搜索结果：rewritten

Radiology reports are typically written in language that is difficult for patients to understand. Large language models (LLMs) excel at simplifying text. We aimed to evaluate the ability of LLMs to improve the understanding of radiology reports. In this systematic review and meta-analysis, we searched CENTRAL, MEDLINE, and Embase from inception to Nov 11, 2025, without restrictions on language. Full-text articles and preprints were considered for inclusion. Eligible studies applied LLMs to simplify radiology reports and had these reports assessed by members of the public or medical professionals. We excluded studies that focused solely on dialogues with interactive chatbots, preimaging leaflets, educational materials, appointment letters, or summarising findings without simplifying them for patients. Search results were screened independently by two authors and full-text review and data extraction were done by three authors; disagreements were resolved by consensus. The main outcomes were patient, public, and clinician evaluations (Likert scores) and text readability metrics. We assessed study quality with the MAIC-10 tool. This study was registered with PROSPERO (CRD420251027489). We identified 2385 records, of which 38 studies were eligible. These 38 studies generated 12 922 simplified reports, assessed by 508 evaluators (387 lay people and 121 clinicians). 35 (92%) of 38 studies used OpenAI GPT models and 29 (76%) produced simplified reports in English. Patients perceived LLM-rewritten reports as significantly more understandable than radiologist reports (mean Likert score 4·04 [SD 1·20] for simplified reports vs 2·16 [SD 0·94] for original reports; mean difference 2·00 [95% CI 1·54-2·46]). Clinicians rated LLM-rewritten reports highly for accuracy (mean 4·45 [95% CI 4·27-4·63]; 27 studies) and completeness (mean 4·53 [95% CI 4·30-4·76]; 14 studies). Readability was improved across imaging modalities, with lower Flesch-Kincaid Grade Level for LLM-rewritten reports, including a mean difference of -6·20 (95% CI -6·91 to -5·48) for CT, -5·07 (-5·99 to -4·15) for x-ray, and -5·0 (-6·0 to -4·0) for MRI. The error rate in LLM-rewritten reports was 7·2% (95% CI 5·1%-10·0%; 13 studies) and 0·9% (95% CI 0·6-1·5%; 2 studies) for clinically significant errors. LLM-simplified radiology reports improved patient-perceived understanding and readability and were rated by clinicians as largely accurate and complete, although a small proportion contained clinically significant errors. LLM-based simplification shows promise for making radiology communication more patient-centred, but further evaluation of its effect on patient outcomes and clinical workflows is required. National Institute for Health and Care Research Sheffield Biomedical Research Centre.

Mental health problems among university students are a growing global concern, yet limited counseling resources and inadequate understanding of counseling procedures often delay timely help-seeking. Informed consent forms (ICFs) are essential for safeguarding autonomy and clarifying counseling procedures, but many universities' counseling ICFs are incomplete, ambiguous, or overly technical. Large language models (LLMs) may offer scalable assistance for improving clarity and accessibility. This study aimed to evaluate whether LLM-based rewriting could improve the structure, readability, content quality, and comprehensibility of university counseling ICFs, and compared 2 advanced models (ChatGPT [GPT-5] and Grok-4). We conducted a comparative evaluation of counseling ICFs collected from 33 Chinese universities (original texts) and generated 2 rewritten versions for each ICF using ChatGPT (GPT-5) and Grok-4. A multidimensional framework assessed (1) textual structure and readability, (2) expert-rated content quality from a counselor perspective, and (3) volunteer-rated reading comprehension from a client perspective. Comparisons between original and rewritten texts were performed using Wilcoxon signed rank tests, with linear mixed-effects models used to validate results while accounting for rater variability. Compared with the originals, both LLM-rewritten ICFs showed significant improvements across all evaluated dimensions. The mean Lee-Yang Readability Index decreased from 28.68 (SD 5.69) to 22.39 (SD 2.13) with ChatGPT (GPT-5) and 24.37 (SD 2.32) with Grok-4 (both P<.001), and mean tone friendliness increased from 2.57 (SD 0.29) to 2.67 (SD 0.12) and 2.67 (SD 0.13), respectively. The mean expert-rated content quality improved from 45.33 (SD 8.74) to 52.54 (SD 7.92) and 55.49 (SD 7.81) (P<.001), driven mainly by higher completeness and specificity of key information. The mean volunteer-rated reading comprehension scores increased from 19.02 (SD 1.32) to 22.33 (SD 0.81) and 22.05 (SD 0.90) (P<.001), indicating improved clarity, readability, and acceptability. Across structural features, Grok-4 tended to produce longer rewritten forms than the originals, highlighting a potential trade-off between added informational content and document length. In this comparative evaluation of 33 Chinese university counseling ICFs, LLM-based rewriting was associated with improved readability, expert-rated content quality, and volunteer-rated comprehension relative to original forms. These findings suggest that LLMs can support the optimization of counseling documentation; however, implementation should consider practical constraints (eg, document length) and retain human oversight.

We reconsider information-theoretic principles, such as the maximum entropy principle/minimum Massieu potential principle, from the perspective of the dual probability distribution. This is introduced through Sanov's Lemma for the multinomial distribution. The dual correspondence becomes asymptotically manifest. The Massieu potential is rewritten as the Kullback-Leibler divergence between the dual probability distribution and the dual reference distribution. Similarly, the dual potential is rewritten as the cumulant generating function with respect to the dual reference distribution. This perspective gives us new insight into information-theoretic principles. As the dual probability distribution naturally arises in data sampling, we anticipate that this new perspective will play a significant role in data analysis.

Psychiatric discharge summaries are vital for ensuring continuity of care, yet they are often written in technical language that can be difficult for patients to understand and may cause emotional distress or reinforce stigma. With increasing patient access to medical records, there is a pressing need to develop communication tools that are both comprehensible and emotionally safe. This study aimed to evaluate the diagnostic fidelity, linguistic clarity, emotional sensitivity, treatment comprehension, and readability of psychiatric discharge summaries rewritten by ChatGPT-4 based on real clinical cases. This was the first study in South America to examine the use of a generative language model for rewriting psychiatric discharge summaries. A mixed-methods, observational cross-sectional design was applied. Twenty-five anonymized clinical cases were rewritten using ChatGPT-4. Three psychiatrists independently assessed each AI-generated summary across four dimensions: diagnostic fidelity, clarity of language, perceived emotional risk, and understanding of treatment. Readability was evaluated using the Fern&#xe1;ndez-Huerta Index and the INFLESZ Scale. A thematic analysis of evaluators' written comments was also conducted. Summaries generated by ChatGPT-4 were rated positively, particularly for clarity and treatment explanation. Significant improvements in readability were observed across all diagnostic groups (p&#x2009;<&#x2009;.001), with mean values surpassing recommended thresholds for general comprehension. However, five summaries remained below those thresholds, and some diagnostic inaccuracies were noted (e.g. omissions in bipolar disorder). Evaluators also highlighted emotionally charged or stigmatizing language in a few cases. ChatGPT-4 can enhance the accessibility and emotional appropriateness of psychiatric discharge communication, supporting more patient-centered care. Nevertheless, professional oversight remains critical to ensure clinical accuracy and contextual sensitivity. Future research should include patient feedback, assess long-term outcomes, and explore hybrid human-AI collaboration models.

Effective communication about clinical trials is essential, as low enrollment undermines scientific validity and contributes to health care inequities. However, recruitment remains a persistent challenge, particularly among older adults, minority populations, and individuals with limited health literacy. Although large language models (LLMs) show promise in understanding and generating health information, it is unclear whether these generative artificial intelligence (AI) tools can improve the content of hospitals' frequently asked questions (FAQ) pages to enhance public attitudes and intentions toward clinical trial participation. This study aimed to compare clinical trial FAQ from health organizations and hospitals with versions rewritten by LLMs to examine whether the generated content improves public attitudes and intentions toward clinical trial participation and to identify the mechanisms underlying these effects. A total of 308 question-answer pairs were collected from the FAQ pages of 38 health organizations and hospitals, categorizing them into 52 types and selecting the 11 most frequent for testing. A comparative survey experiment was conducted with 440 participants randomly assigned to one of the two survey stimuli: the original FAQ versus the GPT-4o-generated answers emphasizing comprehension and empathy. The study compared the impact of AI-generated versus standard FAQ content on attitudes toward clinical trials and examined Theory of Planned Behavior constructs to determine for whom and how AI information is most effective. Participants were recruited through CloudResearch, yielding a 96.94% completion rate, resulting in 440 valid responses across the 2 types of content exposure. Participants who viewed GPT-4o-generated information (mean 0.26, SD 0.65) showed a marginally greater positive change in outcome evaluation attitudes than those who viewed standard FAQ (mean 0.13, SD 0.70; P=.05; 95% CI 0.00-0.25). Follow-up linear regression analyses revealed that several individual factors significantly moderated the effect of the information type (FAQ vs GPT-4o) on attitude change, including age (mean difference 0.87, SE 0.33; 2-tailed t394=2.62; P=.009); race (mean difference 0.36, SE 0.15; t383=2.47; P=.01); risk aversion (B=0.12; SE 0.06; t383=2.23; P=.03); fear of ineffective treatment (B=0.11; SE 0.05; t383=2.03; P=.04); and fear of unknown treatment effects (B=0.21; SE 0.07; t383=3.10; P=.002). This study is the first to apply the Theory of Planned Behavior to compare LLM-rewritten versus original FAQ content for clinical trial communication. The findings show that the GPT-4o-generated responses improved attitudes among traditionally underrepresented groups, including older adults, Black participants, and those with higher uncertainty avoidance or treatment concerns. These attitude gains were positively linked to participation intentions, suggesting that AI-generated language can enhance public attitudes, perceptions, and engagement with clinical research.

Data scarcity is a long-standing challenge in the vision-language navigation (VLN) field, which extremely hinders the generalization of agents to unseen environments. Previous works primarily rely on additional simulator data or web-collected images/videos to improve the generalization. However, the simulator environments still face limited diversity, and the web-collected data often require extensive labor to remove the noise. In this article, we propose a Rewriting-driven AugMentation (RAM) paradigm for VLN, which directly creates the unseen observation-instruction pairs via rewriting human-annotated training data. Benefiting from our rewriting mechanism, new observation-instruction pairs can be obtained in both simulator-free and labor-saving manners to promote generalization. Specifically, we first introduce object-enriched observation rewriting, where we combine vision-language models (VLMs) and large language models (LLMs) to derive rewritten object-enriched scene descriptions, enabling observation synthesis with diverse objects and spatial layouts via text-to-image generation models (T2IMs). Then, we propose observation-contrast instruction rewriting, which generates observation-aligned rewritten instructions by requiring LLMs to reason the difference between original and new observations. We further develop a mixing-then-focusing training strategy with a random observation cropping scheme, effectively enhancing data distribution diversity while suppressing augmentation data noise during training. Experiments on both the discrete environments (R2R, REVERIE, and R4R datasets) and continuous environments (R2R-CE dataset) show the superior performance and impressive generalization ability of our method.

Variants in promoters and enhancers can alter the binding of transcription factors (TFs), but their functional assessment remains difficult. FABIAN-variant is a web application that predicts the effects of DNA variants on TF binding by comparing position weight matrix (PWM) and transcription factor flexible model (TFFM) scores between reference and variant alleles. Here, we present FABIAN-variant 2026, a major update that expands the prediction model library from ~5000 to over 40&#x2009;000 models for >1500 human TFs, sourced from nine PWM databases and including 1290 TFFMs. The application now supports the mouse genome (GRCm38 and GRCm39) with over 35&#x2009;000 models for >1100 mouse TFs. An optional BPNet deep learning scorer provides neural network-based binding predictions for 240 human TFs. Known TF binding site information has been expanded from three to five sources. Predictions for over 1400 heterodimer TF complexes have been added. The web server has been rewritten in Rust and the scoring engine optimized, reducing runtime by ~70%. A RESTful JSON API and a standalone command-line version enable programmatic access and local high-throughput analysis. FABIAN-variant 2026 is available at https://fabianapp.org/variant26/. The web server is free and open to all users and there is no login requirement.

Large language models (LLMs) such as ChatGPT are increasingly used to generate patient education materials; however, default ChatGPT responses often exceed recommended readability levels set by the American Medical Association (AMA) and National Institutes of Health (NIH) health-literacy recommendations. The purpose of this study was to evaluate the readability and educational quality of ChatGPT-generated patient education on meniscal surgery and to determine whether a standardized plain-language prompt could improve readability without compromising accuracy, relevance, or depth. Sixteen standardized patient-focused questions regarding diagnosis, management, and prevention of meniscus tears were submitted to ChatGPT-5 and ChatGPT-4o, with three replicates per question to ensure standardization. Responses were assessed for accuracy against the American Academy of Orthopaedic Surgeons (AAOS) OrthoInfo and scored for relevance and depth using 5-point Likert scales. Readability was assessed using Flesch-Kincaid Grade Level (FKGL) and Flesch Reading Ease Score (FRES). All baseline responses were subsequently rewritten using a plain-language prompt targeting a sixth- to eighth-grade reading level. Pre- and post-prompt readability metrics were compared using paired t-tests. Inter-rater reliability was measured with Cohen's kappa. Both models demonstrated 100% factual accuracy across all baseline responses compared with OrthoInfo. Mean relevance and depth scores were high for ChatGPT-5 (4.49 &#xb1; 0.22; 4.39 &#xb1; 0.26) and ChatGPT-4o (4.55 &#xb1; 0.15; 4.53 &#xb1; 0.08). Baseline readability exceeded recommendations (FKGL 11.4-12.1; FRES 37-40). The plain-language prompt significantly improved readability for both models, reducing FKGL by approximately 5 grade levels and increasing FRES by 30-38 points (P < 0.001), with no loss of accuracy, relevance, or depth. ChatGPT generates accurate and relevant patient-directed content on meniscal surgery; however, readability frequently exceeded established health literacy standards. A simple, reproducible plain-language prompt reliably reduced reading level into the target range, offering a practical strategy for sports medicine surgeons to enhance informed consent discussions and patient education materials. IV.

Patient Information leaflets (PILs) from the British Association of Urologists (BAUS) are commonly used to communicate to patients about surgical procedures, but previous studies have highlighted that they are too difficult for some patients to read. BAUS PILs have since been rewritten using guidelines which emphasize readability. To identify if the readability of BAUS PILs has changed compared to historical versions. Current BAUS PILs (published 2020-2025) were compared with historical PILs with similar titles from 2014 to 2016 using a custom Python script. Readability scores improved by a significant (P&#x2009;<&#x2009;.001) albeit small amount (FKGL 8.83 vs 8.67; SMOG 11.86 vs 11.71; FRE 57.06 vs 57.47), and continue to have a suboptimal readability. The revised PILs had significantly fewer long sentences (10.11% vs 6.84% P&#x2009;<&#x2009;.0001), fewer sentences that used the "passive voice" (30.50% vs 16.22% P&#x2009;<&#x2009;.0001) and were shorter (1875.5 words vs 1726 P&#x2009;=&#x2009;.004). While there have been small improvements in urology PILs for patients, they remain too difficult for many patients, and more work is needed to improve readability.

The cell-based mucin array is a versatile platform for expressing and interrogating recombinant mucin reporter proteins with representative patterning and customizable O-glycan structures. The platform is based on glycoengineered mammalian cell lines (HEK293/CHO), in which the glycosylation machinery is genetically rewritten to enable controlled display of specific O-glycan core structures and terminal epitopes with defined sialylation, fucosylation, and sulfation. Uniquely, the platform presents glycans in their native protein context, enabling investigation of how O-glycan density, clustering, and multivalency influence interactions with mucins. A conceptual framework for recognition of such glycan-context cues is provided by the patterned arrangement of O-glycans within mucin O-glycodomains, often composed of tandem repeat (TR) sequences with distinct O-glycosites resembling molecular barcodes. Mucin reporters mimic the serine-, threonine-, and proline-rich domains of natural mucins and mucin-like proteins or they can be designed as artificial Glycocarriers with model O-glycan cluster motifs. Reporters expressed as membrane-bound forms for cell display or as secreted fusion proteins for production can be applied in diverse bioassays. They have been used to probe glycan/mucin binding by viral and microbial adhesins, as well as human Siglec immune receptors. Moreover, the platform provides defined substrates for functional analyses of mucin and O-glycodomain degradation by microbial O-glycopeptidases or mucinases, revealing information of substrate specificities, cleavage points, and catalytic mechanisms. This chapter describes how the cell-based mucin array can be used to dissect interactions with mucins by glycan- and mucin-binding receptors as well as mucinases. The platform overcomes limitations of contemporary technologies by enabling studies with mimics of natural mucins and O-glycodomains that preserve clustered and patterned O-glycan contexts.

This paper explores how digital disruption necessitates a new understanding of business resilience. The rules of business resilience are being rewritten, rendering traditional disaster recovery inadequate and accelerating corporate obsolescence. Standing on the sidelines is no longer an option; rather, it is a fast track to irrelevance. This paper addresses the critical challenge of navigating an era of unprecedented digital disruption by presenting a holistic, transformational framework for building reimagined resilience. It involves three core phases: align, react and reinvent. The align phase emphasises establishing a clear 'north star' and 'why', driven by C-suite accountability and fostering an engaging, psychologically safe environment for co-creation. Next, the react phase focuses on building a robust digital and data foundation. This includes a business-driven data strategy, building a comprehensive cybersecurity posture beyond mere technical measures, and an agile, platform-based infrastructure. Most importantly, the author presents an argument for cultivating a digital-ready workforce skilled in experimentation and collaboration. Finally, reinvent necessitates the continuous re-architecture of operating models, leveraging digital factories and treating data as a strategic asset through data productisation. Crucially, it involves fostering an adaptive culture that embraces continuous innovation and learns from failure, recognising that building digital resilience is not a one-off exercise, nor is it a hype-induced necessity. Readers will gain actionable insights into navigating this ongoing transformation, understanding that building digital resilience is the fundamental license to operate in our rapidly evolving world. This article is also included in The Business & Management Collection which can be accessed at https://hstalks.com/ business/.

Analogical reasoning involves mapping relational structure from a source domain to a target domain. Distant analogies, compared with near analogies, require reasoners to go beyond surface similarity and extract deeper relational structure. Because utilitarian judgments in sacrificial moral dilemmas also involve evaluating structural trade-offs between costs and benefits, the present study examined whether solving distant analogies increases approval of utilitarian actions. In Experiment 1, 175 Chinese university students were randomly assigned to solve either 40 distant or 40 near verbal analogies and then rated their approval of utilitarian actions across eight sacrificial moral dilemmas. In Experiment 2, 161 students completed the same procedure, except that the dilemmas were rewritten from a second-person to a third-person perspective. Across both experiments, participants in the distant-analogy condition reported significantly higher approval of utilitarian actions than those in the near-analogy condition. This effect remained significant after controlling for analogy accuracy in linear mixed-effects models, and the Condition&#x2009;&#xd7;&#x2009;Accuracy interactions were not significant in either experiment. These findings suggest that solving distant analogies can increase subsequent utilitarian approval in sacrificial moral dilemmas, possibly by promoting a relational, structure-focused processing orientation.

Online patient educational materials (PEMs) have poor readability, limiting their intended purposes in improving patient comprehension of health topics. Orthopaedic oncology PEMs are particularly complex. Although ChatGPT has demonstrated limited success in simplifying PEMs to the recommended sixth-grade reading level, other large language models (LLMs) have not been thoroughly evaluated. The goals of this study were to (1) assess baseline readability of online orthopaedic oncology PEMs, (2) evaluate five LLMs (ChatGPT-4o, Google Gemini, DeepSeek AI, Microsoft Copilot, and Meta AI) for improving readability while preserving accuracy and comprehension, and (3) to examine tradeoffs when PEMs were simplified below the sixth-grade level. Seventy-two PEMs were collected from academic and professional sources. Readability metrics included the Flesch-Kincaid Grade Level (FKGL), Gunning Fog Index (GFI), and Flesch Reading Ease (FRE). Each PEM was rewritten by the five LLMs using the prompt: "rewrite this document to a sixth-grade reading level." Two independent graders then evaluated outputs for comprehension and accuracy (F1 score). ANOVA with pairwise comparisons assessed differences among LLMs and versus baseline (PEMs as written). A secondary analysis evaluated the effect on readability, accuracy, and comprehension of prompts to the fifth-grade, fourth-grade, and third-grade reading level. Baseline FKGL (8.7 &#xb1; 1.5) was between the eighth-grade and ninth-grade reading level, and GFI (10.5 &#xb1; 1.9) was slightly higher. Baseline FRE was 53.9 &#xb1; 8.2. All LLMs significantly improved readability (P < 0.001), and ChatGPT-4o, DeepSeek AI, and Google Gemini conversion produced the most readable outputs. Google Gemini achieved the highest F1 score of 0.986 (range: 0.765-0.986) and 100% comprehension. Accuracy and comprehension were compromised for MetaAI when prompted below sixth grade. ChatGPT-4o, Google Gemini, and DeepSeekAI effectively improved readability while preserving comprehension and accuracy. These findings may guide patient use of LLMs and inform healthcare-AI partnerships.

The clinical presentation of childhood soft tissue sarcoma (STS) creates a diagnostic challenge; however, time to diagnosis is significantly associated with survival. This Delphi consensus process aims to contribute to the development of clinical guidelines for healthcare professionals (HCPs) assessing children and young people presenting with signs/symptoms suggestive of STS. Invitation emails were sent to HCPs to join the Delphi panel. 48 statements were derived from an evidence review by a multidisciplinary team. Participants ranked their level of agreement with each statement on a 9-point Likert scale (1=strongly agree, 9=strongly disagree), with responses &#x2265;7 indicating agreement. Statements not reaching consensus were rewritten and reissued in a subsequent round. In round 1 (R1), 87/117 (74%) participants responded and 68/87 (78%) completed round 2 (R2).In R1, 47/48 (98%) statements achieved consensus, with 16/48 (33%) gaining more than 90% consensus. One statement did not reach consensus, with 26/87 (29%) scoring <7. All statements reached numerical consensus at the end of R2.There was strong consensus for urgent referral and investigations in cases with strong clinical suspicion, urgent discussion with secondary care if clinical uncertainty exists, increasing HCP education and institutional support to overcome potential barriers to timely diagnosis. Dissent was due to logistical and systematic barriers and a concern for misinterpretation if intended setting is not clearly specified. These statements will be included in a clinical guideline for suspected STS for use in both primary and secondary care and translated into public awareness tools as part of the Child Cancer Smart national awareness campaign.

The transition to university is a critical life stage characterized by increased autonomy, identity exploration, and new social and environmental influences. During this period, university students often exhibit low adherence to dietary guidelines. Among the determinants influencing healthy eating, cooking self-efficacy, the central construct of Social Cognitive Theory (SCT), is consistently associated with improved diet quality and is a frequent target of health interventions. However, no validated instrument exists to assess this construct among university students in Spain. Therefore, the goal was to develop and provide preliminary evidence of the Spanish Cooking Self-Efficacy Questionnaire (SCSEQ), a concise SCT-based instrument tailored to Mediterranean university settings. A 32-item questionnaire was developed through a review of existing instruments assessing cooking self-efficacy. Face validity was evaluated with Spanish food and nutrition experts (n&#x2009;=&#x2009;12) to assess the clarity and pertinence of the initial items. The revised Spanish Cooking Self-Efficacy Questionnaire (SCSEQ) was then pilot-tested with Spanish university students (n&#x2009;=&#x2009;73) from four Catalan universities. Exploratory factor analysis (EFA) was conducted to identify the underlying factor structure and detect problematic items. Internal consistency reliability was assessed using McDonald's &#x3c9;, and test-retest reliability over a two-week interval was evaluated using Pearson correlations. Face validity indicated overall clarity and adequacy. Four items were excluded and recombined, two items were added, and nine items were rewritten based on experts' feedback. After pilot testing, the questionnaire overall demonstrated high internal consistency (&#x3c9;&#x2009;=&#x2009;0.9). Items were reviewed based on factor loadings, item redundancy, theoretical relevance, and their contribution to scale-level internal consistency. EFA suggested a two-factor structure with good internal consistency (&#x3c9;&#x2009;=&#x2009;0.88 and &#x3c9;&#x2009;=&#x2009;0.82) and test-retest reliability (ICC&#x2009;=&#x2009;0.91, 95% CI [0.80, 0.96]). Three items with weak loadings were excluded. The final version consisted of 25 items and 2 subscales. The SCSEQ showed favorable preliminary psychometric properties.

A graduate-level master's gross anatomy course was recently rewritten to enhance delivery of its content in a manner that supports student learning. End-of-course evaluations from all 88 students showed highly favorable ratings for the curricular change, but a more detailed analysis is critical to determine whether a safe learning environment had been established. A safe learning environment is established where the curriculum is appropriately challenging, faculty are supportive, and students feel they belong. A measure of the success of a new course is typically the overall performance of the cohort, in addition to the student evaluations. One limitation of this is the use of Likert scores in student evaluations, which are universally devoid of nuanced information. In most instances, students are offered an opportunity to provide written feedback, and beyond a superficial read-through of these comments, they are not typically analyzed for more information or with purpose. An AI-based text analysis tool was used to undertake the underutilized technique of scoring the comments (sentiment data) to further investigate; the data were incorporated in this instance and highlighted several important categories that students deemed necessary to comment on, including: faculty availability, humor and knowledge, effective group dynamics, and a sense of belonging, among others. Based on the assessments of categories in the student-written feedback, a safe learning environment was achieved.

With the widespread adoption of large language models such as ChatGPT, distinguishing AI-generated text from human-written content has become increasingly challenging. Existing detection methods often rely solely on semantic representations and exhibit limited robustness, particularly when texts are paraphrased or rewritten. This study proposes an integrated detection framework that combines contextual semantic embeddings with auxiliary surface-level features, including pronunciation-related textual cues and handcrafted statistical descriptors. Specifically, a RoBERTa encoder is employed to capture deep contextual semantics, while a convolutional neural network aggregates multi-scale representations. In parallel, a set of text-derived structural, lexical, and readability features-serving as proxies for phonetic and stylistic regularities-are incorporated to enrich the representation space. Rather than introducing a fundamentally new detection paradigm, the proposed approach emphasizes feature-level fusion and systematic empirical evaluation. Experiments on the HAGTC dataset and a ChatGPT-written abstract dataset show that the proposed RoBERTa-CNN framework consistently outperforms several strong baselines in terms of accuracy and F1 score. Notably, the model demonstrates improved robustness in detecting rewritten AI-generated texts. Ablation studies further confirm that integrating multiple feature types significantly enhances detection performance. These results indicate that combining contextual representations with auxiliary surface features offers a practical and effective direction for AI-generated text detection.