openaccountants
All posts

The IRS Just Told Every Tax Pro: AI Hallucinates, and You're Liable

Michael Cutajar|24 June 2026|4 min read
aiirscircular-230complianceverification

The IRS just released Introductory Guidelines for Responsible AI Use in Federal Tax Practice (IRS Newswire, Issue 2026-19). It's the clearest signal yet that AI is now part of mainstream tax work — and that the profession is on the hook for getting it right.

The takeaway for practitioners is uncomfortable: you must verify everything AI produces, and when you don't, Circular 230 holds you responsible.

AI is already everywhere — the IRS says so

The guidance opens by stating that "virtually all professional tax firms use some form of AI, whether they are aware of it or not" — from research platforms to generative AI that produces original content and can "make discretionary decisions, devoid of any human interaction."

This isn't a warning about a future technology. It's an acknowledgment that the tools are already in the office.

The limits are the liability

Then it names the problem. Generative AI's "fabricated outputs (or, as commonly termed, 'hallucinations'), bias, and lack of transparency … pose serious ethical and legal risks," plus confidentiality exposure when one client's data gets repurposed into another's answer. The instruction is direct: "it is incumbent on any tax professional using GAI to carefully review all documents crafted by the technology."

The consequences are already real

Courts have sanctioned lawyers for AI hallucinations — fake citations in filings — with financial penalties, public censure, mandatory ethics courses, and bar referrals. And it's spreading beyond law. The guidance points to Deloitte Australia, which agreed to partially refund the Australian government after a 230-page report it delivered contained invented quotes attributed to a judge, references to non-existent reports, and books ascribed to the wrong authors — "all produced apparently by GAI."

A fabricated citation is no longer an embarrassment. It's a refund, a sanction, and a disciplinary referral.

Circular 230 makes the duties concrete

The guidance maps AI use straight onto rules that already exist:

  • Due diligence (§10.22) — verify the accuracy of every fact, citation, and calculation AI produces. "Practitioners cannot rely solely on AI."
  • Competence (§10.35) — understand the technology, its limits, and its risks, not just the law.
  • Written advice (§10.37) — base advice on reasonable, verified assumptions; check citations; read the cases. "If the system's logic is opaque, reliance may be unreasonable."
  • Firm procedures (§10.36) — train staff, set data-handling protocols, vet third-party AI tools, and document all of it.
  • Confidentiality (§§7216, 6713) — never upload taxpayer information to "unsecured or public systems."
  • Fees (§10.27) — pass the efficiencies on; don't bill for time you didn't spend.

The gap nobody's talking about

Here's the structural problem. The IRS tells you to verify everything AI produces — but a general-purpose AI produces tax answers from whatever was in its training data: forum posts, marketing pages, last year's blog articles, the occasional government PDF. There's no authoritative, cited source behind the answer to check it against. You're told to verify, with nothing verified to verify against.

That gap is exactly what OpenAccountants was built to close.

What we're doing about it

OpenAccountants gives AI accountant-verified, source-cited tax rules to read instead of guessing — and it lines up almost one-to-one with the duties above:

  • Cited to statute — every rate, threshold, and treatment links to the section, regulation, or revenue procedure it comes from, so the citations in your AI's output are real and checkable (§10.22, §10.37).
  • Verified by a named, licensed accountant — the rules carry a credentialed professional's sign-off. That's the accountability Circular 230 is built around.
  • Open and inspectable — a third-party tool you can actually vet and document (§10.36).
  • Data-minimal by design — the rules run inside the AI you already control; we don't ingest your clients' returns, which keeps the confidentiality exposure where it belongs (§§7216, 6713).
  • Built for review, not blind reliance — outputs are draft working papers with the assumptions and citations laid out for you to check and sign.

We're explicit about the line. OpenAccountants gives the AI the right things to read and a professional's name behind them. The practitioner still reviews, verifies, and signs. That isn't a workaround of the IRS guidance — it is the IRS guidance.

A standard, not a threat

The guidance isn't anti-AI. It says generative AI "holds promise to improve efficiency and professional services in tax practice" — but that "final decisions must always rest with qualified professionals." That's the whole game: AI for speed, a licensed human for trust, and verified rules in between.

The profession that builds that middle layer — verified rules, named accountants, human sign-off — doesn't just comply with this guidance. It sets the standard everyone else's AI gets measured against.

That's what we're building, one country at a time. If you're a licensed accountant, verify your country's rules. If you use AI for tax, connect it to rules a professional stands behind.

This article is general commentary, not tax or legal advice. Confirm your own obligations under Circular 230 and applicable state law.