Explanations are pervasive in our everyday lives. On account of their ubiquity, explaining processes are a significant research topic, even more so after the technological progress of machine-learned artificial intelligence (AI; Rohlfing et al., 2021). The development of machine-learned AI systems is rapid. AI is increasingly used in different societal contexts, such as education (Zhai et al., 2021), medicine (Liu et al., 2021), law (Lai et al., 2024), or finance (Cao, 2022). Machine-learned AIs are opaque to laypeople as well as experts. Their increasing use in high-stakes domains requires them to be explainable. Thus, the call for explainable AI (XAI), which aims to make AI systems accessible, understandable, and criticizable for humans, is requested by many.

Researchers have provided different advanced theoretical frameworks to make AI more accessible (Miller, 2019). For instance, Ehsan and Riedl (2020) argue that XAI should be more human-centered (HCXAI). In contrast to other theoretical frameworks, HCXAI places significant emphasis on the interaction between AI and the social and contextual factors that influence it. Similarly, Sokol and Flach (2020) posit that more interactive explanations have the potential to increase the transparency of AI systems. Further, they argue that one-directional explanations often fail to meet the diverse needs of users and suggest that interactive explanations can address this issue.

Based on these observations, Rohlfing et al. (2021) point out several limitations in recent research on XAI and explanation and propose the theoretical framework of co-constructivity that addresses these shortcomings. In short, co-constructivity is characterized by a bidirectional and iterative explanation process in which the addressee is actively constructing the explanation in collaboration with the explainer. The agents negotiate both the goal and what is needed to know during the explanation process. In general, the co-constructive framework is characterized by a high agency of both agents involved. Following Rohlfing et al. (2021, 2025), we argue that a more co-constructive approach in human-computer interaction would improve the understanding of AI systems and the quality of the explanation processes. Interestingly, though, most existing approaches to XAI consist of unilateral statements by experts or systems in which the addressees have at most options of personalization, but neither the possibility of actively participating in the explanation nor in identifying the explanandum (Miller, 2019). The tendency of system developers and researchers to develop explanations one-sidedly from the perspective of the explainer seems to be pervasive, and interactive, social, or co-constructive approaches are only slowly and incompletely gaining acceptance.

In the present paper we ask which assumptions about explaining can be found in cultural thinking. We critically scrutinized core assumptions about the explaining process through metaphors in language. It is possible to draw conclusions about cognition through language because metaphors reflect “powerful cultural interpretation patterns” (Schmitt, 2024, p. 220). This consideration is based on the idea that language, culture, and mind are interdependent (Sinha, 2017). This means that we can – to a certain extent – understand conceptualizations of explaining through the examination of figurative language about explaining. Considering the significance of explanations and the use of AI in educational, medical, and judicial contexts (among others), we argue that it is crucial to ensure that the language about explanations is apt. Language patterns, such as metaphors, can contribute to misunderstandings of scientific content (Harrison and Treagust, 2006; Taylor and Drewsbury, 2018) as well as activities and practices (e.g., Hager and Hodkinson, 2009; Paulson and Armstrong, 2011). Explanations or explaining might be one example.

Because metaphors serve as potent representations of cultural and cognitive patterns, they offer a unique perspective into how we conceptualize explanation processes. Metaphors also make it possible to identify the limitations of language and use them as a basis for technological progress. In the present study we will answer the following questions: Which metaphors are commonly used to conceptualize the explaining process? How are the explanation process and the corresponding agents depicted through metaphor? With agency being a central component of co-constructivity, we more specifically asked how the agency of the agents is conceptualized. Overall, we thus study how common explaining metaphors support or impede the co-constructive approach to explaining.

For this purpose, we conducted an empirical study in which the participants produced metaphors for the process of explanation. These metaphors were analyzed with a structured method to elaborate the roles and the agency implied in the metaphors. In the following, we first introduce the notion of metaphor in the context of cognitive linguistics. We then differentiate the necessary schematic structures of metaphor and present studies on cognitive mechanisms and effects of metaphor to show how and to what extent metaphors can influence people’s thoughts. Lastly, we explain the method we used for examining agency, namely transitivity analysis.

300 German metaphor texts and 263 English metaphor texts were collected online in 2022 and 2023 via the service provider Prolific. The participants were at least 18 years old and were native speakers of German or English, respectively.

In accordance with the method Elicited Metaphor Analysis (Low, 2015), the participants were explicitly asked to produce a metaphorical expression about explaining in response to the following prompt:

In order to create a basis for further metaphor interpretation, a standardized method of metaphor identification of the elicited metaphor texts had to be applied. We decided to use the metaphor identification method developed by Steen et al. (2010) called MIPVU. In this method, the meaning of every single lexical unit is compared to the basic meaning found in dictionaries. As recommended by Steen et al. (2010), we used dictionaries for identifying the basic meaning. For English texts, this was the Macmillan English Dictionary for Advanced Learners (Rundell, 2007). The Oxford Dictionary (Oxford University Press, n.d.) was used as a supplement. Comparable dictionaries of the German language were used; the Digitales Wörterbuch der Deutschen Sprache (Berlin-Brandenburgische Akademie der Wissenschaften, n.d.) and as an addition the Duden (Dudenredaktion, n.d.).

Among the metaphors, we analyzed only those related to the target domain explaining. Since our focus was on agency of the explaining process and the usage of transitivity analysis, only verbs and nominalized verbs were identified and coded. For each word, we compared whether the basic meaning corresponded to the meaning of the units in the metaphor texts. If this was not the case, the word was identified as a metaphor.

The following example illustrates the subsequent process of analysis. Based on the basic meanings of the Macmillan (Rundell, 2007), the verbs in italics were identified as metaphorical.

Frames were first reconstructed by grouping the similar meanings of the lexicons used. Because the Macmillan definitions of the terms adjusting (“To change something slightly in order to make it better, more accurate, or more effective”), shaping (“To form something into a particular shape”) and tailoring (“To make or change something especially for a particular person or purpose”) are quite similar, they were summarized as one frame that we called adjusting. Frames were included in the analysis if at least 3 metaphors allocated to them were identified in the 300/269 texts in our corpus.

Frames were then grouped into domains. The grouping of the domains was conducted using the Master Metaphor List (Lakoff et al., 1991), a collection of basic metaphorical concepts, as a reference. In addition to the frame adjusting, several other frames that focus on the modification of objects were identified, such as removing or putting together. These were reconstructed as the domain modification. Again, we decided to include only frames that were present in at least 3 of the texts. Domains were only derived if there were at least two frames which were assigned to the domain.

The frames were added to a coding manual which contains the domains, frames, examples of the metaphorical concepts, and definitions of the metaphorical expressions. With the help of the coding manual, we conducted further reviews of the texts to identify more metaphorical expressions, until no further frames and domains could be reconstructed, and no more metaphorical expressions could be found.

The software MAXQDA (for more information on the usage for qualitative research, see Kuckartz and Rädiker, 2019) was used for the coding of the metaphorical expressions. In each target domain, two researchers coded the metaphors separately. If a metaphor concept occurred several times in a metaphor text, only the first time was coded because we were interested in the frequency of metaphorical concepts across the dataset rather than the frequency of concepts within a text. The intercoder agreement was determined with the help of MAXQDA by calculating the code overlaps in the text.

Once all metaphorical expressions had been identified and the frames reconstructed and coded, they were analyzed on a conceptual and linguistic basis using the transitivity parameters mentioned earlier. The meanings given in the dictionaries were used as the basis for our analysis. The agency of the parameters was compared to the agency of the co-constructive approach. If the agency did not match the co-constructive approach, the parameter was assessed as negative. We provide an open-access corpus with annotated verb metaphors at the Open Science Foundation.² The coding manuals with the domains and frames of explaining are made available in the Supplementary materials.

In the following, we first present the metaphorical domains identified in the data. Secondly, we present the agency of the frames to answer the question of how agency is conceptualized and whether it fits a co-constructive understanding of explaining. Finally, we compare the English and the German metaphors. For the sake of brevity, we present the analysis of the English corpus in detail; the German results differ little.

For a full overview of the conceptual structures and the ratings of all parameters in English and German. The intercoder agreement is 86% for the German data and 91% for the English data, resulting in a Cohen’s κ of 0.85 in the German dataset and 0.9 in the English dataset. According to Landis and Koch (1977) this is rated as an almost perfect agreement. For the final analysis, all disagreements were remedied by one of the coders.

In view of the rapid development of AI and its increasing presence in diverse societal contexts, the importance of XAI and the enhancement of human-computer interaction when explaining AI is steadily rising. We argue that the implementation of social and co-constructive aspects proposed by Rohlfing et al. (2021, 2025) in AI systems is one important component of achieving these improvements.

The goal of the present research was to analyze explaining metaphors with the help of the co-constructive aspects of explaining to find out how the explanation process is conceptualized and whether the co-constructive approach to explaining is existent in the metaphors that are used for explaining. For this purpose, we evaluated the agency of explaining metaphors identified in an English and German dataset collected from native speakers and compared it to the co-constructive view on explaining. Specifically, we identified metaphorical frames of different domains and analyzed them using transitivity analysis, a structured method to analyze the degree of agency or effectiveness of the action in a verbalized event. In the present version of transitivity analysis, the presence and activeness of both explainer and addressee, the duration, iterativeness and bidirectionality of the action as well as the affectedness of the object were examined. Our analysis suggests that common explaining metaphors tend to limit a co-constructive understanding of explaining – in the English and in the German dataset. Their implications hinder a co-constructive understanding.

In more detail: The second participant is either only implicit or does not take an active role in the action (parameter participants). In the theoretical framework of Rohlfing et al. (2021), the addressee is necessarily co-constructing the explanation through collaborative actions and takes on an important role in the explanation process. Although the application to the notion of explaining is new, we are not the first to point out this mismatch. Reddy (1979) also draws attention to the passiveness of the addressee in the conduit metaphor. Also, the psycholinguist Clark (1996) argues that language (which is heavily involved in explaining) itself should be seen as a joint effort.

The analysis of the explaining process further supports the interpretation that explaining metaphors impede co-constructive aspects (parameter temporality). Most metaphorical actions are short-lived, unidirectional and non-iterative rather than iterative and bidirectional as in the co-constructive framework.

In most metaphors, the object remains unchanged by the metaphorical actions (parameter affectedness of the object). The most prevalent change of the objects is a change of location, which is typically realized by the domain transfer or motion. Further modifications are the illumination of objects or changing the viewing direction of the agent to ensure visibility. All of these, however, leave the object itself relatively unaffected. This suggests that the explanandum is typically treated as a predefined, rigid entity that is simply handed over rather than actively shaped in collaboration with the addressee. Again, there is a parallel in earlier discussions: Geeraerts (1993) critiques the conduit metaphor in a similar way, arguing that meaning should be understood as constructed through interaction rather than as a transfer of a fixed object.

An exception is the domain modification with frames such as opening, adjusting or adding. Objects are created from scratch or with the help of parts, they are connected, opened, certain elements are removed, they are adjusted or broken down. To a larger or lesser degree, these metaphors allow for the construction or even co-construction which is one – but only one – of the essential elements of Rohlfing et al.’s (2021) conceptualization of explaining.

In the context of XAI, the findings may explain why many current systems result in a predominantly top-down approach where explanations are “built” by experts or AI and “given” without engaging the addressee. It may also explain why progress in more interactive approaches is so slow, even though these have long been repeatedly demanded. The metaphorical conceptualizations we identified in the present research may be inadvertently built into systems by developers by means of their metaphorical conceptualizations of explaining. This could be all the more true given that the metaphors we have identified are so commonplace that they are probably not even perceived as specific ideas in most cases and therefore cannot be challenged. Metaphors are in fact widely used in technology (e.g., Wilson, 2024) – think of the desktop, surfing, the web, chatting, and the like, and often discussed as helpful (e.g., Kim and Maher, 2020). Researchers have, however, repeatedly pointed out that, although often fostering understanding, metaphors may lead to misconceptualizations that impair the use of technology and have suggested more appropriate metaphors (e.g., Gentner and Nielsen, 1996). Most of this research has addressed metaphors for elements of the technology itself, especially concerning human-computer interfaces. We take this approach a step further to XAI and the practices of explaining.

In order to align with the societal need for understanding, criticizing and co-constructing AI, XAI frameworks need to incorporate more human-centered dynamic, interactive, co-constructive elements that allow users to ask questions, provide feedback, and iteratively refine their understanding. As argued by Ehsan and Riedl (2020), Sokol and Flach (2020) and Rohlfing et al. (2021), this would support users as active participants in the explanation process. This would not only be beneficial to the explanation process but is also ethically urgent because it would empower users to critically and pro-actively engage with AI systems rather than passively receiving their inputs.

Yet, despite this need, we expect that contemporary AI systems are deficient regarding aspects of co-construction. This assumption is supported by research conducted by Lenke and Schulte (2025). In a workshop setting, the theoretical framework of co-construction was introduced. The interaction between ChatGPT and the participants was tested in a pre-post-test design, and the monitoring and scaffolding prompts were then compared. Following the workshop, the participants showed an enhancement in co-constructive prompts. Lenke and Schulte (2025) further posit that the occurrence of co-constructive interaction is not attributable to ChatGPT itself, but that the responsibility for causing such interaction lies with the addressee.

The relevance of our study is not limited to the XAI context. We have focused on it because the considerations of Ehsan and Riedl (2020), Sokol and Flach (2020) and especially the framework of Rohlfing et al. (2021) provided a very precise idea of claims and XAI with which we were able to compare the metaphors. But, of course, some of this can be transferred to educational contexts. Prevailing explaining metaphors may reinforce teacher-centered practices. Educators might adopt methods that prioritize delivering content rather than fostering active dialogue. Duru (2015) for instance has demonstrated that most teacher metaphors of teacher-training students reflect teacher-centered beliefs.

With its critical focus, the present analysis does, in a very specific, empirical way, what computer scientist Agre – to mention only a single researcher – aimed at in his Critical Technical Practice (Agre, 1997). We analyzed metaphors for highlighted and neglected aspects in the notion of explaining. The first goal is to become aware of these metaphors, as a precondition for attempts to adjust or change them and then finally improving explaining practices in XAI and the quality of human-computer interaction. One might take this approach much further to what scholars in, above all, Science and Technology Studies and feminist studies have done (e.g., Latour, 1987; Haraway, 1988; Harding, 1991). For their inventors, metaphors often seem apt and sometimes so self-evident that their metaphoricity is not noticed. However, they may be apt only for certain groups and in certain cultural contexts. Feminist researchers especially have argued that it is necessary to trace possible social and material effects of metaphors (e.g., Cowan and Rault, 2022; Haraway, 1991; Suchman, 2008). Many of them aim at reconfiguring “agencies at the human-machine interface” (Suchman, 2008, p. 150). We consider this to be an important future direction, but one that goes far beyond the current goal.

While our findings suggest that it is important to choose explaining metaphors carefully and analyze them for their potentially undesirable implications, it is important to acknowledge limitations that may impact the interpretation or generalizability of our results. Directly eliciting metaphors in the context of a survey is a standard method in metaphor research (Low, 2015). However, this is an artificial situation, and participants may have used metaphors and sentence structures that diverge from those that they would have chosen in a more everyday discourse. From the perspective of Deliberate Metaphor Theory (Steen, 2011, 2023), our elicited data likely reflect deliberate metaphors, whereas everyday discourse often relies on non-deliberate metaphors. Differences between elicited and everyday metaphor usage could result in an incomplete picture of how explaining is conceptualized in a real-world context. Future Work should therefore distinguish between spontaneously occurring metaphors and deliberately used ones (Reijnierse et al., 2018).

This research should also be supported by the additional use of corpora (Semino, 2008). For example, scientific texts, newspaper articles or educational books could be analyzed to ascertain whether the metaphorical patterns are consistent or if genre-specific contextual factors have any influence on the explaining metaphors. One study from our group strongly indicates that the same frames with the same low agency dominate scientific texts on explaining/XAI (Scharlau and Rohlfing, 2025).

It should further be noted that explaining metaphors could have only a small or even no framing effect on thinking about explaining. There are two meta-analyses that have compared the effects between non-metaphorical and metaphorical utterances. The results are fairly similar – the effect sizes are small and reliable (r = 0.07; Sopory and Dillard, 2002; r = 0.09; Van Stee, 2018). Flusberg et al. (2024) also point out that multiple factors, cognitive, social and pragmatic in nature, influence the metaphorical framing effect. Nevertheless, we would argue that because the metaphors we examined are very commonplace and frequent and because implications regarding agency are very similar, they may have a relevant influence on concepts of explaining, expectations of explaining, and on the actual explaining behavior.

Based on CMT (Lakoff and Johnson, 1980, 2003) and empirical studies on metaphorical framing (Flusberg et al., 2024; Thibodeau et al., 2017, 2019) it is common and reasonable to assume that explaining metaphors influence thoughts and attitudes about explaining. We have identified a lack of collaboration, constructiveness and bidirectionality in the dominant metaphors of explaining. It should now be investigated whether these metaphors actually lead to less co-constructive views of behaviors in explaining than alternative metaphors that contain all these elements. In the event that different metaphors of explaining do affect the perception of the explaining process differently, the choice of metaphors in explaining contexts, whether in XAI, education, or other domains, should be reconsidered to encourage a more co-constructive interaction.

In addition, future research could investigate whether the effects of agentive framing, as reported by Fausey and Boroditsky (2010), extend to the explanation process. Specifically, it should be analyzed whether the agency of the addressee is valued less if the addressee is not explicitly mentioned in the context of an explanation. If metaphors of explanations predominantly focus on the actions of the explainer, then the role and the agency of the addressee may be backgrounded and therefore reinforce a unidirectional transfer of knowledge.

The similarity of metaphorical patterns observed in both German and English suggests that these metaphors might be deeply embedded in the cognitive and cultural frameworks of explaining. However, this study is limited by its focus on only two closely related languages of western culture. Building on Grady’s proposal (Grady, 1997) of near-universal conceptual metaphors that emerge from bodily experiences, it is plausible that transfer, modification, perception and movement metaphors for explaining occur in other languages. Nevertheless, it seems important to investigate whether these patterns extend to diverse linguistic and cultural backgrounds. In the context of XAI, it is crucial to consider how these culturally embedded metaphors influence user expectations and attitudes toward AI generated explanations.

Finally, one can note that there are ways to ameliorate problematic implications or at least draw attention to them. One of them is the metaphor extension strategy (Landau et al., 2017). This strategy retains the metaphor but adds statements that soften its problematic aspects or make alternative descriptions more prominent. One prominent example is the fight metaphor for cancer that cancer patients often reject because it implies that they have not fought enough if the cancer cannot be stopped. An extension here would be to say that it is a fight with unequal means (Wackers and Plug, 2022). This could be transferred to explaining metaphors, especially those that are so common that they cannot be easily avoided. Explaining can, for example, be still metaphorically described as a process of giving, but it could be emphasized that the addressee is actively taking the explanandum and may return it if it does not match their understanding. Similarly, explaining could be described as the process of breaking something down, but the process should be described as a collaborative effort. As mentioned above, the domain modification seems to be the most appropriate domain within our data set. The most agentive metaphors within this domain are building and creating, though the metaphors only emphasize the constructive aspect. The reciprocity and the collaboration would have to be added through extension.

These metaphor extensions and new metaphors which might directly support a co-constructive understanding (think of improvising a piece of music together) may be used in explaining XAI, but also in the future construction of AI systems. This could include developing systems that involve a bidirectional, collaborative, constructive and human-centered conversation, rather than a mere transfer of information. Ultimately, rethinking the metaphors we use for explaining may foster a better communication in both human and AI driven contexts.