The State of Play

Autonomous preliminary radiology reads have moved from laboratory proof-of-concept into production deployment, but remain concentrated at a handful of high-volume operators and specialist vendors. Systems that generate draft reports for radiologist review show measurable efficiency gains in controlled settings, yet fundamental adoption barriers prevent broad clinical scaling. Radiology Partners processes 40-50 million cases annually through RADPAIR, achieving 2-5 second report turnaround with 12% error reduction. Northwestern Medicine's in-house autonomous reporting system achieves 40% productivity boost without accuracy loss; a multicenter evaluation of MIRA across 1.87 million reports from 42 hospitals found 69% of AI-generated impressions rated equivalent to radiologist originals. These results are credible, but they come from operationally mature organisations and domain-specific models trained on large radiology report corpora. Generic foundation models and multimodal LLMs show substantial capability gaps, with vision-language models exhibiting critical grounding failures (text-only baselines reach within 5.7% accuracy of multimodal models; some models ignore images entirely). Cross-lingual systems achieve 98.7% physician-edit rates, and foundation models exhibit severe age-related diagnostic bias. June 2026 evidence amplifies the leading-edge tension: FDA grants Breakthrough designations to Aidoc First Read and Cognita autonomous report generation systems (2,000+ hospitals, 120M+ cases processed), yet peer-reviewed research documents realism-reliability gaps, information degradation in autonomous rewriting (51% entity erosion), and safety validation gaps (missing clinical evidence 1.39× higher recall hazard). July 2026 deepens the crisis: RadLE 2.0 benchmark exposes critical model calibration failures—AI systems deliver incorrect findings with high confidence while human radiologists outperform on combined accuracy and uncertainty admission. Domain-specific models demonstrate speed gains (89% turnaround reduction) but fail to close clinically relevant consistency gaps (25.2% residual inconsistency in chest X-ray interpretation). Across 691 cleared FDA devices audited, only 1.6% cite randomized clinical trial data; the clearance-to-evidence gap reveals that regulatory authorization does not guarantee clinical safety or efficacy. Human factors evidence emerges: 8-month observational studies document that efficiency gains from autonomous draft-generation workflows mask cognitive fatigue, burnout, and measurable decision-quality degradation among radiologists. August 2026 narrows the maturity pathway further: vision-language models show severe confidence miscalibration on brain MRI (33-46% high-confidence errors across frontier models), and forensic reproducibility audits reveal that even careful benchmark evaluations can yield unreproducible results when original artifacts are audited. Agentic AI evaluation standards remain immature, with a scoping review of 557 studies confirming that process reliability, evidence traceability, and external validity are assessed inconsistently across the field. Positive capability development continues (CARE-X demonstrates SOTA report-generation with clinical validation on rare pathologies), and large-scale operational deployments sustain (298,991-exam health check-up study confirms autonomous AI achieving 87.1% sensitivity, 91.8% specificity at programmatic scale); yet structured pre-deployment assessment frameworks show that radiologist adoption requires alignment between predicted and perceived AI value—a critical governance signal that technical capability alone does not predict clinical adoption. The regulatory and liability landscape remains the binding constraint: until governance, legal, and accountability frameworks evolve—and until the field resolves the calibration, reproducibility, and human-factors crisis—autonomous AI remains unlikely to scale beyond niche high-volume screening applications.

Deployment remains concentrated among operationally mature high-volume operators and specialist vendors, with regulatory acceleration in June 2026 offset by emerging evidence of critical failure modes in July 2026 and August 2026 maturity reassessment. Aidoc's First Read received FDA Breakthrough Device Designation (June 26) for autonomous chest X-ray report drafting, deployed to 2,000+ hospitals processing 120M+ cases; Cognita simultaneously received Breakthrough status for autonomous report generation. Radiology Partners (3,900+ radiologists) runs RADPAIR in production; Mosaic Reporting deployed to thousands of radiologists with real-time draft generation during interpretation powered by Cognita foundation models. Yale New Haven Health System (700k+ annual exams across 16 centers) selected Rad AI for infrastructure-scale autonomous reporting rollout. Real-world deployment metrics: Northwestern Medicine's in-house autonomous reporting system achieves 40% productivity boost; 12-hospital academic health system achieved 15.5% documentation efficiency with autonomous draft reports; India autonomous TB screening increased case notifications 80% (162 confirmed cases); a large health check-up cohort (298,991 exams, 2019-2023) achieved sensitivity 87.1%, specificity 91.8% with autonomous AI in double-reading workflows. A systematic review of 11 studies confirms autonomous chest X-ray triage systems ready for clinical implementation, with 42.3% triage rate and 97.8% sensitivity across 500K+ real-world cases. However, July 2026 evidence documents critical reliability barriers blocking further scaling. RadLE 2.0 benchmark testing whether models appropriately defer to humans when uncertain found that frontier AI systems—including Claude Fable 5, Gemini 3 Pro, and medical-specific models—fail this safety requirement, delivering high-confidence wrong diagnoses and underperforming human radiologists on combined accuracy and calibration metrics. Domain-specific chest X-ray report models demonstrated 89% speed acceleration but 25.2% persistent inconsistency; the speed-to-accuracy transfer gap means automation gains do not yield the clinical consistency improvements promised by developers. European regulatory audits of post-market monitoring (EU AI Act Article 72 compliance) found ~50% of deployed systems lacked adequate monitoring infrastructure. Workforce concerns compound adoption risk: 8-month observational studies document that radiologists using autonomous draft workflows report increased cognitive fatigue, burnout, and measurable decision degradation despite aggregate efficiency metrics. August 2026 adds critical calibration evidence: vision-language models on brain MRI show severe confidence miscalibration with 33-46% high-confidence errors (ECE 0.27-0.40); reproducibility audits of radiology VLM benchmarks reveal protocol deviations undermining published claims. Professional society governance positions sharpen: the American College of Radiology endorses augmentation-over-automation models drawing on proven copilot approaches from ophthalmology, dermatology, and pathology. Pilot adoption analysis reveals a paradox: 723 FDA-cleared radiology AI devices yet fewer than 30% underwent clinical testing, and adoption barriers sit in infrastructure, governance, and workflow integration rather than detection performance—fewer than 1 in 3 pilots successfully transition to production.

Technical capability gains compete with deepening reliability concerns. Harrison.Rad 1.5 foundation model passes FRCR 2B professional board exam (86.5 vs 73.2 cutoff), first autonomous report-generation model validated against independent professional standard. Yet peer-reviewed audits reveal critical robustness gaps: vision-language models show fundamental grounding failures—text-only models reach within 5.7% accuracy of multimodal models, some models ignore images entirely, raising questions about whether high accuracy scores reflect genuine multimodal understanding. Information degradation in autonomous report generation is quantified: LLM rewriting erodes 51.4% of clinical entities and 43.7% of hedging language (clinical uncertainty markers) in EHR summarization tasks. Domain-specific tuning proves critical: LLMs trained on 500M+ radiology reports outperform generic GPT-4; autonomy quality depends on specialization not architecture. Yet cross-lingual prospective validation shows 98.7% physician-edit rates with diagnostic performance degrading from 0.2938 (English) to 0.2149–0.2424 (non-English). Foundation models exhibit severe age-related bias. August 2026 capability assessments show continued maturation alongside persistent failure modes: CARE-X (Microsoft Research) demonstrates unified chest X-ray VLM achieving SOTA benchmarks with clinical validation on rare pathologies and measurement-augmented inference (+43.6pp F1 improvement); yet vision-language models on brain MRI exhibit severe confidence miscalibration with 33-46% high-confidence errors, indicating that accuracy metrics mask fundamental reliability failures in uncertainty quantification. Forensic reproducibility audits reveal that even structured benchmark evaluations of radiology VLMs can yield protocol deviations and non-reproducible claims—a meta-level maturity indicator that capability evaluation itself requires strengthened rigor. This evidence pattern—strong domain-specific performance on narrow tasks, critical calibration gaps exposed by robustness audits, benchmark reproducibility concerns—defines the actual capability frontier.

Governance infrastructure is maturing but safety validation gaps persist. The American College of Radiology's June 2026 board chair statement endorses ARCH-AI (quality assurance), Assess-AI (post-deployment monitoring), and Healthcare AI Challenge Consortium (foundation model evaluation) as deployment readiness prerequisites, signaling institutional consensus that governance infrastructure is the frontier. ACR's August 2026 position statement reinforces augmentation-over-automation consensus, drawing on proven copilot models from ophthalmology, dermatology, and pathology—a professional society signal that full autonomy remains unachieved and likely unnecessary for clinical value. Yet structural barriers remain unresolved: JAMA Network Open cohort study of 903 FDA-authorized devices found 30 radiology AI devices recalled (4.3%), with missing clinical evidence associated with 1.39× higher recall hazard, documenting that validation gaps predict downstream safety failures. MIT CSAIL (June 2026) exposes critical regulatory classification loophole: autonomous radiology AI functions as clinical agent while classified as decision support exempt from FDA review. Most FDA-cleared radiology AI classified as 'aid in detection' or 'triage,' not autonomous screening; regulatory framework designed for fixed-functionality devices constrains autonomous system authorization. Peer-reviewed critical assessment recommends autonomous systems deployed as 'constrained, auditable, continuously monitored components of clinical practice that augment—but do not replace—expert judgment.' Fundamental questions about liability, accountability, and post-deployment monitoring responsibility prevent scaling beyond high-volume screening. A scoping review of 557 agentic AI studies in medicine identifies immature evaluation standards: process reliability, evidence traceability, uncertainty quantification, safety, workflow impact, and external validity remain assessed inconsistently across the field—a critical maturity gap for autonomous radiology systems. Pilot-to-production analysis reveals 723 FDA-cleared devices yet fewer than 30% underwent clinical testing; structured pre-deployment assessment frameworks show that radiologist adoption depends on governance, workflow integration, and organizational readiness rather than detection performance alone. The field consensus, articulated by the Radiology Research Alliance in March 2026 and reinforced in August 2026, emphasizes human-AI collaboration over full autonomy. This is the tension defining leading-edge: FDA Breakthrough designations and production deployments at scale clash against documented validation gaps, robustness failures, calibration failures, evaluation inconsistencies, and governance frameworks that remain prerequisites for scaling beyond specialists.

• 2023-H1: Emerging commercial deployment with multiple vendors at scale (3100+ sites, 10.7M scans processed by single vendor). Vendor landscape includes 179 CE-certified products by March 2023, with 67% having peer-reviewed evidence though evidence weighted toward diagnostic accuracy rather than clinical impact. Technical challenges documented: data imbalance in training sets, inadequate evaluation metrics, specific failure modes in non-standard imaging. Safety concerns noted including radiographer override rates and position-dependent accuracy degradation.

• 2023-H2: Substantial real-world deployment acceleration across multiple health systems and use cases. UK NHS trusts (Frimley, Greater Glasgow) deployed qXR for large-scale triage with 99.7% normal-case accuracy and 58% workload reduction potential. India-based TB screening program demonstrated 15.8% incremental yield improvement with autonomous systems. Teleradiology platforms (InHealth) integrated autonomous triage for critical finding alerting. Major commercial deployment at RSNA 2023: Microsoft/Nuance launched PowerScribe Smart Impression on platform used by 80% of radiologists. Adoption sentiment among clinicians remained positive (78% of Chinese residents surveyed support AI embrace), though replacement concerns persist (30% feared workforce reduction).

• 2024-Q1: Consolidation of commercial adoption and emergence of safety frameworks. Providence health system deployed PowerScribe autonomous reporting in largest US platform rollout, signaling mainstream health system adoption. Research momentum continues with IEEE review documenting methodological advances in automatic report generation. Emerging LLM approaches (ChatGPT/GPT-4) explored for autonomous reporting, though professional societies (ACR, CAR, ESR, RANZCR, RSNA) established formal evaluation and monitoring frameworks for autonomous AI tools, emphasizing rigorous safety assessment requirements alongside deployment.

• 2024-Q2: Major vendor expansion and critical deployment-readiness assessment. Microsoft Azure launches Radiology Insights preview service, signaling major cloud vendor entry into autonomous radiology tooling ecosystem. GPT-4 demonstrates capabilities matching radiologists on error detection (82.7% accuracy, lower cost and faster than human review). Independent academic assessment from Imperial College and Royal College of Radiologists documents widespread implementation barriers post-regulatory approval: reliability validation, accountability, trust, and safety governance gaps limit real-world adoption despite regulatory clearance. Asian Oceanian Society of Radiology formalizes regional adoption guidance. Open-source momentum continues with 100+ papers and tools curated in community repositories.

• 2024-Q3: Consolidation of clinical evaluation frameworks and LLM maturation. RSNA and MICCAI publish joint expert consensus on deployment barriers emphasizing trust, reproducibility, and accountability frameworks. LLM-based approaches demonstrate specific clinical capabilities: GPT-4 for synoptic report generation (F1 0.997), Claude-2 for RADS categorization (57% accuracy with prompting), and patient-friendly report generation (improving understanding scores). Radiologist adoption sentiment remains strong (75%+ engagement in AI practices in major markets). NHS Greater Glasgow & Clyde initiates prospective stepped-wedge clinical trial (RADICAL) for rigorous qXR evaluation across 24 months, signaling major health system commitment to evidence-based autonomous read validation. Systematic reviews synthesize efficiency gains (30-40% reporting time reduction) across diverse deployments, though net clinical outcome questions persist.

• 2024-Q4: Infrastructure consolidation and human factors maturation concerns. Largest US radiology practice (Radiology Partners, 3,900+ radiologists) partners with generative AI vendor for co-developed autonomous reporting, signaling shift toward strategic health system integration. Large-scale deployment validation continues: 1.3M chest X-rays processed across 33 UAE visa screening centers documented in peer-reviewed publication. Yet critical limitations emerge: multi-site prospective study documents over-reliance risk when AI provides local explanations (even incorrect ones), and practitioner surveys show mixed adoption sentiment with 100-radiologist cohort expressing concerns about reliability, job displacement, and ethical implications. JMIR viewpoint highlights persistent research-to-practice gap despite 190+ FDA-approved radiology AI devices, suggesting deployment maturity lags regulatory approvals.

• 2025-Q1: Broad hospital adoption and implementation gap evidence. Danish large-scale study demonstrates AI-driven mammography screening with 248,000+ images achieving 48.8% workload reduction while maintaining cancer detection, expanding evidence beyond chest X-ray modality. Pew Charitable Trusts survey (2022 data) documents 44% of US hospitals adopted AI imaging tools, yet critical implementation gaps persist: only 26% piloted before rollout, 34% lack comprehensive validation information, 31% lack monitoring protocols. Over 200 EU-approved radiology AI tools exist with limited real-world adoption, indicating continued research-to-practice gap. Patient acceptance studies show AI-simplified reports improve understanding but patients prefer physician delivery. Major health systems continue infrastructure investments despite unresolved questions about clinical effectiveness versus workflow efficiency gains.

• 2025-Q2: Continued deployment expansion and platform consolidation. UH Cleveland activated qXR for autonomous lung cancer detection on chest X-rays, demonstrating sustained adoption by major US health systems. Intelerad-RADPAIR partnership combines workflow orchestration with generative AI-driven reporting, signaling integration of autonomous reporting into mainstream clinical informatics platforms. Radiologist interviews reveal variability in AI monitoring practices and implementation approaches. Research continues on report generation advancements and faculty/trainee adoption sentiment.

• 2025-Q3: Vendor ecosystem consolidation and agentic AI emergence. RADPAIR-Fireworks partnership demonstrates production-scale autonomous reporting infrastructure at Radiology Partners (40-50M cases annually) with specific metrics: 15-20s reduced to 2-5s report turnaround, 25% time savings per case, 12% error reduction. RADPAIR expands geographic reach via AdvaHealth partnership into Asian healthcare markets. Commentary in Diagnostic and Interventional Radiology explores agentic AI for autonomous radiology workflows (RadGPT example for CT scanning). Academic research examines feasibility and barriers of autonomous CXR reporting in UK context, highlighting unresolved accountability framework gaps, regulatory challenges (IR(ME)R, GDPR), and need for post-market surveillance. RSNA covers autonomous report generation in neuroradiology using LLMs (GPT-4), discussing benefits and limitations including hallucinations and generalizability challenges.

• 2025-Q4: LLM performance maturation and validation gap emergence. Peer-reviewed study demonstrates efficiency gains in semi-automated AI reporting: 6.1-to-3.43 minute turnaround reduction on 100 complex cases with improved accuracy (3.81→4.65/5.0) and confidence (3.91→4.67/5.0, p<0.0001). Scoping review of 67 LLM studies shows strong structured-task performance (>94% accuracy in report simplification) but inconsistent diagnostic performance (16%-86%) with 79% single-center proof-of-concept designs. European adoption expands to 48% of radiologists using AI (up from 20% in 2018); 115+ FDA-approved algorithms by mid-2025. Agentic AI architectural shift: RADPAIR launches PAIRsdk developer framework with industry coalition (Fovia, Interlinx, deepc, Intelerad) planning open-source standards for voice-first autonomous workflows in 2026. Critical expert assessment documents widening validation gap: 100+ tools exhibited at RSNA 2025 lack rigorous clinical validation; algorithms perform significantly worse in diverse populations. Fundamental accountability and clinical outcome questions persist unresolved.

• 2026-Feb: Continued deployment expansion, regulatory acceleration, and LLM capability maturation. Autonomous TB screening in India demonstrates 80.21% increase in case notifications with 162 confirmed cases (44.63% positivity) at deployment sites in Chhattisgarh tribal population. FDA clearing accelerated with 56 new radiology AI devices (1,039 total devices, 80% of all FDA-authorized AI), signaling sustained regulatory momentum. Systematic review of 11 studies concludes autonomous chest X-ray triage systems ready for clinical implementation, with weighted average 42.3% autonomous triage rate and 97.8% sensitivity across real-world datasets (500K+ cases). Implementation barriers evidence: Brisbane hospital study identified 82 barriers and 33 enablers post-deployment using NASSS framework, with sustained adoption constrained by performance inconsistency, weak communication, and medicolegal uncertainty. Industry initiatives: Radiology Partners and Stanford AIDE Lab partnered to develop evidence-based validation and safety monitoring frameworks for autonomous AI tools; practitioner perspective documents job displacement concerns and malpractice exposure constraints on adoption. LLM capability advances: blinded evaluation studies compare LLM-generated to radiologist-written reports across clinical relevance and accuracy; fine-tuned Llama-3-70B demonstrates F1 0.780 error detection capability, outperforming GPT-4 (0.683); web-based automated chest X-ray report generation systems achieve BLEU-4 0.482 and ROUGE-L 0.718. Industry frameworks emerge: Signify Research survey of 150 healthcare organizations identifies seamless workflow integration as 9-10/10 priority and integration barriers as deal-breakers; SATMED and HealthManagement publish frameworks documenting transition from scattered pilots to core infrastructure, citing cloud-native platforms and agentic AI orchestration requirements. Commercial tool proliferation: Qure.ai expands qXR-Detect to 26 FDA indications; Report Rad AI launches claiming 60-95% faster reporting with 10K+ cases. Regulatory moment: ABR evaluates cautious AI adoption for internal functions while assessing competency frameworks for clinical tool use; adoption frameworks systematize integration, governance, and sustainability challenges underlying "pilot purgatory" barriers.

• 2026-Mar: Market consolidation accelerated with RadNet acquiring Gleamer for up to $270M and merging it into DeepHealth (26 FDA-cleared devices, 2,700+ customer contracts across 50 countries), signalling that scale and device breadth are becoming competitive moats. Deployment evidence continued broadening: ARA Health (13 hospitals, 100k studies/month) achieved 20% reporting-time reduction with Rad AI across 79% of radiologists; United Imaging Intelligence unveiled CE-certified uAI Image-to-Report agents generating structured CT and brain MRI preliminary reports across five European countries; and a Ghana clinical study found autonomous AI TB screening achieved 91% accuracy versus 86% for radiologists in a resource-limited setting. The Radiology Research Alliance published multi-institutional consensus in March 2026 that human-AI collaboration produces stronger outcomes than full autonomy, reflecting field caution about replacing radiologist sign-off despite rising deployment velocity.

• 2026-May: Commercial adoption deepened with Rad AI Omni deployed in 8 of 10 largest U.S. private radiology practices, and the RadNet-Gleamer merger creating 700+ customer contracts with autonomous draft reporting as a core capability. DeepHealth revenue grew 51.5% YoY to $29.1M with ARR reaching $96.9M (95% YoY growth) and guidance exceeding $140M, confirming production-scale commercial momentum across 2,890+ customers; RadNet reported record Q1 2026 results with 70%+ of studies expected to run through clinical AI by year-end. ACR launched ARCH-AI and Assess-AI governance programs — formal quality assurance and post-deployment monitoring registries — signalling field consensus that governance infrastructure now defines the frontier. Critical capability limits emerged: the ABRA agentic radiology benchmark found agents achieve 89% success on tool orchestration but only 0-25% on real annotation outcomes, locating the bottleneck in visual perception rather than reasoning; Stanford HAI policy research documented that most deployed radiology AI systems lack robust performance monitoring despite rapid clinical adoption. Independent analysis further documented the evidence gap between validated AI-assisted detection and autonomous-only reads lacking equivalent clinical trial validation.

• 2026-Jun: Regulatory acceleration with dual FDA Breakthrough Device Designations for autonomous report generation. Aidoc's First Read (chest X-ray autonomous reports, 2,000+ hospitals, 120M+ cases, $150M Series E) and Cognita (acquired by Radiology Partners, autonomous report generation) both received Breakthrough status June 25-26, signaling regulatory recognition of autonomous preliminary reads as addressing unmet clinical need. Commercial deployment momentum confirmed: Yale New Haven Health System (700k+ annual exams) selected Rad AI for infrastructure-scale rollout; Mosaic Reporting (Radiology Partners subsidiary) deployed to thousands with real-time autonomous draft generation; HOPPR Presto Agent launched with PowerScribe integration. Northwestern Medicine published case study showing 40% productivity boost with autonomous reporting (JAMA Network). Capability maturation signals include Harrison.Rad 1.5 passing FRCR 2B board exam and MICCAI 2026 research on clinically-controlled autonomous RRG with precision-recall tradeoffs. Yet robustness and safety concerns deepened: peer-reviewed audit documented vision-language models with critical grounding failures (text-only models within 5.7% accuracy of multimodal; some models ignore images); research on LLM-rewritten reports quantified 51.4% entity erosion and 43.7% hedging loss. JAMA Network Open revealed safety validation gaps: 30 radiology AI devices recalled (4.3%), with missing clinical evidence 1.39× higher recall hazard. Peer-reviewed critical assessment identified realism-reliability gap in generative medical imaging and recommended constrained deployment with continuous monitoring. Governance maturation: ACR Board Chair formally endorsed ARCH-AI, Assess-AI, and Healthcare AI Challenge Consortium as deployment readiness prerequisites. Field consensus emphasizes that governance infrastructure, not capability maturation, now defines the frontier.

• 2026-Jul: Continued deployment scaling and governance paradoxes emerged. Natoe AI expanded its AI-native teleradiology platform with production FDA-cleared autonomous pre-read reports across U.S. hospitals and imaging centers; radiologist sign-off remains mandatory. Peer-reviewed validation studies accumulated: the xAID draft-then-sign workflow (758 chest X-rays) achieved 42% reading time reduction (34.2→19.8s) with maintained quality and improved sensitivity (pleural lesions 77.7→87.4%); a Curely deployment-grounded framework synthesized 57 external-validation studies confirming performance-transfer gaps as a core deployment barrier (a sepsis model's AUC dropped from 0.76–0.83 in development to 0.63 in deployment). A governance readiness paradox emerged: a survey of 2,500+ enterprises found 75% rolled back production AI agents, with mature-governance organizations showing an even higher 81% rollback rate — signaling that governance infrastructure surfaces problems ungoverned deployments never detect. Reliability concerns deepened further: a Stanford-Harvard NOHARM evaluation documented a 22% severe clinical error rate across frontier LLMs, with external validation showing deployed systems systematically lack performance monitoring despite rapid clinical adoption. Regulatory ecosystem activity continued: Q2 2026 FDA data showed 86 AI/ML authorizations (59 radiology), and a follow-up report on Aidoc's First Read and Cognita's dual FDA Breakthrough Designations (originally granted in June) reiterated the persistent radiologist-verification burden threatening anticipated efficiency gains. Penn Medicine's Arnie automated radiology recommendation system reported 12 early-stage cancers detected in initial deployment, again underscoring workflow integration as the critical adoption barrier despite autonomous capability. Late-month evidence sharpened both threads: an 8-month observational study found generative AI tools improved radiologist productivity but drove cognitive fatigue and decision-quality erosion; a domain-specific model (M4CXR) cut reading time from 179 to 16 seconds yet retained 25.2% inconsistency, and the RadLE 2.0 benchmark found AI radiology models delivering incorrect findings with high confidence; an independent audit of the FDA device database confirmed only 1.6% of 1,451 authorizations cite RCT data and under 1% report patient outcomes, while EU AI Act audits found roughly half of high-risk deployments lack adequate post-market-monitoring plans — reinforcing that governance and reliability, not capability, remain the binding constraints on autonomous adoption.

• 2026-Aug: ACR governance commentary endorsed augmentation-over-automation, citing successful copilot models in other specialties, while an adoption-barrier analysis found fewer than 30% of 723 FDA-cleared radiology AI devices underwent clinical testing, locating stalled pilots in infrastructure and governance rather than detection performance. Reliability evidence deepened: a safety audit of six vision-language models on 4,102 brain MRIs found 33-46% of high-confidence answers were errors, and a forensic reproducibility audit of a radiology VLM benchmark found unreproducible results leading to withdrawn performance claims. Countervailing capability progress continued: Microsoft's CARE-X chest X-ray VLM achieved SOTA report-generation performance with clinical validation in India, and a 298,991-exam real-world deployment of an autonomous system reported 87.1% sensitivity, 91.8% specificity, and 99.0-100.0% NPV — while a 557-study scoping review of agentic medical AI flagged persistent gaps in process reliability, evidence traceability, and external validity.