When Dr. Lena Okafor was named the first Chief AI Officer of a 14,000-person state health and human services department, the governor's chief of staff shook her hand and asked one question: "So what exactly does this job do?" She did not have a crisp answer. There was no professional license for it, no code of ethics she had sworn, no association that issued a credential, no agreed body of knowledge. She was a respected program executive who had been handed a title that the field itself had not finished inventing. Within a year she would be writing the job description that her successor, and a hundred peers across other agencies, would inherit.

That is the strange position of government AI leadership in this decade. The work is real and consequential. The profession, in the way medicine or accounting or civil engineering are professions, is still being built. This lesson is about doing that building deliberately rather than by accident, because the choices a handful of senior leaders make now will harden into the norms that govern the field for a generation. It is also about being honest, in public, regarding how much of that structure does not yet exist.

What makes something a profession, and why it matters here

A job becomes a profession when four things exist together. There is a shared body of knowledge that members are expected to master. There is an ethical code that members are accountable to, beyond just following the law. There are recognized pathways for entering and advancing. And there are institutions, usually associations, that maintain the standards and speak for the field. Doctors have all four. So do certified public accountants and professional engineers. A "growth hacker" has none of them, which is why nobody trusts the title.

The source material for this lesson states the classical definition slightly differently, and the difference is worth keeping. It defines a profession by specialized knowledge, a code of conduct, peer accountability, and a commitment to the public interest. Peer accountability is the element most often left out of the informal version, and it is the hardest one to build: it means practitioners judging each other's work against a standard, with consequences. Medicine, law, engineering and certified public accounting all followed this path, and each took decades of institutional argument to get there.

Government AI leadership today sits in the awkward middle. The body of knowledge is forming through frameworks like the National Institute of Standards and Technology AI Risk Management Framework, which is voluntary and non-binding, and through Office of Management and Budget guidance on agency AI use. Pathways are appearing, including the Chief AI Officer role that federal agencies were directed to designate. But the ethical code is thin, and the institutions are immature. When a profession lacks shared ethics and standards, every practitioner improvises, the worst outcomes set the public's expectations, and a single high-profile failure tars everyone.

A profession is what a field agrees to be accountable for before anyone forces it to. That is the whole argument for doing this work now rather than waiting. The alternative is that the accountability arrives anyway, in the form of a statute written after a failure by people who do not practice the craft, and the field inherits standards it had no hand in shaping.

The record that motivates a standard of practice

Consider what happened in Okafor's first six months. A vendor pitched a benefits-eligibility model that would, the sales deck claimed, cut case backlogs by 40 percent. A peer at a neighboring agency had bought a similar system, deployed it fast, and was now defending it before a legislative oversight committee after it wrongly flagged hundreds of eligible families for fraud review. The two leaders held the same title and made opposite choices. No shared professional standard said that one of them was practicing the craft correctly and the other was not. That absence is the problem this profession has to solve.

The public record supplies harder versions of the same lesson. The source reports that the Michigan MIDAS unemployment system produced more than forty thousand false fraud determinations, in a vendor-led deployment that ran without meaningful human oversight. It reports that the Internal Revenue Service identity-verification rollout with ID.me was reversed by Treasury after civil rights objections. It reports that the Dutch childcare benefits algorithm contributed to the fall of a national government after twenty six thousand families were wrongly accused of fraud. Each of these is described in the source as a professional failure in the making, not merely a technical one.

Two more cases anchor the legal edge of the field. COMPAS, the criminal risk assessment tool, remains the standing example of a system whose fairness properties were contested long after deployment. Houston Federation of Teachers v. HISD is the case that shows courts will hold an agency accountable when an algorithmic decision arrives without the process the people it acts on are owed. Neither case turned on whether the model was accurate in the aggregate. Both turned on whether anyone could explain, contest, or answer for a specific decision about a specific person.

The counter-example matters as much. In Allegheny County, the source describes a Department of Human Services data leader exercising professional judgment: publishing validation work, engaging the affected communities, running the model in shadow mode before it decided anything, and preserving human decision authority at the end of the process. Nothing in that sequence required a credential. It required someone who believed those steps were obligatory rather than optional, which is precisely what a professional standard encodes.

The role structure you are joining

Government AI leadership is not a single seat, and one of the first professional skills is knowing which chair you are in. OMB Memorandum M-24-10, Advancing Governance, Innovation, and Risk Management for Agency Use of Artificial Intelligence, issued in 2024, gave the field concrete positions. Covered agencies designate a Chief AI Officer. Agencies maintain AI use case inventories. Rights-impacting and safety-impacting systems must follow minimum practices, which the source lists as pre-deployment testing, impact assessment, ongoing monitoring, human oversight, public notice, and operator training.

That role sits inside a crowded governance structure that predates it. The Chief Data Officer is required by the Foundations for Evidence-Based Policymaking Act. The Chief Information Officer traces to the Clinger-Cohen Act. The Senior Agency Official for Privacy owns privacy compliance, and the Chief Information Security Officer is required under FISMA. An agency AI governance board joins an existing risk management council and data governance council. A senior responsible official for AI may sit alongside all of these. The profession's boundary problem is not abstract: it is deciding what an AI leader owns that none of these other officers already own.

The honest answer is that the AI leader owns the integration and the residual. Technical staff own model quality. Counsel owns legality. Privacy owns the privacy assessment. Security owns the authorization. Nobody else owns the question of whether the whole assembly should exist, whether the people it decides about can contest it, and whether the agency can explain it in public. That is the professional territory, and it is also why the role cannot be discharged by any single specialist discipline.

The emerging competency model

The source sets out a competency model with seven pillars, and enumerating them is the most useful self-assessment available while formal credentialing is unsettled. I counted seven items in the source's own list. Read each as a question about what you can currently do without help.

The model is useful precisely because it is unflattering. Most people arriving in these roles are strong in two or three pillars and weak in the rest, and the weak ones are usually procurement and workforce stewardship, because those are the ones no technical training covers. Self-assessment against the seven, followed by a deliberate plan to close the two largest gaps, is more valuable than any credential currently available, and it costs nothing but honesty.

Credentials, standards, and what they actually establish

The credentialing ecosystem for this work is consolidating but unfinished. The source points to established credentials in adjacent fields as the comparison set: ISACA's CGEIT and CISM in technology governance, the Project Management Professional credential, the certified public accountant qualification, bar admission for lawyers, and the Certified Government Financial Manager designation in public finance. It notes that professional bodies including ISC2 and the AICPA have been building AI assurance credentials, and that mid-career university programs at schools of government and public policy now serve the same market.

Be careful about what any of this establishes. A credential attests that a person passed an examination and met an experience requirement at a point in time. It is evidence about the holder, not about any system the holder approves. The same caution applies to management system certification: certification to ISO/IEC 42001 attests to a conforming management system, sampled across a declared scope at a point in time. It is not proof that any given AI system is safe, lawful or fit for its purpose. Leaders who cite certification as an answer to a question about a specific deployment are changing the subject, usually without meaning to.

What the international standard does supply is a management system foundation for AI comparable to what the information security standard supplies for security, which is a real contribution to a young field. It gives an agency a vocabulary for roles, controls, documentation and review that is recognized outside government, and it makes an agency's governance legible to auditors and partners who have never worked in the public sector. That is worth having. It is simply a different thing from an assurance that a model behaves correctly for the people it decides about.

Standards work is the more durable contribution, and it is a different activity from holding a credential. Two lessons in this program treat it directly rather than in passing: Contributing to Standards Bodies (NIST, ISO, IEEE, OECD) and AI Standards Development Participation. If your interest in the profession is in shaping the body of knowledge rather than consuming it, that is where the mechanics live. The relevant point here is simply that a field with no members in the rooms where standards are drafted will be governed by standards written for someone else's context.

The ethical code, which is the part that is missing

Ethics is not a supplement to this profession. The source treats it as constitutive, and that framing is right: a field that cannot say what it will refuse to do has no standard at all. The source proposes a code requiring practitioners to protect rights, prevent harm, be transparent, engage communities, acknowledge uncertainty, and refuse unsafe deployment. That is six duties, counted from the source's own list, and the last two are the ones that distinguish a profession from a job.

Government adds duties that private-sector AI roles do not carry. You serve the public, not shareholders. You are bound by due process, equal protection, transparency, and the principle that the burden of a mistake should not fall on the least powerful person in the transaction. The available foundations are all non-binding: the Blueprint for an AI Bill of Rights is a non-binding blueprint published by the White House Office of Science and Technology Policy, not a statute; the IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems supplies international grounding; the ACM Code of Ethics anchors the computing side. Agency-specific codes of conduct add the enforceable layer.

Because none of these bind you personally, a written commitment does. Okafor settled on one sentence that a legislator, a union steward, and a frontline caseworker could all understand: "I am accountable for ensuring that every automated decision this department makes about a resident is lawful, explainable, contestable, and fair." That sentence did more to orient her team than any org chart, because it told people what she would spend her authority on. It is also falsifiable, which is the property that makes an ethical commitment worth writing down.

Peer accountability, and the honest gap

Peer accountability is what separates a profession from a job market. In medicine and law, licensure boards investigate misconduct, revoke credentials, and publish standards that practitioners can be measured against. Government AI leadership has no licensure board and no revocation mechanism. Saying otherwise would be flattering the field. What exists instead, according to the source, is a set of partial substitutes: oversight by the Government Accountability Office and agency Inspectors General, ordinary agency disciplinary procedures, bar disciplinary rules for attorneys who also serve as AI officials, and the informal norms of communities of practice.

Those substitutes are real but weak, and they share a defect: none of them is triggered by bad professional judgment as such. An audit is triggered by a finding, a disciplinary process by misconduct, a community norm by nothing at all. A practitioner can deploy a system that was never tested for disparate impact, be entirely within agency rules, and face no professional consequence whatever. The source expects these mechanisms to consolidate over time into something more formal, and argues that members of the profession should champion that consolidation rather than wait for it. That is a claim about direction, not a description of current practice.

The version of peer accountability available to you now is voluntary and local. Okafor's version was a peer council across seven agencies where AI leaders reviewed each other's high-impact systems before launch, plus a shared incident log so that one agency's failure became everyone's lesson. Voluntary review has an obvious limit, which is that a council with no authority cannot stop anything. What it can do is make it socially expensive to skip a step, and remove the excuse that nobody knew.

Institutional homes for that kind of peer work already exist and are worth joining rather than duplicating. The source points to the Partnership for Public Service and its leadership programs, and to the GSA AI Community of Practice, as the cross-agency venues where practitioners already meet. It also points to the Department of Defense's Joint Artificial Intelligence Center, which has since been reorganized and is cited here as a historical example of a departmental hub rather than a current destination. Check the current status of any body before you route your team's development through it, because this landscape has changed repeatedly.

Pathways into the field, and continuing education

The field needs to be enterable, and a profession that recruits only from one background will reproduce that background's blind spots. The source lists graduate study in public administration, public policy, computer science or data science; federal and state rotational and fellowship programs; community college technician pathways through registered apprenticeship; and mid-career lateral entry from adjacent professions. I have deliberately not named the individual federal rotational programs the source lists, because several have been reorganized or discontinued and this lesson should not teach a program as current that a reader cannot apply to. The mechanism, which is a rotation or a fellowship that lets a technically capable person learn the institution, is what matters and it survives any particular program's name.

Continuing education is the second half of a pathway and the one agencies forget to fund. The source argues that an annual continuing education requirement of about forty hours, comparable to what other professions require, would be reasonable. Treat that as the source's proposal rather than an existing obligation, because no body currently imposes it on government AI leaders. The underlying point is harder to argue with: the technical and legal ground under this role moves fast enough that a leader who stopped learning three years ago is now giving advice from a world that no longer exists.

Below your own role, define the rungs. Analysts and program managers can see a route into AI leadership only if someone writes down what each step requires a person to know and to demonstrate. That documentation is unglamorous and it is the single most effective retention tool a senior leader has, because the alternative to a visible internal path is watching capable people leave public service to find one.

A usable artifact: the AI leader's professional charter

The single most useful thing a senior government AI leader can produce is a one-page professional charter that names the duties, the ethical commitments, and the competencies of the role. It is the document Okafor wished had existed on day one. Below is a template you can adapt for your agency. Fill it in with your own context, then publish it, because a charter no one can see governs no one.

Treat the charter as a living instrument, reviewed once a year. Walk it through three audiences before you finalize it: your agency's general counsel, who will test the legal duties; a frontline supervisor, who will tell you whether the commitments survive contact with reality; and a resident advocate or community group, who will tell you whether the public would recognize this as fair. Endorsement from all three does not make the charter correct, but it removes the three most common reasons a charter turns out to be unworkable, and it forces the document to be written in language all three can read.

Building the field, not just the career

The leaders who define a profession are remembered differently from those who merely excel within it. The distinction is whether you left structures behind. Okafor's lasting contribution was not the eligibility model she chose not to buy. It was the peer council she convened, the shared incident log her group kept, and the mentoring track she opened so that program analysts could become AI leaders without leaving public service for industry. Those structures are the difference between a field that learns and a field that repeats.

None of them required new legislation or a national mandate. They required a senior person deciding that her job included building the rungs of the ladder below her and the institutions beside her. The source's version of this obligation is a personal professional development plan: a twelve-month plan naming your competency gaps, your continuing education targets, your community of practice involvement, and a short list of peers to whom you will be accountable over the coming year. Naming the peers is the part people skip, and it is the part that makes the rest more than a wish list.

Professions do not form by decree. They form through the accumulation of committed practitioners who hold themselves and each other to a standard before anyone requires it, and who write down what they learned so the next person does not start from zero. That work is available to you this quarter, at no budgetary cost, and it does not wait for anyone's permission.

Anti-Patterns

• Certification treated as competence. Citing a credential or a management system certificate in answer to a question about whether a specific system is safe. Certification to a management standard attests to a conforming management system, sampled across a declared scope at a point in time; it establishes nothing about the deployment being asked about.
• The charter as absolution. Publishing a professional charter and never referencing it in a decision. A charter that has not caused you to refuse something, or to slow something down, is a statement of aspiration that has cost you nothing and protected no one.
• The designated title with no authority. Naming a Chief AI Officer to satisfy a guidance requirement while leaving the role no veto, no budget line, and no route to the agency head. The designation is a decision, not a continuing control, and an org chart entry does not create standing.
• Importing another profession's ethics wholesale. Adopting a computing or engineering code of conduct without adding the public-service duties that make government different: due process, equal protection, and the duty not to place the cost of a mistake on the least powerful person in the transaction.
• The peer council as rubber stamp. Convening a review body that has never asked a project to change. Voluntary peer review buys visibility and standing, not outcomes, and a council that always approves is providing cover rather than accountability.
• Waiting for the profession to be finished. Deferring an ethics code, a competency baseline or a documented pathway until a national body issues one. Every norm this field will have is currently being set by somebody's improvisation, and abstention is a contribution too.
• Treating a failure case as someone else's category error. Reading MIDAS or the Dutch benefits scandal as a technical failure by less capable people. In each case the missing item was a professional judgment that was available to the leaders at the time and was not exercised.
• Recruiting only your own background. Building an entry pathway that admits data scientists and no one else, then discovering that the competency pillars nobody on the team holds are procurement and workforce stewardship.

Practice Prompts

• Score yourself honestly against the seven competency pillars on a three-point scale, then ask a peer who has watched you work to score you on the same sheet. Discuss only the pillars where you disagree by more than one point.
• Draft the one sentence that names what you are accountable for, in language a legislator, a union steward, and a frontline caseworker would each understand without translation. Test it by reading it aloud to one person from each group.
• Write your refusal list: three specific deployments you would decline to approve, and the reason for each. If you cannot produce three, you do not yet have an ethical code, you have a preference.
• Take the professional charter template and complete it for your own role this week. Then walk it through general counsel, a frontline supervisor, and a community advocate, and record what each of them changed.
• Pick one failure case from the record, MIDAS, the identity verification reversal, the Dutch childcare scandal, COMPAS, or the Houston case, and write the specific moment where a professional judgment would have changed the outcome. Then find the equivalent moment in one of your own systems.
• Name the three peers you will be accountable to over the next twelve months, and tell them. Agree what you will show each other and how often.
• Document the rung immediately below your role: what a person must know and demonstrate to be ready for it. Give it to someone who wants that job.

Reflection

Think about the last consequential AI decision you made or witnessed, and ask what standard it was measured against. Was there one, or did the decision rest on the judgment and the nerve of whoever happened to be in the chair? If the standard existed only in that person's head, consider what would have happened had a different person been in the chair that day, with a different tolerance for risk and a different relationship with the vendor. That gap between the person and the standard is exactly the space a profession occupies, and the size of the gap in your own organization is the measure of how much of this work is still ahead of you.

Then ask the harder question, which is about your own accountability rather than the field's. Nobody can currently revoke your ability to practice this work. No board will review your judgment, and no examination stands between you and the next system you approve. Given that, what holds you to a standard? A written commitment you have published, a group of peers who have permission to challenge you, and a habit of documenting your reasoning are the only mechanisms available, and all three are things you have to choose. The professions that came before this one were built by people who chose them before anyone made them.

Glossary

• Profession. A field defined by specialized knowledge, a code of conduct, peer accountability, and a commitment to the public interest, maintained by institutions that outlast individual practitioners.
• Peer accountability. Practitioners judging each other's work against a shared standard, with consequences; the element of professional structure most conspicuously absent in government AI leadership.
• Chief AI Officer. The senior agency role for AI governance and coordination that federal agencies were directed to designate under OMB guidance issued in 2024.
• Professional charter. A one-page statement of a role's mission, ethical commitments, decision authority, required competencies, accountability, and succession, published so that others can hold the holder to it.
• Competency model. The seven-pillar description of what this role requires: technical fluency, policy literacy, ethics, risk and compliance, procurement, workforce stewardship, and communication.
• Rights-impacting AI. An AI system whose output affects a person's civil rights, civil liberties, or access to services, and which federal guidance subjects to minimum practices including testing, impact assessment, monitoring, human oversight, notice, and training.
• Management system certification. Third-party attestation that an organization's management system conforms to a standard, sampled across a declared scope at a point in time; not an attestation about any particular system's safety or lawfulness.
• Shadow mode. Running a model alongside the existing process so its outputs can be compared against human decisions without affecting anyone, before it is given any decision authority.
• Licensure. A legal permission to practice, granted and revocable by a board, which government AI leadership does not currently have in any jurisdiction.
• Community of practice. A voluntary cross-agency group of practitioners sharing methods, templates and incident experience; the main institutional form currently available to this field.
• Continuing education. A recurring requirement to maintain currency in a field, proposed by the source at about forty hours a year for this profession but not currently imposed by any body.

Related Lessons

• Contributing to Standards Bodies (NIST, ISO, IEEE, OECD) covers the mechanics of participating in the standards work this lesson only points at.
• AI Standards Development Participation treats standards involvement as a sustained agency capability rather than an individual activity.
• Mentoring Next-Generation Leaders builds the pathway below your role into a designed practice.
• Personal Leadership: Leading Through Complexity supplies the judgment disciplines that a competency list can name but not teach.
• Building the GOVT.CLUB Community is the community-of-practice side of the same institution-building argument.
• Ethics Boards and Advisory Structures turns an ethical code into a standing decision structure.
• Oversight Mechanisms: IG, GAO, Congress details the external accountability that currently substitutes for peer review.
• ISO 42001: AI Management System Design explains what a conforming management system contains and what certification does and does not cover.
• OMB M-24-10 Deep Dive: Full Implementation works through the role designations, inventories and minimum practices summarized here.
• Building Institutional Knowledge addresses how the documentation that makes a profession legible actually gets produced.

Closing

The uncomfortable fact about this lesson is how little of the structure it describes actually exists. There is no board, no licence, no revocation, no examination, and no enforceable code. A reader who wanted to be reassured that the field has grown up should not be. What the field does have is a small number of senior people in a position to decide what the norms will be, at a moment when those norms are still soft enough to shape. That window closes, and it usually closes after a failure large enough to make legislators write the rules instead.

Okafor could not answer the chief of staff's question on her first day. A year later she could, not because anyone had defined the job for her, but because she had written the definition down, tested it against the people it affected, published it, and invited peers to hold her to it. That is the whole method. It scales from one leader to a field only if enough leaders do it and share what they wrote, which is why the least glamorous parts of this lesson, the charter, the incident log, the documented rung below you, are the parts that actually build the profession.

Key Takeaways

• A profession needs four elements. Specialized knowledge, a code of conduct, peer accountability and a public-interest commitment; government AI leadership has the first forming and the third essentially absent.
• The failure record is the argument. MIDAS, the identity verification reversal, the Dutch childcare scandal, COMPAS and the Houston case each turned on a professional judgment that was available and not exercised.
• Know which chair you are in. The AI leader's territory is the integration and the residual: whether the system should exist, whether it can be contested, and whether the agency can explain it publicly.
• Seven competency pillars. Technical fluency, policy literacy, ethics, risk and compliance, procurement, workforce stewardship and communication; most people arrive strong in two or three.
• Credentials attest to people, not systems. Management system certification covers a declared scope at a point in time and is never an answer to a question about whether a particular deployment is safe or lawful.
• Ethics is constitutive, not supplementary. A code that includes acknowledging uncertainty and refusing unsafe deployment is what distinguishes this from a job, and every available foundation for it is non-binding.
• Peer accountability is the honest gap. There is no licensure board and no revocation; audits, discipline and community norms are partial substitutes that bad judgment alone does not trigger.
• Write a one-page professional charter. Name the mission, ethical commitments, decision authority, competencies, accountability and succession, then publish it and test it with counsel, a frontline supervisor and a community advocate.
• Build the rung below you. A documented entry path is the most effective retention tool a senior leader has, and the mechanism matters more than any particular program's name.
• Legacy is structural. Peer councils, shared incident logs and mentoring tracks turn personal excellence into a field that learns rather than one that repeats.

Frequently Asked Questions

Is government AI leadership actually a profession yet? Not by the definition this lesson uses. The body of knowledge is forming, entry pathways are appearing, and role designations exist in federal guidance. But there is no enforceable ethical code binding practitioners, no licensure, and no peer body that can review or revoke anyone's practice. Calling it a profession today is a statement about trajectory. The practical consequence is that nothing external holds you to a standard, so the standard has to be one you adopt and publish.

Should I pursue a certification to establish my credibility? A credential can be worth holding for what it teaches and for the doors it opens, and the source points to established comparators in adjacent fields. Be clear about the limit, though. A credential attests that you passed an examination and met an experience requirement at a point in time. It says nothing about any system you approve, and citing it in answer to a question about a deployment will read as evasion to an auditor or a legislator. Your published reasoning about a specific system is worth more than any letters after your name.

Does certifying our AI management system to a standard mean our systems are safe? No, and this confusion is common enough to be worth stating flatly. Certification to a management system standard attests that the management system conforms to the standard, sampled across a declared scope at a point in time. It does not test any individual model, it does not evaluate a particular decision, and a certified organization can still deploy a system that harms people. The certificate tells you an organization has a process. It does not tell you the process was followed here.

What distinguishes this role from the Chief Data Officer or the CIO? Those roles have statutory or memorandum-based mandates of their own, and they own data governance and information technology respectively. The AI leader's distinct territory is the integration across all of them plus the residual questions nobody else owns: whether an automated decision system should exist at all, whether the people it acts on can contest it, and whether the agency can explain it in public. In practice you will spend much of your time convening the other officers rather than duplicating them.

Our agency has no ethics code for AI. Where do I start? Start with a refusal list rather than a document. Write down three specific deployments you would decline to approve and why, share it with your leadership, and see whether it survives. Codes that begin as abstract value statements tend not to change any decision; codes that begin as refusals tend to. The available external foundations, including the non-binding federal blueprint and the professional computing codes, are useful for language and legitimacy, but none of them binds you, so the binding has to be local and public.

I am mid-career and not technical. Is this field open to me? Yes, and the competency model is the reason. Two of the seven pillars, procurement and workforce stewardship, are exactly where technically trained entrants are weakest, and policy literacy and communication are two more. A program manager or a policy lead who builds enough technical fluency to interrogate a vendor is often better positioned than a data scientist who has never run an acquisition or a bargaining consultation. The lateral entry path the source describes is real; what it requires is closing the technical gap deliberately rather than deferring to whoever in the room sounds most certain.