The Machine That Writes and the Machine That Chooses

Consider a client who does not exist, although she could walk into any cross-border practice this month.

She moved to the United States on a green card more than twenty years ago, and moved back to Israel with her family more than a decade ago. She kept the card. She kept a savings account at an American bank. Last spring, on advice that she is a resident of Israel under the treaty between the two countries, she filed a nonresident return with the disclosure that a treaty position requires.

In October a letter arrives. A matching program has compared the interest her bank reported with the return she filed, has not found the interest there, and proposes additional tax, with an accuracy-related penalty computed by the same software.

The interest is not missing. If the treaty’s tie-breaker makes her a resident of Israel, she is taxed in the United States as a nonresident, and American law does not tax a nonresident on the interest from an ordinary bank deposit. The return she filed was built on exactly that premise. Whether she is such a resident is a question the treaty answers through a sequence of tests about home, ties, and habit, and it turns on facts no matching program holds. And the position her return already took is not free. For someone who has held a green card as long as she has, claiming residence abroad under a treaty can end her status as a United States resident for tax purposes, with consequences far larger than the letter. Whether anyone weighed that when the return was filed is not a question the letter asks.

So there are two machines in her file. One decided, in effect, that she should receive a letter. The other could help write the answer. The question I want to ask is where each of them puts a person, and it turns out the IRS has already answered it for its own machines, in writing.

---

What the Agency Wrote About Itself

In February the IRS issued, and in August revised, a section of its Internal Revenue Manual titled IRS Policy for Artificial Intelligence (AI) Governance. It applies to all IRS employees, contractors and vendors. Its stated purpose is the trustworthy use of AI in a way that fosters public trust, and the principles it adopts include that AI be traceable, with human roles clearly defined, and accountable, with safeguards that are monitored and audited.

It is a serious document, and I want to give it its due before I say anything about where it draws its lines. For uses that significantly affect people, it requires testing before deployment, documented impact assessments, ongoing monitoring, human oversight, and remedies or appeals. It requires a named official to accept the risk of each use. It warns employees, in plain words, that language models can produce output that is coherent, plausible, and untrue, and tells them to verify it and not to become reliant on it. On paper, it asks a great deal of the agency’s own people.

It is also an internal manual, and that matters for everything that follows. It binds the agency’s people. It does not give a taxpayer anything to enforce; courts have held that the manual lacks the force of law and confers no rights, a point Frank Agostino makes plainly in the survey of AI in tax controversy he prepared for the UCLA tax controversy conference in October. The policy tells us how the agency intends to hold itself accountable. It does not tell a taxpayer what she can demand.

---

The Machine That Writes

The manual’s rules on generative AI hold no surprises. An employee may not use AI output in government business without direct human review to confirm that it is true, may not obscure that content came from AI, and may not use generative AI to make binding determinations about taxpayers’ rights without proper oversight; a violation can be grounds for removal from federal service. The rules for the profession arrived at the same place this year. In June the IRS’s Office of Professional Responsibility told practitioners that every AI-created document must be reviewed before it goes to a client or to the Service, and that the facts, citations and calculations in it must be verified. California, in a statute signed at the end of September and effective in January, bars lawyers from delegating the practice of law to generative AI and from filing in court a citation they have not personally verified. For text a machine writes, the agency and the profession now agree: a person on every output, and the review has to be real. The European regulator, as an earlier essay traced, reached the same standard for when a review counts.

---

The Machine That Chooses

The manual treats a second kind of machine differently, and this is the line I think most readers will walk past.

Some AI does not write anything for a person to read. It produces a score, a ranking, a selection. The manual presumes that AI which informs or influences whether a taxpayer will be audited, or which parts of a return, is high-impact, and it attaches the full set of requirements: testing, impact assessment, monitoring, oversight, remedies. But the person it names is named once. A use case enters operational use on the approval of an accountable official, whose name and approval date are recorded in the agency’s inventory of AI uses. The accountability the manual defines runs to the use case and its mission outcome, and the oversight it requires is oversight of the system; neither is attached to any particular return the use case selects. And whether a use counts as high-impact at all turns on whether the AI’s output serves as a principal basis for the decision, a determination an official must record and, where the official concludes a presumed use is not high-impact after all, an internal assurance team reviews.

In the language of the essay on the assembly line, this is the diverging line: a great many outputs, each one finished as far as the model is concerned. Every year the agency has to decide which few returns, of all the returns it receives, deserve a person’s attention. People still take part in that decision; Agostino notes that the employee who screens a scored return must read it as a whole before selecting it. But the accountability the manual defines attaches to the use case, not to the return.

Our client’s letter came from a different machine, and it may not be AI at all. The manual’s definition excludes systems whose behavior is defined only by rules people wrote. Matching programs and scheduled notice streams can run on exactly such rules, and where they do, the AI policy does not reach them. Where a person does appear in a matching program, Agostino describes the work as verifying one mismatch against one document. The question that decides her case, which country she lives in for the treaty’s purposes, is not part of that comparison.

For penalties, Congress made the choice explicitly: a penalty automatically calculated through electronic means needs no supervisor’s written approval. As Agostino sets out, the Tax Court applied that in 2019 to a penalty a correspondence program computed and no examiner ever reviewed, and a regulation now brings a person back in only when the taxpayer responds, challenges the penalty or the tax behind it, and an employee considers the response.

Follow her file in the ordinary course, and no one at the agency weighs the question that actually decides it until she, or someone she brings, puts it in front of them.

---

Why the Agency Is Probably Right

It would be easy to turn this into an indictment, and I think it would be wrong.

Imagine the alternative: a person signs every notice. At the scale of a national tax system, those signatures would come faster than anyone could read what they approve. That is precisely the cursory approval the European guidance refused to count as human review, and precisely the ratification without ownership that the essay on who has to be in the room argued satisfies the letter of accountability while missing its substance, lending a machine’s output a signature it did not earn. A signature on every file of a mass administration would not restore accountability. It would counterfeit it.

Automation also runs in the taxpayer’s favor, and not only in theory. This summer the IRS began applying penalty relief automatically, at processing, to taxpayers with a three-year history of filing and paying on time, before anyone asks for it. The National Taxpayer Advocate welcomed it as a major win, while warning that relief applied automatically in a year when the taxpayer had reasonable cause anyway can leave nothing for a later year when it is needed.

So I do not think the IRS put the person in the wrong place. If I had to design accountability for a machine that chooses among everyone, I am not sure I would design it differently. But the design has a consequence the agency’s own logic forces, and it is worth stating plainly.

---

If the Signature Lives in the System

If accountability is placed at the level of the system, then the system’s documentation is the signature. It is where the person attached to the machine puts their name to it.

In a report issued in September, the Treasury Inspector General for Tax Administration examined that signature. It identified seven deployed uses, presumed to be high-impact, that drew on the IRS’s compliance data warehouse, and reviewed the five used to select work. Two of the five had no documented impact assessment. Four of the five had no documentation of how the data behind them was tested. Program staff told the auditors the data checks had been done; the documentation varied from one use to the next. The IRS had built its own policy in February, and as of July, the report records, Treasury had not given it guidance on the minimum risk management practices federal policy required by April. The IRS agreed with both recommendations and said it would complete impact assessments for its deployed high-impact uses as of November.

The finding is narrower than it sounds, and I want to keep it that way. The inspector general did not test whether the data was accurate, and nothing in the report says any model selected the wrong returns. What it found is that the signature, in the form the agency’s own policy prescribes, was incomplete. That is a finding about paperwork. In a regime where the paperwork is the accountability, it is the finding that matters.

It is also a finding no taxpayer in our client’s position can use. The manual gives a taxpayer nothing to enforce, and whether a gap in an agency’s AI documentation supports any challenge in a collection hearing is, in Agostino’s careful account, untested, with the Tax Court’s most recent description of that hearing’s verification duty a narrow one. The signature at the level of the system exists, or will, in an inventory she cannot invoke. The only signature at the level of her own file is the one her representative brings.

Agostino closes the body of his survey with the observation that, on both sides of the table, “a person remains answerable on the record for what the machines produce.” I agree. The question this essay has been asking is which record. The agency’s record is its inventory. Hers is her file.

---

The Same Line, Inside Our Own System

Return to the client, and to the second machine in her file.

ktAi is a machine that writes. Its library of law holds the treaty’s tests for a person resident in both countries and Israel’s own definition of residence. The system releases a report drafted from that library to a client only after an adviser approves it, and approval is blocked while a revision request stands unresolved. By the IRS’s own taxonomy, that is the rule for machines that write, and it is where we put the gate.

Her file also shows what the gate is for. The rule that makes her treaty position expensive for a long-term green card holder is not yet in our library of law. Whatever a draft says about it today would rest on no authority the system can cite, and the person who has to catch that is the adviser who reads the draft before approving it. That is this essay’s argument in miniature: the person on the file is the one who knows what the library lacks.

The question of where to place the person comes up in one other part of our system, and I would rather name it than let the gate imply more than it covers. The gate stands on the release of a formal report. The working conversation that comes before a report is answered by the system directly, and the review of each answer belongs to the person reading it, which is where the IRS puts it for its own staff; the system adds no approval of its own, and nothing stops a user from carrying a conversation out of the system. That is a judgment about where a signature belongs. But I would not want a reader to infer, from the gate on the report, that every sentence the system writes has been signed. It has not.

---

One File at a Time

The IRS has placed the person where it can afford to: on every output a machine writes, and once on every machine that chooses. The profession places the person where the client needs one: on the file. Neither placement is a scandal. Both are decisions about the unit of accountability, made in writing and open to audit, and the inspector general has just shown what auditing the system-level one looks like.

The same inspector general has also shown where a choice lands. In March it reviewed a program in which the IRS used a machine-learning model to help choose large partnerships for examination. Of the thirty-six examinations that had closed, thirty-three ended with no change, and the inspector general counted that as a burden on taxpayers who had complied. It is a small sample from one program, and a selection is not a finding. But the cost of each of those thirty-three examinations was borne at the level of a file, by a taxpayer and whoever represented them, not by the model that helped choose it.

What follows is quieter. A taxpayer the machines have already processed gets a person on her own file only if someone brings one. That was always part of what representation was for; the machines have only made it easier to see. No policy manual can promise it, and no audit report can supply it. I supply it the way this profession always has, one file at a time.