The proposed regulatory approach
Regulating the variable that matters.
Artificial intelligence is becoming capable of drafting substantial portions of Qualified Medical Evaluator (QME) reports. As a result, California's workers' compensation system is approaching a question that many professional fields will soon face:
How should accountability be preserved when a machine can generate the text that a licensed professional ultimately reviews and signs?
Existing law was developed for a different technological era. The rules governing report preparation were designed primarily to prevent the delegation of professional judgment to another person. Today, however, text generation and professional judgment are no longer necessarily performed by the same actor.
This proposal argues that regulation should focus on the outcomes the system is trying to protect rather than on the provenance of the text itself.
The relevant questions are:
- Are the facts accurate and traceable to the record?
- Have the required legal and evaluative frameworks been addressed?
- Does the report reflect the physician's independent professional judgment?
- Can accountability be meaningfully enforced?
These questions are ultimately what determine whether a report is fair, reliable, and trustworthy.
The original purpose of the rule
Labor Code §4628 requires that a physician personally examine the applicant, take the history, review the records, and "compose and draft the conclusions" of the report.
When enacted, this requirement served an important purpose. Drafting assistance generally meant delegation to another person. Requiring the physician to draft the report personally helped ensure that the conclusions reflected the physician's expertise and judgment rather than that of an undisclosed intermediary.
In practice, authorship functioned as a proxy for accountability.
The physician who wrote the report was presumed to be the physician who reasoned through the case. The physician who reasoned through the case was presumed to be accountable for the result.
For many years, that proxy worked reasonably well.
Why the proxy is breaking down
Artificial intelligence changes the relationship between authorship and judgment.
A drafting system can generate narrative text without exercising medical judgment. A physician can independently evaluate and adopt conclusions that were initially drafted by software. Conversely, a physician can sign prose that appears professionally written without carefully evaluating the reasoning that produced it.
As a result, the question of who generated the text no longer reliably answers the question the law is trying to ask.
A report may be trustworthy even if portions were drafted by software. A report may be untrustworthy even if every word was personally typed by the physician.
The policy objective has not changed. The relationship between the proxy and the objective has.
A different regulatory approach
Rather than regulating authorship as an end in itself, this proposal explores a framework that regulates the characteristics most closely associated with reliable medical-legal evaluations. Under this approach:
- Physicians remain responsible for their conclusions.
- Physicians continue to attest that the report reflects their independent professional judgment.
- AI-assisted drafting may be permitted under defined conditions.
- Systems used for drafting are evaluated against objective standards for factual reliability and framework compliance.
- Accountability remains with the physician rather than with the technology.
The objective is not to reduce physician responsibility. The objective is to make physician responsibility more meaningful by ensuring that the material being reviewed and signed can be independently evaluated for quality and reliability.
An illustrative statutory framework
This proposal does not advocate for specific legislative language. However, it is useful to illustrate what a shift from authorship-based regulation to accountability-based regulation might look like.
Illustration — not proposed legislative text
To sign the report, the physician must have personally examined the applicant, taken the history, reviewed the medical records, and composed and drafted the conclusions of the report.
Assistance from others is allowed only for taking the history and reviewing records, and only if disclosed. The conclusions may not be delegated.
Regulates who performed each act. The fourth requirement is the clause that blocks a tool from drafting analytical sections.
To sign the report, the physician must have personally examined the applicant and composed and drafted independently formed and adopted the conclusions as their own.
The physician may use an automated drafting tool to prepare any portion of the report — including analytical sections — as input to the physician's review, provided that: its use is disclosed; the physician verifies the factual content against the record; and the tool conforms to a recognized evaluation standard administered by an independent body.
The physician declares under penalty of perjury that the conclusions reflect their own independent judgment and that every fact is traceable to the record.
Regulates whether the judgment is the physician's and whether every fact is verifiable. Provenance of the text stops mattering.
In such a framework, the physician would remain personally responsible for the report's conclusions and would continue to certify that those conclusions reflect independent professional judgment.
The use of AI-assisted drafting could be permitted if factual assertions are verifiable, required evaluative frameworks are addressed, and the system satisfies an established evaluation standard administered by an independent body.
The focus would move from who generated the words to whether the resulting report satisfies measurable standards of reliability and accountability.
Potential benefits
A framework of this type offers several potential advantages.
- It directly targets the factors most closely associated with trustworthy evaluations: factual accuracy, framework compliance, and professional accountability.
- It may improve protection for parties who lack the resources to identify and challenge factual errors through litigation or deposition. Requiring upstream validation of report quality reduces reliance on downstream error correction.
- It may help address evaluator capacity constraints by allowing technology to assist with drafting while preserving physician responsibility for the final product. Reduced delays can improve access to evaluations, treatment decisions, and benefits determinations.
- It may encourage transparent and auditable uses of AI rather than undisclosed and unregulated adoption occurring outside any formal oversight framework.
Risks and tradeoffs
The proposal also raises important concerns.
The most significant risk is automation bias. As drafting systems become more capable, physicians may become more likely to accept generated content without sufficient scrutiny. Even highly accurate systems can create risk if users become overly reliant on them.
The proposal depends on credible governance. Any framework built around certification or evaluation standards is only as strong as the institutions responsible for administering those standards. Weak oversight could undermine the protections the framework is intended to create.
Policymakers must also consider public confidence, implementation costs, unequal access to compliant technologies during a transition period, and the replacement of a relatively simple authorship rule with a more complex performance-based framework.
The core claim
The central claim of this proposal is straightforward.
The policy question is not who drafted the words. It is whether the physician's judgment is genuinely their own, whether the facts are reliable, and whether the report can be trusted.
Technology changes how reports are produced. It does not change the need for physician accountability.
The remaining sections describe a possible framework for measuring, certifying, and governing that accountability in an era of AI-assisted report generation.