The increasing deployment of Automated Vehicles (AVs) necessitates efective Explainable Artificial Intelligence (XAI) models to enhance user trust through transparent decision-making. While these models employ diverse techniques, they often prioritize algorithmic interpretability over human cognitive and linguistic processes. This paper introduces PsyLingXAV, a framework that integrates psycholinguistic principles into AV explanations, leveraging the psychology of language to optimize communication with users. By grounding explanation design in cognitively and linguistically informed strategies, PsyLingXAV tailors outputs to the driving context, advancing AV explainability. This work paves the way for more intuitive, human-centered explanations in AVs, with future studies poised to validate its impact.
Automated Vehicles (AVs) ofer substantial potential to improve road safety, trafic eficiency [
AV decision-making relies on ML/DL models [9] that manage either the entire driving task [12, 13] or only specific components [14]. Unlike generic AI applications, AVs operate in dynamic, safety-critical environments where decisions must be rapid, reliable, and contextually appropriate [3, 12]. Current XAI methods, dominated by algorithmic processes such as SHAP [15], provide general algorithmic transparency but are insuficient for the unique demands of AVs. These methods often fail to address domain-specific challenges, such as interpreting sensor data under varying road conditions or balancing trade-ofs in real-time collision scenarios [ 16], which require tailored explanations for stakeholders. Thus, AVs necessitate application-specific explainability frameworks that deliver interpretable and trustworthy explanations suited to their complex, high-stakes context.
As stated, most studies on XAI are algorithm-centric, focusing on model interpretability rather than user comprehension [18]. While these approaches enhance technical transparency, they may not align with how humans naturally understand AI decisions or what they expect in order to trust AI systems [9, 19]. As a result, these approaches often fail to account for human cognitive [10] and ethical considerations [5]. This failure can lead to reduced user trust [20, 21], increased uncertainty in AV behavior [22, 23], and dificulty in attributing responsibility in critical scenarios [ 8, 6]. Without explanations that align with human reasoning and ethical expectations, users may find it challenging to predict or rely on AV actions, potentially leading to slower adoption [3], regulatory challenges [17], and increased skepticism about the safety and reliability of AVs [5]. Therefore, efective AV explanations require a human-centric approach to address the aforementioned concerns.
XAI research often relies on researchers’ subjective assumptions about “good”explanations rather than established human explanation mechanisms [17]. To address this gap in AV development, we introduce a psycholinguistics approach that explores how humans process language, providing a foundation for designing cognitively aligned explanations [24]. Figure 1 illustrates XAI’s interdisciplinary nature at the intersection of AI, Human-Computer Interaction, and Social Sciences (adapted from [17]). Our expanded framework highlights key social science subfields: Philosophy, Linguistics, Psychology, and Social Foundations, with psycholinguistics bridging cognitive processing and linguistic communication. While philosophy and social foundations address the “why”(transparency, trust) and “what”(ethical concerns) of AV explainability, they often overlook the “how”—structuring explanations to be linguistically accessible. This research concentrates exclusively on textual explanations despite AVs’ capacity for various communication formats. This choice leverages text’s precision for AV explainability, with potential for later audio conversion. Specifically, we focus on sentence structure, leaving other psycholinguistic aspects (word recognition, phonological processing, lexical access, discourse analysis) for future research. The ultimate goal is to develop theoretically “user-aligned”explanations that efectively bridge the gap between algorithmic outputs and human comprehension, ensuring that these explanations align with the needs, expectations, and values of AV users. This approach not only represents a foundational step but also contributes to advancing a fully human-centric paradigm for AV explanation design. To the best of our knowledge, this integration has not been previously explored, especially in AVs.
Studies emphasize that timing is critical for efective explanations in AVs. Proactive (pre-action) explanations—delivered before an AV executes an action—build driver trust, reduce anxiety, and lower mental workload more efectively than post-action explanations [ 25, 20, 26]. This feedforward approach helps drivers to respond appropriately, though its impact varies by age: younger and middle-aged drivers show lower trust when AVs seek permission pre-action, while older drivers favor it [26]. Post-action explanations, however, consistently undermine trust, especially for older and middle-aged groups, suggesting timing should align with driver demographics [26].
While pre-action explanations are generally preferred, their necessity depends on driving context or scenario. Research shows explanations are more critical in emergencies or near-crashes than routine driving, with demand shaped by both scenario and driver traits [22]. Regarding explanation types, “why-only” explanations (e.g., “The car slowed to avoid a pedestrian”) elicit the most positive emotions, and both “why-only” and “why+how” explanations (e.g., “The car slowed to avoid a pedestrian by braking”) boost acceptance over no explanation [25]. Yet, “why+how” types enhance safety most, despite raising cognitive load, while contrastive “why not” explanations (e.g., “The car didn’t swerve due to a truck nearby”) improve intelligibility and accountability in crises [23]. Trust also hinges on risk perception: simple explanations work best as risk rises, but at extreme levels, no explanation may sufice [ 21, 22]. A key trust factor is anthropomorphism—attributing human-like traits to AVs [27]. Speech-based explanations increase perceived anthropomorphism and trust compared to non-speech interfaces [28]. Figure 2 captures how timing, demand, and explanation type shape human factors in AV trust, drawing on the aforementioned user-centered research.
Sentence processing occurs in three main stages: conceptual, syntactic, and phonological. The conceptual stage focuses on forming the intended meaning of a sentence by identifying events and their participants, known as thematic roles [29]. Each component of a sentence serves a specific function, and thematic roles are essential in shaping our understanding of sentences. They provide a structured framework for linking semantic meaning to syntactic structure, making them fundamental to language comprehension [30]. The agent (who performs the action) and the patient (who receives the action) are common roles, but others exist. The syntactic stage structures the sentence by arranging key components such as the subject, verb, and object, though assigning roles can be complex. Finally, the phonological stage prepares the sentence for speech by determining syllables, stress patterns, and intonation, ensuring smooth articulation [29, 31].
Psycholinguists believe sentence comprehension involves syntax (structure) and semantics (meaning), but they debate whether this occurs in one or two stages [29]. The two-stage model suggests syntax is processed first, then meaning, while the constraint-based model (one stage) argues both happen simultaneously. A “Garden-path” sentence initially leads the reader or listener to an incorrect interpretation before requiring reanalysis due to an unexpected structure or meaning [31, 32, 33]. For example, in “While Sarah bathed her baby played on the floor ,” we first assume “baby” is the object of “bathed,” but “played” disrupts this, requiring reinterpretation [34]. The garden path model is typically associated with the two-stage approach. Research indicates people read ambiguous sentences more slowly and often reread them [35], which supports the garden path model, where syntax is analyzed before meaning. Because language is processed rapidly and working memory is limited, we rely on heuristics—quick mental shortcuts—that usually work but sometimes cause errors, as with garden path sentences.
The garden path model explains sentence structure assignment using two heuristics. One is “Late Closure,” which means we keep adding new words to the current phrase unless there is strong evidence to start a new one. This often leads to misinterpretations because our brains tend to assume a subjectverb-object structure [29, 31, 32, 33]. Even when we detect a parsing error, the initial interpretation can linger. Research shows people who hear ambiguous sentences often recall them incorrectly later, demonstrating we rely on a “good enough” approach in everyday communication, where speed can outweigh perfect accuracy [36]. Late closure can also cause early misinterpretation by prematurely “closing” a structure [29, 33]. Some sentences contain complex grammar, like reduced relative clauses, which can be dificult to process. When a word is syntactically ambiguous, the brain tends to pick the more common structure, sometimes leading us down the wrong garden path [29].
“Minimal Attachment” is another garden path heuristic, which assumes the simplest possible sentence structure [32, 33]. This strategy can lead to initial interpretations that later need correction if they clash with the intended meaning [29]. Minimal attachment assumes a prepositional phrase attaches to the main verb rather than to an object noun phrase, simplifying processing but occasionally causing misunderstandings. There are two types of attachment: high attachment which links the phrase to the verb, and low attachment which links it to the object [37]. In syntactic theory, the verb is structurally higher than the object, so high attachment is simpler [29]. However, context also strongly influences how people interpret ambiguous sentences, guiding them to the correct attachment [29, 32, 33].
“Syntactic Priming” is the tendency to repeat a previously encountered sentence structure [38]. Studies show repeated structures are processed more easily, especially if the verb is also repeated [29, 32, 39, 40]. During comprehension, listeners often anticipate upcoming words while constructing sentence structures, a phenomenon known as “Cloze Probability” [41]. Research using the visual world paradigm reveals that visual context helps predict words [42, 43]. Event-Related Potential (ERP) studies show that unexpected words trigger an N400 response, basically a negative brain wave peaking at 400 milliseconds after a stimulus; in this context, it means the brain detects a meaning mismatch [44], reflecting semantic dificulty, while unexpected sentence structures elicit a P600 response, basically a positive brain wave peaking at 600 milliseconds after a stimulus; in this context, it means the brain struggles with sentence structure [45], indicating syntactic processing issues. Expectation thus plays a central role in language comprehension: the brain continuously makes predictions about upcoming words and structures, enabling eficient speech processing despite noise and distractions [46].
As shown in Section 2.1 and Figure 2, the time of provision, demand requirements, and diferent types of AV explanations shape driver trust, mental workload, and other related factors. Psycholinguistic insights from section 2.2 reveal how human can process these explanations in natural language comprehension. The garden path model suggests complex syntactic structures may mislead drivers, requiring reinterpretation that raises cognitive load and anxiety during driving [31]. Heuristics like minimal attachment and late closure show that drivers may favor simple interpretations, and mismatches with expectations may erode trust [29].
Demand for explanations ties to language anticipation. The constraint-based model, blending syntax and meaning, implies unclear or context-lacking explanations can confuse drivers, especially in highstakes scenarios [29]. Experienced drivers, leveraging cloze probability, may need less detail, while novices demand clear, structured explanations to align with their cognitive adaptability [41]. Explanation type afects acceptance and accountability. Syntactic priming indicates familiar structures boost intelligibility by easing efort, while N400 and P600 responses flag processing spikes from unexpected phrasing, potentially lowering acceptance [39, 44]. Misrecall of ambiguous sentences underscores the need for concise, well-timed explanations to ensure accurate understanding [36]. Thus, psycholinguistics presents a valuable opportunity to enhance the efectiveness of AV explanations by ofering a user-aligned approach that improves trust, reduces anxiety, and can optimize cognitive load in driver-AV interactions.
Figure 3 illustrates the PsyLingXAV framework, a theoretically grounded approach that integrates psycholinguistics with XAI to transform raw decision outputs from AV systems into cognitively optimized, linguistically structured explanations for the AV’s HMI. The framework operates through three interconnected phases.
In Phase 1, the AV collects multisensory inputs, which are processed by an AI black-box model to generate decision outputs. In Phase 2, the XAI module extracts critical decision information and refines it using a psycholinguistics module, implemented through natural language processing techniques or large language models. To ensure rapid comprehension during driving, PsyLingXAV employs predefined explanation templates that simplify syntactic structures based on minimal attachment and late closure heuristics, avoiding complexity associated with garden-path sentences. The system also leverages cloze probability and constraint-based parsing to align explanations with user expectations, and syntactic priming to maintain consistent sentence patterns, thereby reducing cognitive load and minimizing trust erosion. The implementation of this phase can follow a similar approach to RAGDriver [47], a retrieval-augmented, multi-modal large language model that employs in-context learning to generate explanation patterns similar to those found in the BDD-X dataset [48]. In Phase 3, the HMI delivers these explanations through visual displays (text) or audio outputs, enabling real-time driver comprehension. For example, a raw AV output such as “There is no obstacle detected ahead. The car is initiating acceleration.” would be translated into a psycholinguistically optimized explanation: “I’m speeding up since ahead is clear.” This transformation enhances syntactic simplicity, preserves familiar sentence patterns, and supports faster driver understanding. The framework also integrates a continuous Feedback & Iteration process, collecting passive behavioral data (e.g., reaction times, steering corrections, braking) and active user inputs (e.g., survey responses, verbal feedback). These insights allow dynamic adaptation of explanation syntax, timing, and detail to individual user profiles while maintaining psycholinguistic alignment.
Although the PsyLingXAV framework is based on established psycholinguistic principles, its efectiveness in driving scenarios requires empirical validation. A user study with a VR-based driving simulator will expose participants to three conditions: (1) no explanation, (2) GPT-generated explanations, and (3) psycholinguistically optimized explanations (PsyLingXAV). This controlled setup will allow direct comparison of explanation styles under realistic but safe conditions. Success will be evaluated through trust, cognitive workload, comprehension, and user preference metrics. Findings will inform framework refinements, and future work will explore multimodal extensions, including spoken outputs, with additional research needed to align auditory explanations with psycholinguistic principles.
This paper introduced PsyLingXAV, a psycholinguistic framework for XAI in AVs, designed to enhance AV explanation efectiveness by aligning them with users’ cognitive and linguistic processes. PsyLingXAV integrates insights from user-centered studies on AV explainability, grounded in robust psycholinguistic principles. While further empirical validation via real-world testing and human-centered evaluations is needed, PsyLingXAV ofers a promising approach to improving AV explainability. It lays a foundation for advancing human-centered explanations in AVs through psycholinguistics, paving the way for more intuitive user-AV interactions.
This research was supported partially by the Australian Research Council Discovery projects funding scheme (DP220102598).
The authors used GenAI for grammar check and language assistance during the preparation of this work. They reviewed, edited, and take full responsibility for the final content. [3] A. M. Nascimento, L. F. Vismari, C. B. S. T. Molina, P. S. Cugnasca, J. B. Camargo, J. R. de Almeida, R. Inam, E. Fersman, M. V. Marquezini, A. Y. Hata, A systematic literature review about the impact of artificial intelligence on autonomous vehicle safety, IEEE Transactions on Intelligent Transportation Systems 21 (2019) 4928–4946. [4] S. Tekkesinoglu, A. Habibovic, L. Kunze, Advancing explainable autonomous vehicle systems: A comprehensive review and research roadmap, ACM Transactions on Human-Robot Interaction 14 (2025) 1–46. [5] W. J. Von Eschenbach, Transparency and the black box problem: Why we do not trust ai, Philosophy & Technology 34 (2021) 1607–1622. [6] C. Rudin, Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead, Nature machine intelligence 1 (2019) 206–215. [7] A. Preece, D. Harborne, D. Braines, R. Tomsett, S. Chakraborty, Stakeholders in explainable ai, arXiv preprint arXiv:1810.00184 (2018). [8] P. Linardatos, V. Papastefanopoulos, S. Kotsiantis, Explainable ai: A review of machine learning interpretability methods, Entropy 23 (2020) 18. [9] D. Calvaresi, A. Najjar, M. Schumacher, K. Främling, Explainable, Transparent Autonomous Agents and Multi-Agent Systems: First International Workshop, EXTRAAMAS 2019, Montreal, QC, Canada, May 13–14, 2019, Revised Selected Papers, volume 11763, Springer Nature, 2019. [10] I. Lage, E. Chen, J. He, M. Narayanan, B. Kim, S. Gershman, F. Doshi-Velez, An evaluation of the human-interpretability of explanation, arXiv preprint arXiv:1902.00006 (2019). [11] A. Explainable, The basics policy briefing, Available at royalsociety. org/ai-interpretability (2019). [12] S. Kuutti, R. Bowden, Y. Jin, P. Barber, S. Fallah, A survey of deep learning applications to autonomous vehicle control, IEEE Transactions on Intelligent Transportation Systems 22 (2020) 712–733. [13] S. Hecker, D. Dai, L. Van Gool, End-to-end learning of driving models with surround-view cameras and route planners, in: Proceedings of the european conference on computer vision (eccv), 2018, pp. 435–453. [14] L. Claussmann, M. Revilloud, D. Gruyer, S. Glaser, A review of motion planning for highway autonomous driving, IEEE Transactions on Intelligent Transportation Systems 21 (2019) 1826–1848. [15] H. A. Tahir, W. Alayed, W. U. Hassan, A. Haider, A novel hybrid xai solution for autonomous vehicles: Real-time interpretability through lime–shap integration, Sensors 24 (2024) 6776. [16] K. Panduru, J. Walsh, et al., Exploring the unseen: A survey of multi-sensor fusion and the role of explainable ai (xai) in autonomous vehicles, Sensors (Basel, Switzerland) 25 (2025) 856. [17] T. Miller, Explanation in artificial intelligence: Insights from the social sciences, Artificial intelligence 267 (2019) 1–38. [18] D. Wang, Q. Yang, A. Abdul, B. Y. Lim, Designing theory-driven user-centric explainable ai, in:
Proceedings of the 2019 CHI conference on human factors in computing systems, 2019, pp. 1–15. [19] Q. Meteier, M. Capallera, L. Angelini, E. Mugellini, O. A. Khaled, S. Carrino, E. De Salis, S. Galland, S. Boll, Workshop on explainable ai in automated driving: a user-centered interaction approach, in: Proceedings of the 11th International Conference on Automotive User Interfaces and Interactive Vehicular Applications: Adjunct Proceedings, 2019, pp. 32–37. [20] N. Du, J. Haspiel, Q. Zhang, D. Tilbury, A. K. Pradhan, X. J. Yang, L. P. Robert Jr, Look who’s talking now: Implications of av’s explanations on driver’s trust, av preference, anxiety and mental workload, Transportation research part C: emerging technologies 104 (2019) 428–442. [21] T. Ha, S. Kim, D. Seo, S. Lee, Efects of explanation types and perceived risk on trust in autonomous vehicles, Transportation research part F: trafic psychology and behaviour 73 (2020) 271–280. [22] Y. Shen, S. Jiang, Y. Chen, E. Yang, X. Jin, Y. Fan, K. Campbell, To explain or not to explain: A study on the necessity of explanations for autonomous vehicles, arxiv, arXiv preprint arXiv:2006.11684 (2020). [23] D. Omeiza, H. Web, M. Jirotka, L. Kunze, Towards accountability: Providing intelligible explanations in autonomous driving, in: 2021 IEEE Intelligent Vehicles Symposium (IV), IEEE, 2021, pp. 231–237. [24] A. Garnham, Psycholinguistics (PLE: Psycholinguistics): Central Topics, Psychology Press, 2013. [25] J. Koo, J. Kwac, W. Ju, M. Steinert, L. Leifer, C. Nass, Why did my car just do that? explaining semiautonomous driving actions to improve driver understanding, trust, and performance, International Journal on Interactive Design and Manufacturing (IJIDeM) 9 (2015) 269–275. [26] Q. Zhang, X. J. Yang, L. P. Robert Jr, Drivers’ age and automated vehicle explanations, Sustainability 13 (2021) 1948. [27] N. Epley, A. Waytz, J. T. Cacioppo, On seeing human: a three-factor theory of anthropomorphism.,
Psychological review 114 (2007) 864. [28] Y. Forster, F. Naujoks, A. Neukum, Increasing anthropomorphism and trust in automated driving functions by adding speech output, in: 2017 IEEE intelligent vehicles symposium (IV), IEEE, 2017, pp. 365–372. [29] D. Ludden, The psychology of language: an integrated approach, Sage Publications, 2015. [30] C. Anderson, Essentials of linguistics, McMaster University, 2018. [31] T. A. Harley, The psychology of language: From data to theory, Psychology press, 2013. [32] M. J. Traxler, Introduction to psycholinguistics: Understanding language science (2011). [33] P. Warren, Introducing psycholinguistics, Cambridge University Press, 2013. [34] N. D. Patson, E. S. Darowski, N. Moon, F. Ferreira, Lingering misinterpretations in garden-path sentences: evidence from a paraphrasing task., Journal of Experimental Psychology: Learning, Memory, and Cognition 35 (2009) 280. [35] L. Frazier, K. Rayner, Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences, Cognitive psychology 14 (1982) 178–210. [36] I. Ivanova, M. J. Pickering, H. P. Branigan, J. F. McLean, A. Costa, The comprehension of anomalous sentences: Evidence from structural priming, Cognition 122 (2012) 193–209. [37] M. J. Pickering, J. F. McLean, H. P. Branigan, Persistent structural priming and frequency efects during comprehension., Journal of Experimental Psychology: Learning, Memory, and Cognition 39 (2013) 890. [38] J. K. Bock, Syntactic persistence in language production, Cognitive psychology 18 (1986) 355–387. [39] K. M. Tooley, M. J. Traxler, T. Y. Swaab, Electrophysiological and behavioral evidence of syntactic priming in sentence comprehension., Journal of Experimental Psychology: Learning, Memory, and Cognition 35 (2009) 19. [40] K. Segaert, G. Kempen, K. M. Petersson, P. Hagoort, Syntactic priming and the lexical boost efect during sentence production and sentence comprehension: An fmri study, Brain and language 124 (2013) 174–183. [41] E. W. Wlotko, K. D. Federmeier, Age-related changes in the impact of contextual strength on multiple aspects of sentence comprehension, Psychophysiology 49 (2012) 770–785. [42] M. K. Tanenhaus, M. J. Spivey-Knowlton, K. M. Eberhard, J. C. Sedivy, Integration of visual and linguistic information in spoken language comprehension, Science 268 (1995) 1632–1634. [43] M. Tanenhaus, Sentence comprehension, Handbook of Perception and Cognition 11 (1995). [44] L. Hunt III, S. Politzer-Ahles, L. Gibson, U. Minai, R. Fiorentino, Pragmatic inferences modulate n400 during sentence comprehension: Evidence from picture–sentence verification, Neuroscience Letters 534 (2013) 246–251. [45] L. Osterhout, P. J. Holcomb, Event-related brain potentials elicited by syntactic anomaly, Journal of memory and language 31 (1992) 785–806. [46] E. Gibson, L. Bergen, S. T. Piantadosi, Rational integration of noisy evidence and prior semantic expectations in sentence interpretation, Proceedings of the National Academy of Sciences 110 (2013) 8051–8056. [47] J. Yuan, S. Sun, D. Omeiza, B. Zhao, P. Newman, L. Kunze, M. Gadd, Rag-driver: Generalisable driving explanations with retrieval-augmented in-context learning in multi-modal large language model, arXiv preprint arXiv:2402.10828 (2024). [48] J. Kim, A. Rohrbach, T. Darrell, J. Canny, Z. Akata, Textual explanations for self-driving vehicles, Proceedings of the European Conference on Computer Vision (ECCV) (2018).