Is the IRS using AI to pick returns?

The agency was unusually direct about it. IR-2023-166 describes improved technology and artificial intelligence helping compliance teams detect tax cheating, identify emerging compliance threats, and improve case selection so as to avoid burdening taxpayers with needless no-change audits.

It goes further and names the target: an expansion of a pilot focused on the largest partnerships, leveraging artificial intelligence. Selecting those returns, it says, came out of collaboration between data science and tax enforcement staff applying machine learning across partnership tax, general income tax and accounting, and international tax.

The agency also publishes Internal Revenue Manual 10.24.1, its policy for AI governance, covering how these systems are managed internally.

Read the target carefully

Almost every headline about this dropped the qualifier. The announced deployment was aimed at large partnerships, large corporations and high-income individuals, in segments the release describes as historically subject to limited examination coverage.

If your return is wages, a mortgage and a couple of 1099-INTs, you were not the subject of that announcement.

What has watched your return for decades

Long before anyone marketed AI, the IRS was matching documents. Every W-2, 1099, 1098 and K-1 issued with your Social Security number on it arrives at the IRS independently of your return. A computer compares the two.

When they disagree you get a CP2000, a notice of underreported income. It is the most common letter the IRS sends about an individual return, and it is worth understanding three things about it:

• It is a proposal, not a bill. Topic 652 describes it as proposing changes, and it can be disagreed with.
• It is frequently wrong in your favour. A sale reported on a 1099-B shows proceeds, not gain, so the proposed tax often ignores your basis entirely.
• It has a deadline, and missing the deadline converts a proposal into an assessment.

This is the automation that touches ordinary taxpayers, and no language model was involved in building it.

What it means in practice

The practical implication is unglamorous: file a return that agrees with the documents already on file, and keep the records for the parts that are not on any document.

You can check what the IRS has for a given year. Get Transcript gives you the wage and income transcript, which lists every information return filed under your number. Reading that before you file is the single most effective way to avoid a matching notice.

Both sides of the table

Here is the part worth sitting with. The agency selecting returns is deploying machine learning, staffed by data scientists, with access to every information return ever filed about you.

On the other side, increasingly, sits a taxpayer who asked a chatbot.

That is not an argument that AI is bad. The IRS's tooling is a reason to care more about the quality of the return and about who can speak for it. When a notice arrives, the response has to come from someone who can read the file, take a position, and be held to it. Circular 230 restricts that role to attorneys, certified public accountants, enrolled agents and a few other licensed categories, and the restriction is the point: it means someone with standing is answering.

A model can help you understand the notice. It cannot answer it.