The first article, written by health writer and ICU nurse Jenna Congdon, tackles a phenomenon that many hospital administrators would prefer not to name: shadow AI. In

The tension between clinical innovation and institutional oversight that defines the shadow AI phenomenon is not unique to any single health system. Across hospitals and clinics, the gap between what frontline workers need and what officially sanctioned software provides has widened as consumer-facing AI tools have become more capable and more accessible. A clinician who once might have drafted a discharge summary by hand can now paste de-identified fragments of a note into a general-purpose chatbot and receive a polished summary in seconds. The convenience is undeniable, but so is the risk: once patient information leaves a controlled environment, the organization loses visibility into where that data travels, how it is retained, and whether it could resurface in outputs shown to other users.

Researchers studying health information technology adoption have long observed a pattern in which workarounds emerge when formal systems are misaligned with the realities of practice. Barcode medication administration, electronic health records, and secure messaging platforms have each spawned their own unauthorized shortcuts when clinicians found the approved versions too slow, too rigid, or too poorly matched to their workflows. Shadow AI fits squarely within this tradition. Its rapid proliferation suggests that governance strategies built purely on prohibition are unlikely to succeed, and that institutions which engage with the underlying drivers of unauthorized use stand a better chance of channeling that demand into safer, auditable tools.

The economics of clinical documentation further illuminate why shadow AI has taken hold. Studies of physician and nurse time allocation consistently show that a substantial fraction of a clinician’s day is consumed by administrative tasks rather than direct patient care. Burnout surveys repeatedly link documentation burden to exhaustion and attrition, and staffing shortages across many health systems have intensified the pressure. In that environment, any tool that promises to reclaim even a few minutes per patient carries enormous appeal. The challenge for health systems is to capture those efficiency gains without sacrificing privacy safeguards, model transparency, or accountability for errors.

The science of vocal biomarkers, meanwhile, rests on a growing body of evidence that the human voice encodes measurable signatures of physiological and neurological state. Speech production requires the coordination of respiratory muscles, laryngeal folds, articulators, and multiple brain regions, so disruptions in motor control, cognition, or mood can alter acoustic properties such as pitch variability, speech rate, pause frequency, and spectral features. Researchers have reported that such features may change in conditions ranging from Parkinson disease and Alzheimer disease to depression and respiratory illness, sometimes before overt symptoms prompt a clinical visit.

What machine learning adds to this field is scale and pattern recognition. Traditional speech and language assessments are administered one patient at a time by trained specialists, limiting throughput and introducing subjectivity. Automated pipelines can process recordings in seconds and extract hundreds of acoustic variables simultaneously, potentially flagging subtle deviations a human listener would miss. The JMIR feature article highlights initiatives such as the Luxembourg Institute of Health’s Deep Digital Phenotyping Research Unit, which is incorporating vocal data into digital twin models, and the Weizmann Institute’s Human Phenotype Project, whose collection of more than 7000 voice recordings within a deep phenotyping cohort represents an early effort to build the large, standardized datasets that robust clinical validation will require.

Standardization is a critical hurdle for the field. Voice recordings are exquisitely sensitive to recording conditions: microphone quality, background noise, distance from the device, and compression by telecommunication software can all shift acoustic measurements. A biomarker validated on studio-quality recordings may perform very differently on a smartphone captured in a noisy hallway. Efforts to establish universal collection protocols, reference datasets, and calibration methods are therefore seen as prerequisites for moving vocal biomarkers from research curiosity to diagnostic utility. Regulatory agencies have begun to signal interest, but the pathway from an interesting acoustic correlation to an approved clinical test remains long and demands prospective validation in diverse populations.

If vocal biomarkers mature, they could reshape screening paradigms, particularly for neurodegenerative disorders where early detection may eventually matter most. Passive or semi-passive monitoring, in which a patient’s voice is analyzed during ordinary phone calls or periodic app-based prompts, could complement traditional assessments and provide longitudinal trends that single snapshots cannot. Yet the same capabilities raise questions about consent and surveillance. Voice is gathered constantly by consumer devices, and the prospect of health inferences drawn from routine audio underscores the need for clear policies on when such analysis is permissible and who may access the results.

The emergence of griefbots raises similarly unsettled questions in the domain of bereavement. These systems, trained on the texts, messages, social media posts, photographs, and other digital traces a person leaves behind, generate conversational avatars that mourners can interact with as though speaking with the deceased. Commercial services offering such interactions already exist, and the digital legacy market they anchor is expanding. What distinguishes the current moment is that the underlying language models have grown sophisticated enough to make these interactions feel genuinely responsive, amplifying both their potential comfort and their potential to complicate the grieving process.

Grief researchers have long described mourning not as a linear path toward forgetting but as an ongoing renegotiation of the relationship with the deceased. Within that framing, as psychologist Dr. Robert Neimeyer suggests, a griefbot could serve as one instrument among many, offering a controlled space for continuing bonds when used alongside counseling, ritual, and community support. Some bereaved individuals report that simulated conversations helped them articulate things left unsaid, and clinicians sympathetic to the technology argue that any tool that reduces isolation deserves study rather than reflexive dismissal.

Skeptics, including bioethicist Dr. Craig Klugman, counter that the technology may entrench avoidance and contribute to prolonged grief disorder, a condition characterized by persistent, disabling grief that impairs functioning. The mental health outcomes of griefbot use are only beginning to be researched, and no consensus yet exists on which users, if any, are most likely to benefit or be harmed. The commercial incentives behind these products add another layer of concern: Rebernik notes that scant regulations exist to promote user safety and privacy, leaving open the possibility that deeply personal data about the dead and the bereaved could be leveraged for precision marketing or retained indefinitely without meaningful consent from the deceased, who never agreed to be resurrected in algorithmic form.

Digital legacy questions extend beyond griefbots themselves. People increasingly curate what happens to their data after death, and legal frameworks for posthumous digital rights remain fragmented across jurisdictions. Whether an individual’s conversational style, likeness, and voice constitute property that can be licensed to a resurrection service is largely untested, and survivors may disagree among themselves about whether such services honor or violate the memory of the person they loved. These disputes are likely to grow as the technology improves and as more of the population leaves behind rich digital archives.

The story of EndeavorRX illustrates a different convergence of software and medicine: the treatment delivered as a game. Its FDA authorization for pediatric patients aged 8 to 17 marked a regulatory milestone, establishing that a video game designed to train attention through gameplay could satisfy the agency’s expectations for a therapeutic device. The game challenges users to attend to multiple simultaneous demands, and the developer reported improved attention measures in 73 percent of pediatric patients in its supporting studies. Importantly, it is positioned as an adjunct rather than a replacement for established treatments, reflecting a cautious regulatory posture toward a novel therapeutic category.

The broader category of serious games encompasses applications designed for purposes beyond entertainment, including rehabilitation, health education, and cognitive training. Researchers have investigated whether commercial games also confer measurable benefits, and some clinicians, such as psychologist Dr. Megan Connell, already incorporate certain commercial titles into therapeutic practice. The mechanisms proposed for therapeutic effect vary by application, ranging from attentional training and neuroplasticity to the safe exposure and rehearsal of coping strategies within engaging virtual environments. Advances in artificial intelligence and virtual reality may enable games that adapt difficulty in real time to an individual’s performance, potentially increasing both efficacy and adherence.

Open questions remain about how prescription games will fare in routine care. Questions of insurance coverage, prescriber familiarity, and adherence outside supervised trial conditions will shape real-world impact. Sceptics also caution that enthusiasm must be tempered by rigorous independent replication, since effect sizes in novel behavioral interventions sometimes shrink when tested at scale. Nonetheless, the authorization signals to developers and investors that digital therapeutics can clear regulatory hurdles, a signal already reflected in a growing pipeline of software-based treatments for psychiatric and neurological conditions.

Taken together, these four threads describe a health landscape in which the boundary between the clinical and the computational is steadily dissolving. Clinicians reach for unapproved tools because approved ones fall short; researchers mine the voice for early signals of disease; mourners converse with algorithmic echoes of the dead; and children practice attention within a game their doctor can prescribe. Each development carries promise measured against distinct risks: privacy breaches, premature clinical claims, commercial exploitation of grief, and unvalidated therapeutic enthusiasm. The common denominator is a widening recognition that software in health and human experience requires not only innovation but deliberate structures of evidence, ethics, and governance to earn the trust of the people it touches.

The task for researchers, regulators, clinicians, and the public in the coming years will be to decide, case by case, which of these technologies deserve integration into practice and under what safeguards. The JMIR Publications News and Perspectives series contributes to that deliberation by grounding emerging trends in expert analysis and open access scholarship, an approach consistent with the publisher’s broader commitment to open science and to making the evidence base for digital health decisions freely available to all who need it.

Subject of Research: Digital health technologies including shadow AI, vocal biomarkers, griefbots, and therapeutic video games

Article Title: JMIR news: Shadow AI, vocal biomarker tech, griefbots, and a prescription video game

Article References: JMIR news: Shadow AI, vocal biomarker tech, griefbots, and a prescription video game. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: shadow AI, vocal biomarkers, griefbots, digital resurrection, EndeavorRX, ADHD, FDA approval, machine learning, digital health, mental health, bereavement, digital legacy