Generative AI in Healthcare Statistics 2026
By Axis Intelligence Research
Co-authors: Jennifer Miller (Digital Health) · Sarah Mitchell (AI & Machine Learning) | Last updated: August 4, 2026 | License: CC BY 4.0
Four in five U.S. physicians now use AI professionally, and the two things they use it for most — summarizing medical literature and drafting clinical notes — are both generative tasks. Yet Axis Intelligence Research counted 92 AI-enabled device authorizations on the FDA’s list in Q1 2026 and found zero identified as using large language models or foundation models.
Quick Answer
Generative AI has become routine clinical infrastructure in the United States without passing through the regulatory gate that governs every other class of clinical AI. Physician AI use reached 81% in 2026, up from 38% in 2023, per the American Medical Association’s annual survey. Of the 92 AI-enabled medical device authorizations Axis Intelligence Research counted on the FDA’s list for Q1 2026, none were identified as incorporating large language models — the FDA has stated it is still developing methods to tag such devices. Axis Intelligence Research puts the resulting Generative AI Governance Gap Index (AGGI) at a baseline reading of 60.0 as of August 4, 2026, with the widest gaps in patient-facing use, where 32% of U.S. adults now consult AI chatbots for health information under no clinical oversight at all.
Key Findings
- According to Axis Intelligence Research’s count of the FDA AI-Enabled Medical Device List, 92 AI-enabled devices received marketing authorization with Q1 2026 decision dates, and none were identified as incorporating large language models or foundation models.
- Axis Intelligence Research finds that radiology accounted for 69 of those 92 authorizations (75.0%), meaning the FDA’s AI pipeline remains overwhelmingly an imaging pipeline while generative deployment concentrates in documentation and patient communication.
- Physician AI use reached 81% in 2026, more than double the 38% recorded in 2023, with an average of 2.3 use cases per physician, according to the American Medical Association’s 2026 Physician Survey on Augmented Intelligence.
- Axis Intelligence Research estimates the Axis Generative AI Governance Gap Index (AGGI) at a baseline of 60.0, with patient-directed general health queries scoring highest (72.8) and FDA-authorized generative clinical output scoring lowest (35.0) — the only domain with full oversight is the only domain with no deployment.
- 41% of adults who use AI for health information have uploaded personal medical records into a chatbot — 13% of all U.S. adults — while 77% of the public says it is concerned about the privacy of medical information given to AI tools, per KFF’s March 2026 Tracking Poll.
How Many Physicians Use Generative AI in 2026?
The American Medical Association’s 2026 Physician Survey on Augmented Intelligence, fielded among 1,692 physicians between January 15 and February 2, 2026, found 81% using AI in a professional capacity. That is the headline. The composition of that 81% is the story.
The two most common applications are summarizing medical research and standards of care (39% of physicians) and creating discharge instructions, care plans, or progress notes (30%). Both are text generation. Neither existed as a mainstream clinical workflow before large language models. The average physician now runs 2.3 distinct AI use cases, up from 1.1 in 2023 — adoption has not just widened, it has deepened per user.
Physician AI adoption, 2023–2026
| Metric | 2023 | 2024 | 2026 | Source |
|---|---|---|---|---|
| Physicians using AI professionally | 38% | 66% | 81% | AMA Physician Survey on Augmented Intelligence |
| Average use cases per physician | 1.1 | — | 2.3 | AMA Physician Survey on Augmented Intelligence |
| Believe AI improves ability to care for patients | 65% | — | >75% | AMA Physician Survey on Augmented Intelligence |
| Survey sample (n) | 1,081 | 1,183 | 1,692 | AMA Physician Survey on Augmented Intelligence |
What physicians are worried about
Enthusiasm and unease arrive together. Seventy percent of physicians see AI as a way to automate the tasks that drive burnout. Eighty-eight percent are concerned about skill loss, a worry concentrated among those with ten years or less in practice. Eighty-five percent want to be consulted on adoption decisions at their own institution, and the AMA reports that clear liability frameworks rank highest among the regulatory actions physicians say would increase their trust.
Sarah Mitchell, AI & Machine Learning: The liability finding is the one to sit with. Physicians are not asking for better benchmarks or bigger context windows. They are asking who is legally holding the bag when a generated note contains something the encounter never produced. That question has no answer in the current U.S. framework, because the FDA pathway that would normally supply one has not been extended to the generative function. Eighty-eight percent flagging skill loss is a related signal: it is a workforce telling you it has already changed its practice around a tool nobody has certified.
How Many FDA-Authorized Medical Devices Use Generative AI?
None that the FDA has identified. Axis Intelligence Research downloaded the FDA AI-Enabled Medical Device List on August 4, 2026 — a list the agency last updated on June 16, 2026, carrying decisions through March 30, 2026 — and counted every authorization with a Q1 2026 decision date.
The FDA states on that page that it “will explore methods to identify and tag medical devices that incorporate foundation models,” spanning large language models to multimodal architectures, and encourages sponsors to include the relevant information in their public summaries so such devices can be identified in a future update. In plain terms: as of the current list, the agency does not yet have a way to tell the public which authorized devices contain an LLM, and no device is so labeled.
Q1 2026 AI-enabled device authorizations by specialty panel
| Panel | Authorizations | Share | Source |
|---|---|---|---|
| Radiology | 69 | 75.0% | FDA AI-Enabled Medical Device List (Axis count) |
| Cardiovascular | 10 | 10.9% | FDA AI-Enabled Medical Device List (Axis count) |
| Neurology | 5 | 5.4% | FDA AI-Enabled Medical Device List (Axis count) |
| Orthopedic | 2 | 2.2% | FDA AI-Enabled Medical Device List (Axis count) |
| All other panels combined | 6 | 6.5% | FDA AI-Enabled Medical Device List (Axis count) |
| Total | 92 | 100% | FDA AI-Enabled Medical Device List (Axis count) |
Other panels each recording a single Q1 2026 authorization: ear, nose and throat; general hospital; gastroenterology-urology; general and plastic surgery; hematology; microbiology.
The Generative AI Regulatory Coverage Ratio
Axis Intelligence Research computes a simple cross-source ratio to make the asymmetry citable:
Generative AI Regulatory Coverage Ratio = (authorizations identified as using LLM or foundation models) ÷ (total AI-enabled device authorizations in period)
For Q1 2026: 0 ÷ 92 = 0.0%.
Set that against physician deployment. Thirty percent of physicians use AI to generate discharge instructions, care plans, or progress notes. The gap between generative deployment in clinical documentation and generative coverage in FDA premarket review is 30 percentage points against a coverage base of zero.
Jennifer Miller, Digital Health: The precise limitation worth naming: FDA’s list is assembled by scanning marketing-authorization summaries for AI terminology, and the agency says plainly it is not comprehensive. So “zero LLM-tagged devices” is a statement about what the regulator can currently see, not proof that no authorized device contains a language model anywhere in its stack. That distinction matters, and it cuts the wrong way for reassurance. If the tagging method does not exist yet, neither does the surveillance. A clinician reading a device’s decision summary today cannot determine whether generative components are present.
How Widely Have Hospitals Deployed Clinical AI?
The most rigorous national measurement of institutional AI use is federal. ASTP/ONC Data Brief No. 80, drawn from the American Hospital Association IT Supplement, found 71% of non-federal acute care hospitals using predictive AI integrated with their EHR in 2024, up from 66% in 2023.
That brief measures predictive AI, not generative — a distinction Axis Intelligence Research preserves rather than blurs, because merging survey instruments with different definitions produces a number that means nothing. But ASTP itself draws the connective line, noting that the fastest-growing predictive use cases parallel the increasing use of generative AI for administrative processes, and describing the pattern as spillover.
Hospital predictive AI adoption and the digital divide, 2024
| Hospital characteristic | Predictive AI use | Locally evaluated for accuracy | Source |
|---|---|---|---|
| Multi-hospital system member | 86% | 76% | ASTP/ONC Data Brief No. 80 |
| Independent hospital | 37% | 41% | ASTP/ONC Data Brief No. 80 |
| Large (400+ beds) | 96% | 74% | ASTP/ONC Data Brief No. 80 |
| Small (<100 beds) | 59% | 65% | ASTP/ONC Data Brief No. 80 |
| Urban | 81% | 75% | ASTP/ONC Data Brief No. 80 |
| Rural | 56% | 60% | ASTP/ONC Data Brief No. 80 |
| Critical access hospital | 50% | 59% | ASTP/ONC Data Brief No. 80 |
| Market-leading EHR vendor | 90% | 80% | ASTP/ONC Data Brief No. 80 |
| All other EHR vendors | 50% | 52% | ASTP/ONC Data Brief No. 80 |
The administrative shift is unambiguous. Hospital use of predictive models to simplify or automate billing rose from 36% to 61% in a single year, and scheduling from 51% to 67%. Use for treatment recommendations moved two percentage points. The money-and-paperwork layer is where the automation is landing.
Nationally, 82% of hospitals evaluated at least some predictive models for accuracy in 2024, 74% for bias, and 79% conducted post-implementation monitoring. The distribution is the problem: independent hospitals sit at 41% for local accuracy evaluation against 76% for system members. Roughly the same institutions with the least AI have the least capacity to check the AI they do have.
What Does the Clinical Evidence Actually Show?
Two studies published within two months of each other reached compatible conclusions with very different force, and reading them together is more useful than picking one.
The larger enthusiasm signal comes from a multicenter quality improvement study published in JAMA Network Open in October 2025. Across 263 physicians and advanced practice practitioners at six U.S. health systems, self-reported burnout fell from 51.9% to 38.8% after 30 days of ambient AI scribe use — a 13.1-percentage-point drop, with improvements also reported in cognitive task load, after-hours documentation, and focused attention on patients.
The tighter methodological signal comes from a randomized trial published in NEJM AI in November 2025 (Lukac et al., doi:10.1056/aioa2501000). At UCLA Health, 238 outpatient physicians across 14 specialties were randomized to one of two ambient scribe products or usual care. Time spent writing each note fell by an estimated 41 seconds in the Nabla arm — from 4 minutes 30 seconds to 3 minutes 49 seconds — against 18 seconds in the control arm. Net effect: roughly 23 seconds per note.
Ambient documentation evidence at a glance
| Study | Design | n | Headline result | Source |
|---|---|---|---|---|
| Olson et al., 2025 | Pre/post quality improvement, 6 systems | 263 | Burnout 51.9% → 38.8% at 30 days | JAMA Network Open 2025;8(10):e2534976 |
| Lukac et al., 2025 | Randomized, 1:1:1, single center | 238 | −41s vs −18s per note (Nabla vs control) | NEJM AI 2025;2(12), doi:10.1056/aioa2501000 |
Jennifer Miller, Digital Health: These are not contradictory findings, and anyone citing one against the other is misreading both. The pre/post study measures how clinicians feel after a month; the randomized trial measures keystrokes. Both can be true — a tool can meaningfully change the experience of documentation while shaving seconds off the log. But note the study designs before you quote a number in a board deck. The burnout study is uncontrolled, self-reported, single-vendor, and carries a vendor-affiliated co-author; the randomized trial is single-center over two months. The honest version of the 2026 evidence base is: promising on wellbeing, modest and unconfirmed on time, and thin on downstream outcomes like note accuracy, coding integrity, and diagnostic error. Nobody has published the study that would tell you whether a generated note is a safe note at scale.
How Many Patients Use Generative AI for Health Advice?
This is the layer with no clinician in it, and it is now the largest layer by population.
KFF’s Tracking Poll on Health Information and Trust, fielded February 24 to March 2, 2026 among 1,343 U.S. adults, found 32% had turned to AI chatbots for health information in the past year — about the same share that uses social media for the purpose. Twenty-nine percent sought physical health information; 16% sought mental health information.
Patient generative AI use, 2026
| Measure | Share | Source |
|---|---|---|
| Adults using AI chatbots for health information (past year) | 32% | KFF Tracking Poll, March 2026 |
| Used AI for physical health information | 29% | KFF Tracking Poll, March 2026 |
| Used AI for mental health information | 16% | KFF Tracking Poll, March 2026 |
| Adults under 30 using AI for mental health | 28% | KFF Tracking Poll, March 2026 |
| Adults 50+ using AI for mental health | 8% | KFF Tracking Poll, March 2026 |
| Uninsured adults using AI for mental health | 30% | KFF Tracking Poll, March 2026 |
| Insured adults using AI for mental health | 14% | KFF Tracking Poll, March 2026 |
| AI health users who uploaded personal medical records | 41% | KFF Tracking Poll, March 2026 |
| All adults who uploaded personal medical records to AI | 13% | KFF Tracking Poll, March 2026 |
| Public concerned about privacy of medical data given to AI | 77% | KFF Tracking Poll, March 2026 |
Why patients turn to chatbots
Sixty-five percent of users cite wanting quick or immediate information. Forty-one percent used AI before deciding whether to see a provider. Thirty-six percent said they felt more comfortable looking things up privately. Then the access findings: 19% cite not being able to afford a health professional and 18% cite having no regular doctor or no available appointment. Among users under 30, those figures rise to 29% and 38%.
The follow-up numbers are the ones that should travel. Fifty-eight percent of those who asked AI about mental health did not follow up with a health professional afterward. For physical health, 42% did not.
Jennifer Miller, Digital Health: Read the access data as clinical substitution, not curiosity. A 38% access-barrier rate among under-30 users, combined with 58% of mental health queries ending without a clinician, describes a chatbot functioning as a first — and frequently only — point of contact for a population that is disproportionately uninsured, young, Black, and Hispanic. Whatever one thinks of the technology, that is triage happening outside the health system, with no adverse-event reporting pathway, no scope-of-practice boundary, and no obligation on the tool to escalate. The privacy finding compounds it: 77% of the public is concerned about medical data given to AI, and 41% of users upload their records anyway. Stated concern is not restraining behavior, which means disclosure-based policy will not restrain it either.
Who Actually Regulates Generative AI in Clinical Care?
Not the FDA, for the generative function. The binding rules that do exist are largely at the state level, and the earliest is narrow by design.
California’s Assembly Bill 3030, effective January 1, 2025 and codified at Health and Safety Code §1339.75, requires health facilities, clinics, physicians’ offices, and group practices that use generative AI to produce patient communications about clinical information to include a disclaimer that the communication was AI-generated, plus clear instructions for reaching a human. Violations by physicians fall under the Medical Board of California.
Two exemptions define the law’s actual reach. Communications read and reviewed by a licensed human provider are exempt, and communications unrelated to clinical information — scheduling, billing — are outside scope. In practice, the statute regulates unreviewed clinical output only. Reviewed output, which is how most health systems deploy ambient documentation, sits outside it.
The oversight stack for generative AI in U.S. healthcare, August 2026
| Layer | Binding premarket review? | What governs it | Source |
|---|---|---|---|
| Generative clinical device output | Yes, in principle | FDA 510(k)/De Novo/PMA — no LLM-tagged authorization identified | FDA AI-Enabled Medical Device List |
| Clinician-reviewed AI documentation | No | Professional and accreditation guidance; AB 3030 exempts reviewed communications | California Health & Safety Code §1339.75 |
| Unreviewed AI patient communications (California) | No | AB 3030 disclosure requirement | California Health & Safety Code §1339.75 |
| Point-of-care literature summarization | No | No domain-specific rule | Axis Intelligence Research assessment |
| Direct-to-consumer health chatbots | No | General consumer protection law | Axis Intelligence Research assessment |
The Axis Generative AI Governance Gap Index (AGGI)
Adoption statistics and regulatory statistics are usually published separately, which is why the mismatch between them is rarely quantified. AGGI exists to quantify it.
AGGI stands for the Axis Generative AI Governance Gap Index. It scores individual generative-AI application domains in U.S. healthcare on how far measured deployment and clinical consequence have run ahead of binding oversight. Higher readings mean a wider gap.
Formula
AGGI(d) = 0.40 × D + 0.35 × C + 0.25 × (100 − O)
Where, for each domain d, all components run 0–100:
- D — Deployment intensity. The measured share of the relevant population using generative AI in that domain, taken directly from a sourced fact-table row. No estimates.
- C — Clinical consequence. A disclosed ordinal: 100 = output can directly drive a patient’s care decision with no licensed clinician necessarily in the loop; 80 = output enters the legal medical record or drives clinical action with clinician review expected; 60 = output informs clinician judgment while the clinician remains decision-maker; 40 = administrative only.
- O — Oversight coverage. A disclosed ordinal: 100 = FDA premarket authorization required and granted for the generative function; 75 = binding federal statute or regulation specific to the generative function; 50 = binding state statute specific to the generative function; 25 = professional or accreditation guidance only, non-binding; 0 = no domain-specific rule, general consumer law only.
The composite is the unweighted mean of all six domain scores.
AGGI domain readings, as of August 4, 2026
| Domain | D | Deployment source | C | O | AGGI |
|---|---|---|---|---|---|
| Patient general health queries | 32 | KFF Tracking Poll, March 2026 | 100 | 0 | 72.8 |
| Patient mental health queries | 16 | KFF Tracking Poll, March 2026 | 100 | 0 | 66.4 |
| Patient upload of own medical records | 13 | KFF Tracking Poll, March 2026 | 100 | 0 | 65.2 |
| Point-of-care knowledge summarization | 39 | AMA Physician Survey, 2026 | 60 | 0 | 61.6 |
| Clinical documentation drafting | 30 | AMA Physician Survey, 2026 | 80 | 25 | 58.8 |
| FDA-authorized generative clinical output | 0 | FDA AI-Enabled Medical Device List | 100 | 100 | 35.0 |
| AGGI composite | 60.0 |
This is a baseline reading. AGGI has no prior history, and no reading before August 4, 2026 exists to compare it against.
Worked example (clinical documentation drafting): 0.40 × 30 = 12.0; 0.35 × 80 = 28.0; 0.25 × (100 − 25) = 18.75. Sum = 58.75, reported as 58.8.
Sensitivity check. Weighting the composite by deployment intensity rather than treating domains equally yields 64.7 instead of 60.0. The direction does not change: the domains carrying the most users carry the least oversight.
Sarah Mitchell, AI & Machine Learning: The shape of that table is the finding, not the composite. The single domain scoring 35 — the low, well-governed end — is the one with zero deployment. Every domain where people are actually using the technology sits between 58.8 and 72.8. That is not a regulator falling behind a fast-moving field, which is the usual framing. It is a regulatory perimeter drawn around a category of product that generative AI in healthcare has mostly declined to become. Scribes are not devices. Chatbots are not devices. Literature summarizers are not devices. The perimeter did not move; the technology walked around it.
Stated limitations
Two of the three AGGI components are ordinal assignments by Axis Intelligence Research, not measurements. C and O are expert judgments applied against the published scales above; a reasonable analyst could assign different values and produce a different composite. D is measured and sourced in every case. AGGI also does not capture deployment outside the United States, enterprise deployments not visible in physician or consumer surveys, or the quality of any individual tool. The index measures the distance between deployment and binding oversight — nothing about whether a given system is safe or effective.
Where Is the Money Going?
Capital has stopped treating AI as a category. U.S. digital health startups raised $7.4 billion across 244 deals in the first half of 2026, per Rock Health, against $6.4 billion across 245 deals in H1 2025 — a billion more dollars on effectively the same deal count. Mega-deals of $100 million or more absorbed 45% of all capital, up from 22% in 2024. Median deal size rose to $14 million.
The most telling data point is a measurement decision rather than a number: Rock Health stopped labeling startups as “AI-enabled” in Q1 2026, on the grounds that AI is no longer a distinguishing product or strategy. For a tracker that had reported AI deals as a separate cohort for years, retiring the category is a statement that the cohort is now the whole market.
| Metric | H1 2025 | H1 2026 | Source |
|---|---|---|---|
| U.S. digital health VC raised | $6.4B | $7.4B | Rock Health |
| Deal count | 245 | 244 | Rock Health |
| Mega-deal share of capital | 42% (FY2025) | 45% | Rock Health |
| Median deal size | $12M (FY2025) | $14M | Rock Health |
| Acquisitions | — | 115 | Rock Health |
For context on the broader capital picture, see Axis Intelligence Research’s AI investment statistics and digital health statistics.
Methodology
Collection. All figures were retrieved from primary sources during the production session on August 4, 2026, and logged with source URL and retrieval date in the accompanying CSV. Sources used: the FDA AI-Enabled Medical Device List (list last updated June 16, 2026); the AMA 2026 Physician Survey on Augmented Intelligence (fielded January 15–February 2, 2026, n=1,692); ASTP/ONC Health IT Data Brief No. 80, drawn from the 2023–2024 AHA IT Supplement; KFF’s Tracking Poll on Health Information and Trust (fielded February 24–March 2, 2026, n=1,343, ±3 percentage points); Olson et al., JAMA Network Open 2025;8(10):e2534976; Lukac et al., NEJM AI 2025;2(12), doi:10.1056/aioa2501000; California Health and Safety Code §1339.75 (AB 3030); and Rock Health’s H1 2026 funding recap.
Axis-calculated figures. Q1 2026 device counts were produced by Axis Intelligence Research by filtering the FDA list to decision dates between January 1 and March 31, 2026 and tallying by lead specialty panel. The Generative AI Regulatory Coverage Ratio and all AGGI domain scores and composites were computed from the formulas disclosed above. All arithmetic was independently recomputed before publication.
Definitions. “Generative AI” refers to systems producing novel text, summaries, or communications, including large language models. “Predictive AI,” as used in the ASTP data, refers to statistical and machine learning models producing risk scores or classifications. Axis Intelligence Research does not merge these two categories into a single adoption figure, because the underlying survey instruments define them differently and averaging them would produce a number with no referent.
Limitations. The FDA list is compiled by scanning marketing-authorization summaries for AI terminology and the agency states it is not comprehensive; devices whose summaries omit such language may be undercounted, and the FDA sometimes adds older devices retroactively, so Q1 2026 counts are a snapshot as retrieved. ASTP hospital figures reflect 2024 survey data, the most recent federal measurement available, and are flagged as such. The two ambient documentation studies differ in design, setting, and vendor, and neither measures clinical outcomes. AGGI’s consequence and oversight components are disclosed ordinal judgments, not measurements. Physician and consumer survey figures are self-reported.
About This Dataset
This dataset compiles 120 verified statistics on generative AI adoption, regulation, clinical evidence, and investment in U.S. healthcare, spanning 2023 through August 2026, with full provenance for every row — 99 drawn from primary sources and 27 calculated by Axis Intelligence Research from disclosed formulas. It includes the component inputs and readings for the Axis Generative AI Governance Gap Index (AGGI).
License: CC BY 4.0. You may share and adapt this data for any purpose with attribution: Axis Intelligence Research, “Generative AI in Healthcare Statistics 2026,” axis-intelligence.com, 2026.
Dataset distribution: Published to Hugging Face, Kaggle, and GitHub under CC BY 4.0.
Cite This Research
APA: Axis Intelligence Research. (2026). Generative AI in Healthcare Statistics 2026: Adoption, Evidence, and the Oversight Gap. axis-intelligence.com/generative-ai-in-healthcare-statistics/
MLA: Axis Intelligence Research. “Generative AI in Healthcare Statistics 2026: Adoption, Evidence, and the Oversight Gap.” Axis Intelligence Research, 2026, axis-intelligence.com/generative-ai-in-healthcare-statistics/.
Chicago: Axis Intelligence Research. “Generative AI in Healthcare Statistics 2026: Adoption, Evidence, and the Oversight Gap.” axis-intelligence.com/generative-ai-in-healthcare-statistics/. Published 2026.
Frequently Asked Questions
Has the FDA cleared any large language model for clinical use?
No device on the FDA’s AI-Enabled Medical Device List is identified as incorporating a large language model or foundation model. The agency states it is developing methods to tag such devices and is asking sponsors to disclose the relevant information in public summaries. Because the list is built by scanning authorization summaries for AI terminology and is not comprehensive, “none identified” describes the limits of current regulatory visibility rather than a verified absence.
Why are ambient AI scribes not regulated as medical devices?
Ambient documentation tools generally position themselves as drafting support for a clinician who reviews and signs the note, which places them outside the device definition that triggers FDA premarket review. California’s AB 3030 follows the same logic in reverse: it regulates AI-generated patient communications about clinical information but exempts anything a licensed provider has read and reviewed. The clinician’s review is the legal firewall in both frameworks.
Do ambient AI scribes actually save clinicians time?
The randomized evidence says seconds per note, not hours. The UCLA trial published in NEJM AI found note-writing time fell 41 seconds in one intervention arm against 18 seconds in usual care. The larger wellbeing effect is better supported than the efficiency effect: a six-system quality improvement study in JAMA Network Open found self-reported burnout dropping from 51.9% to 38.8% over 30 days, though that study was uncontrolled and single-vendor.
What is AGGI and what does a reading of 60.0 mean?
AGGI is the Axis Generative AI Governance Gap Index, an Axis Intelligence Research metric scoring how far generative AI deployment and clinical consequence have outrun binding oversight across six healthcare domains. The August 4, 2026 composite of 60.0 is a baseline — the first reading, with no prior history. Its useful content is the spread: patient-facing domains score 65.2 to 72.8 with zero domain-specific regulation, while the only fully regulated domain scores 35.0 because nothing has been authorized in it.
Are patients using AI chatbots instead of seeing a doctor?
Frequently, yes. KFF found that 58% of adults who consulted AI about mental health and 42% who consulted it about physical health did not follow up with a health professional. Access and cost are explicit drivers: 19% of users cite affordability and 18% cite lacking a regular doctor or appointment, rising to 29% and 38% among users under 30.
How many hospitals evaluate their AI models locally?
Among U.S. non-federal acute care hospitals in 2024, 82% evaluated at least some predictive AI models for accuracy, 74% for bias, and 79% conducted post-implementation monitoring. The distribution is uneven: 76% of multi-hospital system members reported local accuracy evaluation against 41% of independent hospitals. Note that this federal measurement covers predictive models, not generative ones.
Which medical specialties dominate FDA AI authorizations?
Radiology, by a wide margin. Axis Intelligence Research counted 69 of 92 Q1 2026 authorizations (75.0%) in the radiology panel, followed by cardiovascular at 10 and neurology at 5. The FDA’s AI pipeline remains an imaging pipeline, while generative deployment concentrates in documentation, literature summarization, and patient communication — none of which produce device authorizations.
Does HIPAA cover medical records uploaded to a consumer chatbot?
HIPAA obligations attach to covered entities and their business associates, not to a patient’s own voluntary disclosure to a consumer product. When an individual uploads their test results or clinician notes into a general-purpose chatbot, that transfer generally falls outside the HIPAA framework and is governed by the product’s terms and general consumer protection law. KFF found 41% of AI health users — 13% of all U.S. adults — have done exactly this.
What would close the oversight gap?
Three levers appear in the current data. The FDA tagging foundation-model devices, which the agency has stated it intends to explore, would create visibility. Extending premarket review to generative clinical functions would create a binding federal layer where none exists. And the AMA’s finding that clear liability frameworks rank as physicians’ highest regulatory priority points at the third: an allocation of legal responsibility for generated clinical content, which no current U.S. framework supplies.
Related Research from Axis Intelligence Research
- Digital Health Statistics 2026 — market size, telehealth, wearables, and funding data
- Telehealth Statistics 2026 — utilization, mental health dominance, and Medicare policy
- AI Hallucination Statistics 2026 — hallucination rates by domain, including clinical settings
- Healthcare Cybersecurity Statistics 2026 — breach costs and threat data for health systems
- AI Governance by Industry 2026 — cross-sector regulatory frameworks
- AI in Legal Statistics 2026 — the comparable adoption-versus-oversight picture in law
- AI Drug Discovery Statistics 2026 — AI-originated drugs in clinical trials