
Abstract Background Children's accounts in legal contexts can be influenced by leading and repetitive questioning. Best‐practice guidance recommends minimizing yes/no questions, although some are still asked once open‐ended prompts are exhausted. The Bonn Test of Statement Suggestibility (BTSS) quantifies interrogative suggestibility via Yield (endorsing misleading suggestions) and Shift (changing answers after repetition and implied disagreement). In Mandarin, polar questions can be asked with a sentence‐final particle (–ma; e.g., ‘The little girl in the story is called Xiao Bei, ma?’, roughly ‘…, right?’) or with an A‐not‐A structure (e.g., ‘The little girl in the story shi‐bu‐shi (is or isn't) called Xiao Bei?’), which makes both alternatives explicit. Aim We tested whether this linguistic difference affects children's suggestibility in response to yes/no questions. Materials and Methods Sixty‐eight Mandarin‐speaking children (4–8 years) were randomly assigned to a BTSS version asking all yes/no questions in either –ma or A‐not‐A form; forced‐choice questions were identical across versions. Yield and Shift were analysed separately for yes/no questions using binomial logit models with age, intelligence and age × format terms. Results Yield decreased with age but did not differ by question format. For Shift on repeated yes/no questions, the Original (–ma) version produced more answer changes than the A‐not‐A version (OR mean_age = 2.16, 95% CI [1.24, 3.76]). This effect survived bootstrapping; the exploratory age‐dependent format effects did not. Discussion The linguistic form of yes/no questions in Mandarin can matter for children's answer stability under pressure. Conclusions Where yes/no questions are used in Mandarin‐speaking interviews despite best‐practice guidance, A‐not‐A formulations may help children maintain their answers under repeated questioning, although this is not a recommendation to use them.
Abstract Purpose Interviewer questioning style, including the types of questions asked, represents one measurable and trainable component of child investigative interview quality. Although automated approaches to question classification have been explored, previous systems have shown only moderate performance, particularly when distinguishing fine‐grained categories of questions. The present study aimed to improve the automatic multi‐class classification of interviewer questions and the detection of repetitive questioning patterns in avatar‐based simulations of child sexual abuse investigative interviews using transformer‐based architectures. Method The system was trained on 17,425 annotated questions derived from controlled avatar‐based interview simulations and categorized into 14 question‐category labels. A transformer model (F2LLM‐0.6B) was fine‐tuned for multi‐class classification using weighted loss to address class imbalance. A separate LLM‐based module (Qwen3.5‐9B) detected semantically repeated or reformulated questions within interview sequences. Performance was evaluated using accuracy, macro‐ and weighted F1‐scores, and confusion matrix analyses. On a held‐out test set ( N = 2, 614), the classification model achieved an overall accuracy of 86.8% and a macro‐F1 score of 79.0%, while the repetition‐detection module achieved an accuracy of 84.9%. Conclusions These findings indicate that AI systems can approximate the available human‐coded reference labels for interviewer question types and identify repeated or reformulated questioning patterns. The system supports taxonomy‐based coding of interviewer questions; it does not independently determine their pragmatic or contextual appropriateness within the broader interview sequence, nor does it assess overall interview quality. The proposed approach provides a scalable framework for automated question‐type coding in avatar‐based training and research settings while preserving the central role of professional judgement.
Purpose This study investigated the role of varied legal training in rule-based legal thinking (RLT)-the capacity to apply legal rules and principles even when they conflict with lay conceptions of justice in legal decision-making.Methods To address these objectives, we conducted two online studies involving a diverse sample of participants (NStudy 1 = 281; NStudy 2 = 13,690). In Study 1, participants evaluated case-specific civil vignettes to determine defendant liability and compensation amounts, followed by a self-report measure capturing the underlying reasoning behind their verdicts. In Study 2, we evaluated RLT using a broader, more generalized assessment instrument.Results Study 1 showed that legally trained participants (including law practitioners and law students) exhibited a significantly stronger propensity to apply RLT in legal judgements and made more lenient judgements regarding liability and compensation. Study 2 showed that both formal and informal legal training were significantly associated with higher levels of RLT. Furthermore, the association between legal training and RLT was statistically accounted for primarily by differences in applied legal knowledge and, to a lesser extent, by a serial indirect pathway involving applied legal knowledge and endorsement of Care and Fairness as measured by the Moral Foundations Questionnaire.Conclusion These findings provide support for the potential role of legal training in RLT and offer insights into potential pathways for strengthening laypeople's capacity to engage in rule-based legal decision-making.
Purpose Restorative justice (RJ) offers an alternative or complement to criminal justice in sexual abuse cases, yet the public remains ambivalent. Given that survivors increasingly turn to large language models (LLMs) for information and support when coping with sensitive justice-related decisions, the present study examined whether the guidance LLMs provide regarding RJ in sexual abuse cases resembles human views.Methods Participants (N = 469) were randomly assigned to read a vignette describing intra- or extrafamilial sexual abuse and rated offence severity and support for RJ. Their responses were compared with those generated by Gemini, Grok and DeepSeek (120 total outputs).Results Both humans and LLMs perceived intrafamilial sexual abuse as more severe than extrafamilial abuse. However, in both contexts, LLMs produced significantly higher numeric ratings of RJ support than did humans under the structured prompting conditions used. Among humans, perceived severity fully mediated the relationship between abuse context and RJ support, such that higher severity reduced openness to RJ. Among LLMs, this mediation was partial, such that heightened severity did not serve as a categorical veto against RJ. Significant differences were observed between models.Conclusions Under structured prompting conditions, LLM-generated outputs acknowledged abuse severity while remaining relatively open to RJ, a pattern consistent with a cautious, descriptive use of the 'mirror cleaner' metaphor. The findings have important implications for survivors who turn to AI systems for information and support, underscoring both their potential and the need for careful ethical curation.
Purpose AI-assisted adjudication is increasingly embedded in judicial decision-making, yet whether legal professionals and laypersons can independently guard against overreliance remains contested. This study examines how people with varying legal expertise respond to demonstrably flawed AI recommendations and whether reliance differs across legal-error types.Methods Legal professionals, law students and laypersons each adjudicated one criminal and one civil case after random assignment to correct or incorrect AI-generated recommendations. Incorrect advice involved two error types: misapplication of statutory provisions and erroneous citation of judicial precedents. Responses were coded using two independent adoption indicators. Explicit adoption captured direct incorporation of the AI-provided legal basis; implicit adoption captured a final decision materially aligned with the AI recommendation without requiring explicit incorporation of that basis.Results Explicit adoption showed a floor effect: participants rarely reproduced, cited or endorsed the AI-provided legal basis, regardless of advice accuracy or legal expertise. Implicit adoption was significantly less frequent for erroneous than correct advice. Legal expertise showed neither a significant main effect nor an interaction with advice condition, and no detectable expertise differences emerged in either adoption outcome. Subjective trust similarly declined significantly in response to AI errors, with no detectable main effect of legal expertise or interaction with advice condition.Conclusions These findings are consistent with selective adoption under the tested conditions: across expertise levels, participants aligned less often with erroneous than with correct AI recommendations. Nevertheless, further scrutiny is needed of how AI systems influence legal reasoning in practice.