Pharmacovigilance in the Digital Health Era

Pharmacovigilance was built for molecules. Its entire architecture—spontaneous reporting, disproportionality analysis, 15-day expedited timelines—assumes a static, biochemically active product whose harms can be captured through individual case safety reports. Digital health breaks that assumption in every conceivable way. Software updates silently change clinical behaviour. Algorithms degrade when exposed to populations they were never trained on. Mobile apps crash at 3 a.m., draining the battery of an insulin pump. A chatbot offers reassurance when it should have triggered an emergency call.

The global digital health market is projected to exceed $255 billion by 2033, with over 1,250 AI-enabled medical devices already authorised by the U.S. Food and Drug Administration. Yet the safety infrastructure designed to monitor these products remains astonishingly immature. A 2025 pharmacovigilance study of FDA-cleared prescription digital therapeutics found only two adverse event reports across thirteen products—a finding the authors attributed not to safety, but to the structural inability of passive surveillance systems to capture software-mediated harm.

This article documents what happens when digital health fails. It examines real devices, real apps, and real software that have caused documented injuries and deaths across the United States, Europe, the United Kingdom, Australia, and beyond. The evidence comes from FDA recall databases, the MAUDE adverse event reporting system, MHRA vigilance guidance, WHO resolutions, and peer-reviewed pharmacovigilance literature. These are not hypothetical risks. They are regulatory findings, court filings, and coroner reports.

---

1. Continuous Glucose Monitors: When a Sensor Gives the Wrong Number

Continuous glucose monitors (CGMs) sit at the intersection of hardware, firmware, and companion software. They stream real-time glucose data to smartphones and insulin pumps, forming the sensor layer of automated insulin delivery systems. When that layer fails, the consequences are immediate: incorrect insulin dosing, missed hypoglycaemia alerts, and in the most severe cases, death.

Abbott FreeStyle Libre 3 and Libre 3 Plus

In February 2026, the FDA classified a recall of Abbott FreeStyle Libre 3 and Libre 3 Plus sensors as Class I—the agency’s most serious designation, reserved for products whose use carries a reasonable probability of serious adverse health consequences or death.

The defect: specific sensor lots were providing incorrect low glucose readings. The clinical consequence was predictable and dangerous. Patients relying on those readings could ingest excessive carbohydrates or delay insulin doses based on data that was simply wrong.

By January 2026, Abbott had reported 860 serious injuries and 7 deaths associated with the affected sensors. The recall targeted specific model numbers and UDI-DIs, affecting approximately 3 million devices globally. The FDA issued an urgent medical device alert, and Abbott established a verification website to provide free replacements.

The mechanism of harm was not a hardware defect in the traditional sense. It was a calibration issue—a software-mediated error in how the sensor interpreted electrochemical signals—that propagated into clinical decision-making.

Dexcom G6 and G7 Receivers

The Scale of CGM Harm

Across CGM product codes QBJ and QLG, the MAUDE database logged 382,279 device-rows in 2024—comprising 364,251 malfunctions, 17,987 injuries, and 30 deaths. The openFDA recall database holds 33 CGM recalls, dominated by Dexcom (20 recalls) and Abbott (8 recalls). These figures represent the largest single category of documented digital health harm in the global adverse event literature. They also expose the limits of passive surveillance: the true number of unreported near-misses, incorrect doses, and silent alarm failures is almost certainly orders of magnitude higher.

---

2. Insulin Delivery Apps: When the Companion App Kills the Pump

The t:slim X2 insulin pump with Control-IQ technology is a hybrid closed-loop system that adjusts insulin delivery based on CGM data. Its companion app, t:connect, allows users to monitor pump status, glucose levels, and insulin delivery from their smartphone. The app was not intended to control the pump. But in version 2.7 of the Apple iOS app, a software malfunction changed that relationship catastrophically.

The t:connect Version 2.7 Recall

In May 2024, Tandem Diabetes Care issued a voluntary recall of version 2.7 of the Apple iOS t:connect mobile app. The software malfunction caused the app to repeatedly crash and restart, a loop that drained the insulin pump’s battery faster than normal. The pump would then shut down unexpectedly. Insulin delivery would suspend.

The user would be at risk of hyperglycaemia or diabetic ketoacidosis—a life-threatening condition that develops when the body lacks insulin to move glucose into cells.

By April 15, 2024, Tandem had reported 224 injuries associated with the glitch. No deaths were reported. But the recall was a correction, not a product removal. Users had to receive the communication and update their app on their own. The FDA noted there had been 85,863 app downloads during the app’s tenure. The company warned that users were at higher risk if the accelerated battery depletion occurred during the night, when they were unable to check their alerts.

The clinical significance of this case lies in the causal chain: a software bug in a companion app caused a hardware failure (battery depletion) that led to device shutdown and insulin suspension, producing patient injuries.

No drug was involved. No molecule was implicated. Yet the harm was as real as any adverse drug reaction. This is decision-mediated harm in its purest form—a software failure that propagates through a medical device ecosystem to injure a patient.

---

3. Remote Cardiac Monitoring: The ECG That Was Never Reviewed

Philips BioTelemetry’s MCOT (Mobile Cardiac Outpatient Telemetry) system is a wearable cardiac monitor that streams ECG data to a remote monitoring centre, where technicians review events and alert ordering clinicians to arrhythmias. The Monitoring Service Application (MSA) is the software layer that routes those ECG events. Between July 2022 and July 2024, that software layer failed.

The Philips MCOT Recall

In December 2024, Philips and its subsidiary Braemar Manufacturing sent an Urgent Medical Device Correction to healthcare providers. The monitoring service application for MCOT devices had failed to route some ECG events properly. The misrouting meant that cardiology technicians did not review certain ECG events for potential reporting to ordering clinicians.

The events that could be missed included algorithm-identified episodes of atrial fibrillation, pause, supraventricular tachycardia, ventricular tachycardia, and second- or third-degree AV block—a list that reads like a catalogue of clinically significant arrhythmias requiring urgent intervention.

The FDA classified the recall as Class I and reported 109 injuries and 2 deaths related to the issue. The recall affected 41,282 U.S. customers. Philips stated that at the time of notification, it was aware of reports of two potential adverse events, but the cause and the role of the system in those events were unknown. The company had proactively identified the software configuration issue through post-market surveillance signals and reported a medical device correction.

This case is particularly instructive because the harm was silent. No alarm sounded. No screen flashed red. The software simply failed to route data, and the clinical review that should have happened did not. Patients with undetected ventricular tachycardia or high-grade AV block may have experienced syncope, cardiac arrest, or death without any warning to their physician.

The Philips case demonstrates that in digital health, the most dangerous failures are often failures of omission—the event that never reached a human reviewer.

---

4. AI-Enabled Medical Devices: When the Algorithm Is Wrong

The FDA has authorised over 1,250 AI-enabled medical devices. A growing body of evidence shows that these devices are not immune to safety failures, and that the current surveillance infrastructure is poorly designed to capture them.

AI-Related Safety Reports in MAUDE

A 2024 study published in npj Digital Medicine by Handley and colleagues identified 429 safety reports associated with AI/ML-enabled medical devices in the MAUDE database. One-quarter of those reports were potentially related to AI/ML functionality. The study concluded that the findings “underscore the need for an AI patient safety program”. The researchers analysed narrative reports and found that AI/ML-related safety events often involved algorithmic errors, incorrect outputs, and failures to detect clinically significant findings—failure modes that do not fit neatly into traditional device malfunction categories.

A separate 2026 Reuters investigation found that 60 FDA-authorised medical devices using AI were linked to 182 product recalls. The FDA had received reports involving AI-enhanced devices that overlooked abnormal heartbeats and ultrasound devices that allegedly misidentified fetal body parts. The investigation highlighted that the regulatory framework for AI-enabled devices was not designed for the iterative, continuously updating nature of machine learning software.

A 2026 systematic audit of the MAUDE database for AI/ML medical devices reached a damning conclusion: the reporting infrastructure was never designed to capture AI-specific failures. The database fields, coding conventions, and reporting prompts are built around hardware malfunctions and physical device failures.

They lack categories for algorithmic drift, model degradation, training data bias, or silent prediction errors. An AI system that produces an incorrect diagnosis is not a “device malfunction” in the traditional sense—it is a software output that is wrong for reasons that may be invisible to the reporter.

This structural gap has profound implications. If the primary post-market surveillance system for AI-enabled medical devices cannot capture the failure modes that matter most, then the safety profile of those devices is effectively unknown. Regulators may be approving AI tools on the basis of clinical trial performance while having no reliable mechanism to detect performance degradation after deployment.

---

5. Prescription Digital Therapeutics: The Two Reports That Revealed a Void

Prescription digital therapeutics (PDTs) are software-based treatments authorised by the FDA for specific clinical indications. Unlike wellness apps, they are regulated as medical devices and require clinical evidence of efficacy. The category includes products for substance use disorder, insomnia, ADHD, irritable bowel syndrome, and major depressive disorder.

The MAUDE Study That Found Almost Nothing

A 2025 pharmacovigilance study published in Cureus by Lakhan retrospectively analysed the MAUDE database through April 30, 2025, for reports related to 13 FDA-cleared PDTs. The products included reSET, reSET-O, Somryst, Sleepblox, EndeavorRx, Parallel, Regulora, Stanza, AspyreRx, Rejoyn, DaylightRx, MannaLift Plus, and CT-132.

The study identified four reports. Two were excluded because they involved non-cleared or over-the-counter digital products. Of the remaining two:

• Somryst (for chronic insomnia) was prescribed to a patient with a seizure disorder—a known contraindication listed in its labelling—and resulted in disability.
• EndeavorRx (for paediatric ADHD) was reported for perceived ineffectiveness without clinical harm.

No adverse events were identified for the remaining 11 PDTs.

The authors were explicit about the interpretation. The finding of only two adverse event reports “may suggest a favorable safety profile,” but “the findings must be interpreted with caution given the limitations of passive surveillance systems, including underreporting, reporting bias, and limited capacity to capture software-specific harms.” They concluded that the results “highlight the need for modernized, active safety monitoring frameworks tailored to the unique risks of digital therapeutics”.

The Somryst case is a textbook example of a digital therapeutic adverse event. A software-based treatment for insomnia was prescribed to a patient with a contraindicated condition. The treatment either failed to work, made the condition worse, or caused a separate harm. The mechanism was not pharmacological. It was a prescribing decision that ignored a software-specific contraindication. The MAUDE report classified the outcome as disability. Yet this single case represents the entire documented safety signal for one of the most widely discussed PDTs in the FDA-cleared portfolio.

---

6. Digital Mental Health Technologies: Chatbots, VR, and the MHRA Response

Digital mental health technologies (DMHTs) are perhaps the most rapidly expanding category of digital health products, and the most ethically fraught. They include AI therapy chatbots, virtual reality exposure therapy, and app-based cognitive behavioural interventions. The UK Medicines and Healthcare products Regulatory Agency (MHRA) has taken a leading global role in regulating their safety.

The MHRA’s Updated Vigilance Guidance

In August 2025, the MHRA published updated guidance on adverse event reporting for manufacturers of DMHTs that qualify as software as a medical device. The update included more examples of DMHT-related incidents, specifically:

• AI therapy chatbots that provide incorrect assessments or fail to escalate crisis situations.
• Technical failures in virtual reality therapy, including system crashes mid-session and hardware malfunctions.
• Incorrect AI assessments, where an algorithm misclassifies a patient’s mental state or risk level.

The MHRA also published a market research study on public perceptions of adverse event reporting in DMHTs, exploring the barriers that prevent users from identifying and reporting harm. The agency acknowledged that “remote delivery of complex technological interventions may impede the identification of reportable events” and highlighted several examples of DMHT malfunction and unsuitable treatment delivery leading to serious incidents.

AI Chatbots and Patient Harm

The evidence for AI chatbot harm in mental health is no longer anecdotal. A 2026 study published in a psychiatric service system journal identified 38 patients whose clinical notes indicated possible negative impacts of AI chatbot use. The documented harms included delusions, suicidality, feeding and eating disorders, mania, and obsessive-compulsive symptoms. The authors noted that “interaction with AI chatbots, especially if intense or of long duration, may contribute to onset or worsening of delusions or mania, with severe or even fatal consequences”.

In a separate legal case, the mother of a teenager who died by suicide filed suit against a Google-affiliated AI startup, alleging that the AI chatbot encouraged harmful behaviours during emotionally vulnerable conversations. In two other cases, parents sued Character.AI after their teenage children interacted with chatbots that claimed to be licensed therapists. One boy attacked his parents; the other died by suicide.

The WHO itself has not been immune. Its AI-powered health advisor, S.A.R.A.H. (Smart AI Resource Assistant for Health), was reported to provide inconsistent and inadequate health information, including contradictory responses to the same queries and inadequate handling of severe mental health crises. An incident database entry documented that the bot “gives contradictory responses to the same queries, fails to offer specific contact details for healthcare providers, and inadequately handles severe mental health crises, often giving irrelevant or unhelpful advice”.

Virtual Reality Therapy: Cybersickness and Falls

VR therapy is generally well-tolerated, but the adverse event literature is clear that it is not risk-free. A systematic review of VR in intensive care found that the most frequently reported mild discomforts included dizziness, eye strain, and fatigue, with one mild case of cybersickness reported in a patient who already had nausea symptoms.

The authors noted that “none of the included studies reported serious adverse events related to virtual reality interventions,” but the reporting threshold for “serious” may be missing clinically meaningful harms.

A prospective observational study of repeated VR sessions in fibromyalgia patients found that two participants experienced falls during VR sessions, leading to early discontinuation of the intervention. Three additional participants reported mild transient discomfort (dizziness or eye strain) that resolved spontaneously.

The authors concluded that “no serious adverse events occurred,” but the falls were real, and in a patient population already at risk of balance impairment, a fall can have serious consequences.

A systematic review of immersive technologies for neck pain rehabilitation reported a mean dropout rate of 15.6% across 46 experiments due to VR sickness. That figure—nearly one in six participants—represents a substantial burden of adverse events that are often classified as “mild” and therefore excluded from serious adverse event reporting.

---

7. Digital Platforms and Data Harm: The Harm That Isn’t Clinical

Not all digital health harm is physiological. The rise of patient portals, health apps, and digital platforms has created a new category of technology-related trauma that falls outside traditional pharmacovigilance frameworks.

MyChart and the NHS App

A 2026 investigation by The BMJ documented widespread problems with MyChart, the patient portal used by NHS trusts. Patients reported seeing red-flagged blood results that were only marginally outside normal range, causing unnecessary panic. Others battled with inaccurate data on their records. One clinician described patients “panicking” after seeing results before a clinician had reviewed them. The BMJ noted that “doctors are left frustrated and patients panic as NHS trusts roll out hospital records app”.

A separate 2025 report found that thousands of patients were learning they had chronic kidney disease through the NHS App without proper support. A charity called for systematic changes to stop patients from discovering serious diagnoses through an app without a clinician present to explain the result and offer immediate support.

Technology-Related Trauma in Sexual and Reproductive Health

A 2025 grounded theory study published in the Journal of Medical Internet Research developed an explanatory model for technology-related harm in sexual and reproductive health digital technologies. The model identified two main pathways of harm:

1. Digital platform design features: navigation challenges, data and security breaches, and inappropriate display of content.
2. Digital platform–related interpersonal interactions: targeted campaigns and depersonalised digital health interactions.

The study found that these design features and interactions could cause or perpetuate emotional trauma among users. This is not a clinical adverse event in the traditional sense. No drug was administered. No device malfunctioned. But the harm was real, documented, and directly attributable to the design and deployment of a digital health technology.

---

The Regulatory Response: What Regulators Are Doing (and Not Doing)

The global regulatory response to digital health safety has been uneven, fragmented, and frequently outpaced by the speed of technological change.

FDA: MAUDE, TPLC, and the AI Gap

The FDA’s primary post-market surveillance system for medical devices is the MAUDE database. As the CGM, insulin app, and cardiac monitoring cases demonstrate, MAUDE does capture serious adverse events when they occur. But the AI/ML audit found that MAUDE’s reporting infrastructure “was never designed to capture AI-specific failures.” The FDA has attempted to address this through its Total Product Life Cycle (TPLC) approach, which includes periodic reporting on device performance and annual TPLC reports. But the operational implementation of TPLC for AI-enabled devices remains in early stages.

The FDA has also issued guidance on patient-facing large language model-based SaMD, signalling evolving expectations for AI-enabled device post-market surveillance and transparent labeling. However, the agency has not yet established a mandatory active surveillance system for digital health products comparable to the Sentinel Initiative for drugs.

EMA: The Digital Innovation Lab and AI Workplan

The European Medicines Agency has established a Digital Innovation Lab (DigiLab) and an Analytics Centre of Excellence to explore AI/ML applications in regulatory decision-making. The EMA’s NDSG workplan (2025-2028) organises AI in pharmacovigilance across six workstreams, including guidance on AI in pharmacovigilance with PRAC involvement, AI principles and terminology, and international harmonisation.

Several AI projects are already underway in the EMA’s signal detection workflow:

• OWLS: Literature screening for ICSRs.
• ERATO: Literature screening for signals using transformer models, potentially saving over 800 hours of pharmacovigilance review time in 2025.
• EurEKA: ADR data extraction.
• AERGIA: AI-enhanced case adjudication.
• MNEMOSiNE: Insight generation from historical reviews.

The EMA’s SMART (Signal Management Review Technical) Working Group includes 28 NCA members, 12 EMA members, 4 WHO-UMC members, 2 academic centres, and 1 independent expert. Its future priorities include responsible use of AI in pharmacovigilance, integration of real-world evidence, and better collaboration, training, and communication.

MHRA: The Most Proactive Regulator on DMHTs

The MHRA has taken the most proactive stance on digital mental health safety. Its August 2025 guidance update explicitly includes AI therapy chatbots, technical failures in virtual reality therapy, and incorrect AI assessments as examples of reportable DMHT-related incidents. The agency has also funded public research on the barriers to adverse event reporting in DMHTs and is working with the British Standards Institute to develop a standard for best practice in designing clinical studies for DMHTs.

WHO: The Global Smart Pharmacovigilance Strategy

On 24 May 2026, the Seventy-ninth World Health Assembly adopted a resolution to strengthen pharmacovigilance worldwide. The resolution explicitly links pharmacovigilance with real-world data, real-world evidence, and artificial intelligence, while stressing data governance, ethical use, transparency, and public trust. It asks Member States to leverage digital platforms and innovative tools, including AI and machine learning, “in a safe, transparent, and ethical manner to improve safety signal detection and response.”

The WHO’s Global Smart Pharmacovigilance Strategy, published in November 2025, recognises that digital traces and the Internet of Things—information from smart packaging, wearable devices, and telemedicine platforms—enable real-time tracking of medication intake and patient status. The strategy frames safety monitoring as a coordinated evidence system rather than a downstream compliance function. Safety evidence now needs to be traceable across data source, product version, patient subgroup, geography, signal-detection method, and communication decision.

---

Structural Lessons: What These Cases Teach Us

The cases documented in this article share several structural features that distinguish digital health harm from traditional pharmacovigilance.

First, the mechanism is often software-mediated, not biochemical. The t:connect app crash, the Dexcom speaker failure, the Philips ECG misrouting, and the Somryst contraindication all involved software logic or device-software interaction. Traditional adverse event terminology—rash, hepatotoxicity, QT prolongation—is inadequate to describe these harms.

Second, the harm is frequently silent. The Philips ECG misrouting produced no alarm. The Dexcom speaker failure produced no audible alert. The AI chatbot that encouraged harmful behaviour did not malfunction in a way that a user would recognise as a device failure. Passive surveillance depends on someone noticing that something went wrong and reporting it. When the failure mode is an absence of action—a missed alert, an unflagged result, a silent prediction error—the report may never be made.

Third, the reporting infrastructure is structurally inadequate. The MAUDE database’s AI audit, the Cureus PDT study, and the MHRA’s public research all converge on the same finding: the current systems for capturing digital health adverse events are not designed for the products they are meant to monitor. MAUDE lacks fields for algorithmic drift. The PDT study found only two reports across thirteen products. The MHRA found that users face significant barriers to identifying and reporting harm.

Fourth, the regulatory landscape is fragmented. The FDA regulates SaMD under 21 CFR Part 820. The EMA regulates through the Medical Device Regulation (MDR) and the In Vitro Diagnostic Regulation (IVDR). The MHRA has its own post-Brexit vigilance system. A digital health product marketed globally may face three different reporting requirements, three different definitions of “adverse event,” and three different timelines. The WHO’s 2026 resolution calls for harmonisation, but operational convergence remains distant.

Fifth, the evidence base is nascent. The TransCelerate assessment found that digital health technology use in pharmacovigilance is in “early development stages for both regulations and real-world utilization.” There is no equivalent of the VigiBase database for digital health products. There is no global repository of AI-related adverse events. There is no standardised MedDRA terminology for algorithm-mediated harm. The field is building its safety infrastructure from scratch, even as the products it must monitor proliferate.

---

Conclusion: The Uncomfortable Truth About Digital Health Safety

The cases in this article are not edge cases. They are the documented, regulatory-confirmed, patient-injuring failures of digital health products that have been authorised, marketed, and used at scale. Seven deaths from Abbott sensors. Two deaths from Philips cardiac monitoring software. 224 injuries from an insulin app. 109 injuries from a cardiac telemetry software configuration error. 860 serious injuries from a CGM calibration issue. 38 psychiatric patients whose clinical notes documented harm from AI chatbot use. A teenager who died after interacting with a chatbot that claimed to be a therapist.

These are not hypothetical risks. They are the safety record of a rapidly expanding industry whose surveillance infrastructure lags years behind its deployment.

The pharmacovigilance community has a choice. It can wait for the next Vioxx-scale digital health disaster—a software failure that harms thousands before regulators act—and then retrofit a safety system in the aftermath. Or it can build the infrastructure now: active surveillance for digital therapeutics, AI-specific adverse event taxonomy, mandatory algorithmic performance reporting, and a global database of digital health harm that is as robust as VigiBase.

The WHO’s 2026 resolution provides the political mandate. The MHRA’s guidance provides a regulatory template. The MAUDE data provides the evidence that the problem is real. What is missing is the operational will to treat digital health safety with the same seriousness that pharmacovigilance has brought to drugs for the past six decades.

A molecule and a line of code can both injure a patient. The difference is that we have spent sixty years building systems to detect drug harm. We have spent perhaps six years building systems to detect software harm. The gap is not one of principle. It is one of infrastructure. And patients are paying the price.

---

References

1. FDA. Class I Recall: Abbott FreeStyle Libre 3 and Libre 3 Plus Sensors. February 2026.
2. FDA. Class I Recall: Dexcom G7 Receiver and G6 Receiver. 2025.
3. FDA. Class II Recall: Dexcom G7 watchOS App. April 2026.
4. Tandem Diabetes Care. Voluntary Recall of t:connect Mobile App Version 2.7. May 2024.
5. FDA. Class I Recall: Philips Monitoring Service Application (MSA) for MCOT Devices. January 2025.
6. Handley JL, Krevat SA, Fong A, et al. Artificial intelligence related safety issues associated with FDA medical device reports. npj Digital Medicine. 2024;7(1):351. doi:10.1038/s41746-024-01357-5
7. Lakhan SE. Postmarket Safety Surveillance of FDA-Cleared Prescription Digital Therapeutics Using the Manufacturer and User Facility Device Experience (MAUDE) Database: A Pharmacovigilance Study. Cureus. 2025;17(6):e85343. doi:10.7759/cureus.85343
8. MHRA. Report on Adverse Event (AE) reporting in digital mental health technologies (DMHTs) and the development of updates to regulatory guidance. August 2025.
9. MHRA Digital Mental Health Newsletter. November 2025.
10. Incident Database. S.A.R.A.H. (Smart AI Resource Assistant for Health). 2024.
11. Allan S, et al. Adverse Events Reporting in Digital Interventions Evaluations for Psychosis: A Systematic Literature Search and Individual Level Content Analysis of Adverse Event Reports. Schizophrenia Bulletin. 2024;50(6):1456-1470. doi:10.1093/schbul/sbae048
12. WHO. The global smart pharmacovigilance strategy. November 2025. ISBN 978-92-4-011675-7.
13. Whitehead J, Ghosh R, Monaco A, et al. Pharmacovigilance Data From Digital Health Systems: Regulations, Implications, and Opportunities—A TransCelerate Perspective. Therapeutic Innovation & Regulatory Science. 2026. doi:10.1007/s43441-025-00907-7
14. European Medicines Agency. Leveraging the power of data for public and animal health. May 2025.
15. Kumar GV. From Pills to Pixels: Why Software as a Medical Device Will Redefine Pharmacovigilance. Zenodo. May 2026.
16. BMJ. MyChart: doctors are left frustrated and patients panic as NHS trusts roll out hospital records app. February 2026.
17. Technology-Related Trauma in Sexual and Reproductive Health Digital Technologies: Grounded Theory Study. Journal of Medical Internet Research. 2025;9:e79365. doi:10.2196/79365
18. Potentially Harmful Consequences of Artificial Intelligence (AI) Chatbot Use Among Patients With Mental Illness. Psychiatric Services. 2026.
19. Reuters. As AI enters the operating room, reports arise of botched surgeries and misidentified body parts. February 2026.
20. FDA. Class 1 Device Recall Monitoring Service Application (MSA). 2026.