All traces

Half of included studies had identified risk of bias and clinical adoption remains limited, requiring multicenter prospective validation and interpretable frameworks before bedside use.

Visual HRCT scoring remains reader-dependent and AI outputs lack prospective multicenter validation and protocol harmonization needed to serve as treatment-triggering biomarkers.

Chatbots sometimes confused similar congenital anomalies and provided less detailed anatomical descriptions in complex tumor cases, requiring caution and verification by qualified professionals before clinical use.

Participants identified trust as the central barrier to using an AI companion, driven by privacy concerns and vulnerabilities specific to schizophrenia.

Students with medium and low baseline history-taking proficiency showed relatively limited score improvements from LLM-VSP self-practice, with practice frequency alone not independently predicting final performance.

Most hand surgery AI tools are deployed in unaudited workflows and rarely remeasured after release after testing only on training-like data, leaving the hand surgeon accountable for patient outcomes shaped by opaque models.

AI tools translating radiology reports into plain language can produce translation errors that compromise comprehension and safety, causing reports to be graded unsafe and warrant withholding from patients.

In medical education, the same AI tools may reduce training safety by quietly reshaping how future clinicians think and act, functioning as a wedge that undermines training.

Machine learning models for autism spectrum disorder prediction that use data augmentation and feature selection have limited external validation and inadequate evaluation frameworks, reducing confidence in reported performance improvements and model generalizability.

AI-based assessment introduces distinct validity threats across scoring, generalisation, extrapolation and implications, including contamination, instability, inequities, automation bias and deskilling when used for consequential learner progression decisions.

The auxiliary model cannot replace pathological and molecular diagnosis and relies on tissue-derived IDH and Ki-67 markers, with development limited to a single-center retrospective cohort of 400 patients and internal validation only.

Accuracy and clinical utility of AI-assisted non-contact neonatal heart rate monitoring remain unvalidated for routine implementation pending future multicenter studies.

The announcement raised concerns that in-progress work stored in OpenAI's Codex model was potentially visible to OpenAI, with OpenAI stating it could not rule out that the pair's product use helped improve its models, alongside a prior disclosure of agents hacking into Hugging Face.

Even when incorrect, ChatGPT and Claude produced thorough justifications, creating risk of convincing but wrong guideline-based information, while Gemini showed formatting deviations.

Educators and students report concern about misuse of GenAI in education and a lack of classroom and institutional preparedness to manage it in the writing process.