{\text{b}}} = 0.54\),\(p=0.04\)). High positive correlations were also found between Self-learning and Autonomous (\(\tau _{{\text{b}}} = 0.54\),\(p<0.001\)), as well as Self-learning and Anthropomorphic (\(\tau _{{\text{b}}} = 0.37\),\(p<0.001\)), indicating large co-occurrence of these features. Less strong but still statistically significant correlations were found between Human-controlled and Rule-based (\(\tau _{{\text{b}}} = 0.20\),\(p=0.02\)), Autonomous and Data-driven (\(\tau _{{\text{b}}} = 0.18\),\(p=0.03\)), and Self-learning and Data-driven (\(\tau _{{\text{b}}} = 0.17\),\(p=0.04\)). All other pairwise correlations were nonsignificant (see Table 6).

To explore the distinct emotional orientation profiles towards AI among non-experts, a hierarchical clustering analysis was conducted using the four affective dimensions of Exciting, Frightening, Not for me, and Trust humans more than AI. The goal of this analysis was to identify meaningful subgroups who share similar patterns of responses across the emotional orientation variables.

The analysis ended up in a four-cluster solution (\(R^2=0.43\), silhouette score 0.53, Calinski-Harabasz 34.3), with the four emotional orientation profiles (Fig. 4) named to depict the variable profile as follows:

Bild vergrößern

In sum, the clustering analysis revealed four distinct emotional orientation profiles towards AI among non-experts. These groups illustrate the diversity of affective stances, ranging from enthusiastic engagement to wary openness, and from disengagement to indifference.

Logistic regression analyses examined whether certain features in non-experts conceptions of AI (RQ1) predicted their emotional orientation profile (RQ2). The models provided little evidence that non-experts conceptions of AI would reliably predict their emotional orientation towards AI.

The model with Optimists cluster membership as a dependent variable showed modest fit and above-chance accuracy (McFadden \(R^2=0.09\), balanced accuracy = 0.58), suggesting that those viewing AI as anthropomorphic (\(Est.=1.31, SE=0.47, p<0.01\)) and probabilistic (\(Est.=1.47, SE=0.64, p=0.02\)) were more likely to belong to the Optimists profile.

The model with Cautious cluster membership as a dependent variable showed weak fit and accuracy on par with chance (McFadden \(R^2=0.05\), balanced accuracy = 0.50), and revealed no predictors reaching conventional significance.

For the Disengaged and Indifferent profiles, no predictors were retained in the final models, indicating that the measured features did not significantly differentiate these groups from others (see Table 7).

We have identified several limitations and methodological considerations that should be taken into account while interpreting our results.

First, due to the open nature of the lecture series, we used convenience sampling, that is, participation was based on self-selection, where individuals chose to register for the lecture series and thus became study participants. This approach limits generalizability as it is likely to attract those with a pre-existing interest in the topic. Our sample is predominantly female, from a single Nordic country, and largely middle-aged or older. Consequently, the results should not be assumed to apply to other populations. Nevertheless, we argue that this focus on a non-expert population offers a valuable point of comparison to the large body of research centred on other groups.

Second, while the amount of data analyzed in this study in total is non-trivial, it remained a limiting factor. Due to demographic imbalance, we were not able to make credible analysis based on age and gender. Moreover, the sample size leaves open the possibility of Type II errors, and some marginal trends, particularly the potential role of anthropomorphic and probabilistic conceptions of AI as predictors of emotional orientation, warrant further investigation in future research.

Third, coding in RQ1 was particularly challenging for the Anthropomorphic feature. In our initial discussions, we agreed that a code along the lines of ‘excessive anthropomorphism’ would have been interesting, and it appeared somewhat intuitive to all of us. While a certain degree of anthropomorphism arises from the scientific and technical terminology used in relation to AI systems (e.g., “intelligence,” “learning,”), there is clearly a level that extends beyond this baseline (e.g., “wants”). However, in practice, distinguishing between the technical and non-technical terms was difficult in the terse responses: many words such as “knows” or “understands” can be used in a technical context (e.g., “knowledge base,” “natural language understanding”), but they could also have been used in a highly anthropomorphic fashion, with little to distinguish between the two in the plain text of the participants’ responses. Despite the difficulties, the interrater reliability (Fleiss’ Kappa = 0.43) may be considered moderate according to Landis and Koch (1977).

Fourth, trust in AI was measured in relative terms, comparing it to trust in human performance for the same task. This framing means that our results reflect perceptions of both AI and human trustworthiness, and cannot be interpreted as absolute trust in AI. A participant s high relative trust in AI could stem either from strong confidence in AI or from low confidence in humans (or both). Future research could combine both relative and absolute measures of trust to provide a more comprehensive understanding of how people evaluate AI systems.

Our findings suggest that people approach AI with different baseline mental models of how the technology functions. These different kinds of mental models have contrasting consequences for both AI literacy and public discourse. A deterministic, rule-based and human-controlled view of AI may risk downplaying concerns about bias, transparency, and unpredictability by creating an illusion of human control. However, it may be rather easily corrected through targeted teaching about data-driven learning, model training, and the probabilistic nature of AI in general. In contrast, anthropomorphic conceptions may fuel either dystopian fears of machines taking over or utopian hype about AI as a superhuman partner. In AI literacy initiatives, these kinds of conceptions could be addressed by focusing besides the technical aspects of AI, also on the relational aspects, such as the human agency as creators and users of AI systems, and providers of training data.

Regarding the emotional orientations, our four profiles show that public orientations toward AI are more granular than a simple utopia dystopia split. AI literacy interventions should optimally recognise baseline emotional differences as different profiles have different kinds of needs for the intervention. Optimists benefit from critical reflection on limits and risks of AI to avoid over-reliance and hype, whereas people with disengaged or indifferent emotional orientation first need relevance and motivation. People with cautious orientation balance interest and concern, which may be seen as an emotional level target of an AI literacy intervention. Furthermore, a mix of participants with varying emotional profiles may provide a chance for fruitful debates that broaden perspectives. Regarding societal participation in AI policies, the disengaged and indifferent groups may be the most challenging to reach, which risks public debate being dominated by the most enthusiastic and the most anxious voices. This dynamic can reduce perceived legitimacy and inclusiveness of policy processes and increase polarisation.

While we found little evidence that non-experts conceptions of AI reliably predict their emotional orientations, weak models could be formed for the Optimists and Cautious, whereas the Disengaged and Indifferent yielded only null models. This might suggest that for the latter groups, emotional orientations are shaped less by specific conceptions and more by broader factors such as prior experiences, media exposure, or sociocultural narratives. In practice, this means that interventions aimed at correcting misconceptions may influence the already engaged, but reaching disengaged or indifferent individuals likely requires strategies that highlight personal relevance. At the societal level, this underscores that cultivating broad-based engagement with AI cannot rest solely on correcting misconceptions, but must also address the motivational and relational conditions that enable even the indifferent or disengaged to see themselves as stakeholders in technological futures.