Provider Identity in the AI Era: Why Accurate Healthcare Data Infrastructure Matters

Pouyan Golshani, MD · Interventional Radiologist; Founder, GigHz · October 2026

Written in a personal capacity. Disclosure: the author is a partner in a California medical group and the founder of Guide.MD, a provider-data platform, and of clinical decision-support products. The measurements in Section 2 were made by Guide.MD on its copy of public records and have not been independently reviewed.

Summary

Patients increasingly meet their doctors through a search result or an AI-generated answer before they meet them in person. Those answers are assembled from public provider records that were never designed for that purpose. When the record is wrong, the answer is wrong, and the clinician usually has no way to fix it. California has begun to legislate on AI in health care. This brief describes the data problem underneath, measures it on California's own public records, and offers options that reuse what the state already has.

Three measured facts frame the problem. Of the 993,162 federal NPI records for clinicians with a California practice address, 42 percent carry no state license number. Where a federal record could be tied to a state license, about one in fourteen of those licenses is delinquent or inactive, which the federal record does not show. And 649,899 people hold a current California health license that could not be matched to any NPI record, so a system keyed only to the federal identifier may not see them.

1. The record behind every clinician

No single source says who a clinician is, where they practice, and what they are qualified to do. The pieces sit in different places:

• National Provider Identifier (NPI) registry. Federal, public, and self-reported. An NPI does not confirm that a provider is licensed or credentialed, and many licensed nurses, hygienists, assistants, and technicians do not have one.
• State licensing boards. Authoritative for licensure and discipline. The address a board holds is an address of record, which may be a home or mailing address and not a place of practice. Boards generally do not record specialty or services.
• Medicare enrollment and other CMS public files. Cover only clinicians who bill Medicare.
• Health plan directories. Maintained plan by plan, with different update cycles.

Each source is updated on its own schedule by a different owner. Linking them requires matching people and organizations across records that use different names, addresses, and identifiers. There is no reliable shared key: the federal record has a field for the state license number, and for California clinicians it is empty about two times in five.

2. What goes wrong

• Stale practice locations and phone numbers after a clinician moves.
• Duplicate or merged records for clinicians with similar names.
• Confusion between an individual clinician and the organization that bills for them.
• Specialty information that reflects what was entered years ago, not current practice.
• License standing that the federal record does not carry, so a lapsed license is not visible to a directory that relies on the federal record alone.

What the records show in California

On September 27, 2026, Guide.MD matched the federal NPI registry (NPPES) against the licensee files that the California Department of Consumer Affairs publishes for its health licensing boards. A federal record was tied to a state license when the license number on the federal record matched the board's number and the surnames agreed, or, failing that, when the full name was unique on both sides at the same ZIP code and the profession was consistent. Nothing looser was accepted. These are measurements of the public records as published, not statements about any individual.

Method and date: Guide.MD; federal NPPES records and California Department of Consumer Affairs public licensee files as held on September 27, 2026; matching rules as described above. Counts are of records, not people, except where stated.

Five readings of the table matter for policy.

First, the two systems record different things about place. The state holds an address of record; the federal registry holds a practice address. They name different cities for about half of the clinicians they share. That is not evidence that either address is out of date. It does mean that a directory or AI system which treats a board address as a practice location will often be wrong.

Second, the federal record is silent on license standing. 7,656 physicians who have an NPI record with a California practice address hold a Medical Board license that is delinquent. Many of these are likely to be physicians who have retired or left the state. The point is that the federal record still lists them in California and gives no indication of their license status.

Third, the federal identifier covers only part of the licensed workforce. About nine in ten California registered nurse licenses could not be tied to an NPI record.

Fourth, many federal records have not been touched in years. Nearly half of the federal records for California clinicians show no update in five years, and about a quarter show none in ten. An old record is not necessarily wrong, since a clinician who has not moved has nothing to change. It does mean that no one has confirmed the record in that time.

Fifth, duplicates inside the federal registry are rare; errors arise when sources are joined. Fewer than one federal record in a thousand shares a full name and ZIP code with another. Guide.MD's own matching supplied an example of what can go wrong at the join. A rule that treated one name at one ZIP code as one person combined 435 nurses into 382 profiles. Most of them shared a single surname that is common in several Central Valley cities. The rule was tightened on September 27, 2026 to require a distinct license type or middle initial. Name-based matching fails more often for communities in which many people share a name, and an automated system joining the same records would make the same error without anyone noticing.

For comparison, a federal review of Medicare Advantage online directories conducted from November 2017 to July 2018 examined 5,602 providers at 10,504 locations across 52 plans and found that 48.74 percent of listed locations had at least one inaccuracy. That review measured something different from the table above: it tested directory entries by contacting the listed offices.

3. Why AI raises the stakes

A wrong directory entry used to cost a patient a phone call. AI systems now read the same records and present a confident summary of a clinician: specialty, location, affiliations, sometimes reputation. Three things change:

• Errors travel. One stale record can be repeated across many AI products.
• Errors are harder to see. A summary does not show which source it came from, or that the source holds a mailing address and not a clinic.
• Clinicians cannot easily correct them. There is no standard route for a physician to fix how an AI system describes them.

4. What California has already done

• AB 489 (Bonta; Chapter 615, Statutes of 2025) extends existing bans on misleading health care titles to AI systems, so that an AI product may not imply that its advice comes from a licensed professional. Licensing boards enforce it.
• SB 660 (Menjivar; Chapter 325, Statutes of 2025) concerns the Data Exchange Framework, which is now administered by the Department of Health Care Access and Information.
• SB 503 (Weber Pierson; Chapter 857, Statutes of 2026) requires developers and deployers of clinical decision-support systems to identify, document, and monitor risks of biased impacts, starting January 1, 2027.

These measures address what AI systems may claim and how clinical tools are monitored. None of them addresses the accuracy of the provider records that AI systems read.

5. The link to health data exchange

The Data Exchange Framework depends on knowing who its participants are. Its Stakeholder Advisory Committee is working on how to identify which organizations are required to sign the data sharing agreement and how to hold them accountable. That is a provider-matching task: the same physician organization can appear under different names and identifiers in licensing, billing, and plan data. The measurements above show the size of the task for individuals. Two sources the state already relies on name the same city for a clinician only about half the time, and one of them has no matching record for about 650,000 California license holders.

6. Implementation burden

Every new directory, attestation, or registry asks practices to enter the same facts again. Large systems absorb this with staff. Small practices do it after hours. Any new requirement should be weighed against that cost, and should reuse an existing identifier where one exists.

7. Options for California

• Reuse existing identifiers, together. Build on the NPI and the state license number instead of creating a new state provider ID, and treat the pair as the key. The license number is the identifier the state controls, and it is the one the federal record most often lacks.
• Publish data-quality measures for state-maintained provider files, such as completeness and match rates, so that users know how far to trust them. The table in Section 2 is an example of such a measure. The state can produce it from files it already publishes.
• Give clinicians a correction route. A clinician who finds an error in a state-maintained record should be able to fix it once and have the fix carried forward.
• Ask where AI summaries get their facts. When the state evaluates AI products that describe clinicians to patients, it is reasonable to require the source and date of the underlying record, and whether the location shown is a place of practice or an address of record.

These are options for discussion. The first two follow from how the existing data is structured. The last two would benefit from measured evidence on error rates before any requirement is written. Section 2 is a first measurement.