Artificial intelligence is now deeply embedded in professional work, from legal research and financial analysis to education, healthcare, hiring, and corporate decision-making. As its use expands, the most important question is no longer whether professionals should use these systems, but how responsibility should be handled when their output influences real decisions.
California’s recently passed SB 574 brings that question into sharp focus. The legislation addresses lawyers’ use of generative AI by requiring reasonable steps to verify generated content, including obtaining legal citations, restricting the delegation of legal practice to AI systems, addressing disclosure in court filings, and limiting the use of confidential or nonpublic information. Although the bill is specifically directed at the legal profession, the principle behind it reaches much further.
What California is really confronting is the boundary between assistance and accountability!
Professionals have always relied on tools, research systems, software, databases, and other forms of technological support. What makes generative AI different is the extent to which it can produce outputs that look complete, confident, and authoritative even when they are inaccurate, incomplete, or based on flawed reasoning. That creates a new kind of professional risk because the output may look ready for use long before it is ready to be trusted.
This distinction matters because producing an answer is not the same as establishing that the answer is reliable. A legal citation can look legitimate and still be fabricated. A financial recommendation can sound persuasive and still rest on incomplete assumptions. A medical summary can appear coherent while omitting an important detail. A hiring recommendation can appear objective while reflecting problematic patterns in the underlying data.
The deeper issue, therefore, is not simply whether these systems make mistakes. Every professional tool has limitations. The real question of governance is what happens between the moment a system produces an output and the moment a human decides to rely on it.
That point of reliance deserves far more attention. It is also central to the SAFER AI™ Protocol framework, which I developed around five areas: Scope, Authority, Failure Awareness, Evidence, and Record. The purpose of the framework is not to slow professionals down or create unnecessary layers of approval around everyday work. It is to make responsibility more deliberate at the point where the generated information begins to influence judgment.
Scope asks whether the system is being used for an appropriate purpose. Authority asks who is responsible for accepting, rejecting, or escalating the output. Failure Awareness requires professionals to consider how the system could be wrong in that context. Evidence asks what level of verification should be required before the output is trusted. Record addresses whether the reasoning and verification behind a consequential decision should be documented. These questions become increasingly important as the consequences of a decision increase.
Using a generative system to brainstorm a marketing slogan does not carry the same level of risk as using one to support a legal filing, evaluate a borrower, recommend a medical intervention, assess a student, screen an employee, or advise an executive on a high-impact decision. Responsible use should therefore be proportional. The greater the potential consequence, the greater the need for verification, evidence, and accountable human judgment. This is why California’s action should matter well beyond the legal profession.
Healthcare organizations should be asking what clinicians must verify before relying on generated recommendations. Financial institutions should be asking what evidence is necessary before machine-generated analysis influences lending, investment, or risk decisions. Universities should be asking what faculty and students remain responsible for when generative systems contribute to academic work. Corporate leaders should be asking where technological assistance ends and accountable authority begins.
The common thread across all these professions is that responsibility cannot disappear simply because part of the work was delegated to a system.
That may be one of the most important lessons emerging from the current debate over the use of professional AI. Organizations have spent years asking what these systems can do, how much time they can save, and how quickly they can increase productivity. Those questions will continue to matter, but they are no longer enough.
The more consequential question is whether organizations have established clear expectations for the point at which a person decides that a generated output is sufficiently reliable to influence action. This is where responsible practice becomes real.
California’s proposed requirements for lawyers may eventually be remembered as part of a broader shift in professional standards, one in which the use of advanced technology does not reduce human responsibility but instead makes the boundaries of that responsibility more important.
The future of responsible AI will not depend only on better systems. It will also depend on professionals who understand when to question an output, when to verify it, when to escalate it, and when not to rely on it at all.
Technology can support professional judgment, but it should never eliminate accountability.
