Primary Source vs Secondary Source AI Tax Answers and Why It Matters for Practitioners

Citation-backed AI tax research for federal and state questions requires more than an answer with footnotes. Marble provides answers grounded in primary authority, while Thomson Reuters’ CoCounsel Tax offers a Citation Viewer for inspecting cited documents. Whether AI answers draw on primary-source or secondary-source material affects whether practitioners can defend their work, but they still need to read the controlling text whatever tool or citations they use.

Why the source of a tax answer has always determined whether a practitioner can rely on it

An IRC section and a practice guide serve different functions in tax research. The section supplies legal authority, while the guide helps the practitioner understand and navigate it. Treasury Regulations, IRS rulings, and binding case law also have a place in the authority hierarchy, separate from the commentary that explains it. A useful explanation does not acquire the legal weight of the provision it describes.

That distinction matters when research becomes advice or a return position. The position must be traceable to primary authority, and secondary commentary alone does not complete the underlying research. Practitioners have always had to move from explanation to authority before standing behind a conclusion. AI makes that duty more urgent because its answers can hide whether they draw on authoritative text, commentary, or neither.

How general-purpose AI generates a tax answer and why that process is structurally different from retrieval

The change begins with how a language model produces an answer. General-purpose large language models generate tokens through probabilistic text prediction, rather than retrieving and verifying each statement against authoritative text. A Treasury Regulation and a practitioner blog post can both enter that process as training data without being ranked by legal weight. The architecture predicts language independently of the place a document occupies in the tax authority hierarchy.

A model may also skip real-time retrieval when it treats its training data as sufficient. The resulting answer can look just like one supported by current research, leaving the practitioner unable to distinguish the two from fluency alone. Asking the model to rate its confidence does not resolve that uncertainty. Its rating comes from the same predictive process that produced the answer, creating a “confidence mirage”.

The resulting “hierarchy error” is more fundamental than reliance on a weak source. AI-generated output is not itself a source in the professional sense, even when its explanation is correct. The contrast is clearest in tools such as Marble, which surface primary authority for practitioner review rather than leaving retrieval and synthesis indistinguishable.

Why hallucination in tax is more dangerous than hallucination in other domains

In tax, an error that sounds right can become a return position, an IRS representation, or advice to a client before anyone spots what is wrong. Fixing it may then bring a risk of penalties as well as a need to edit the answer. Under tax’s self-assessment system, a conclusion must hold up to later checks, so practitioners need authority they can trace it to. Fluency offers no protection at that stage.

Zihao Li, Weiwei Yi, and another researcher describe hallucinated content that remains persuasive and contextually appropriate while conveying distortions that escape conventional accuracy checks. The tax-specific version includes “citation escape,” in which an AI cites a real IRC section or ruling but draws a conclusion the text does not support. The citation passes an existence check while the reasoning fails. That makes a genuine section number potentially more misleading than a plainly fictitious one.

Dollar thresholds, phase-out ranges, effective dates, and provisions that have recently changed pose a particular risk because training data can fall behind changes in the law. A prior-year figure may appear without any signal that it is stale. The usual ways of spotting errors can fall short when an answer is well written and sounds authoritative. Verification must therefore test both the currency of the authority and the proposition attributed to it.

What the courts and disciplinary bodies have established when practitioners relied on unverified AI output

Clinco (2026) involved briefs citing nonexistent cases that appeared to Judge Holmes to have been hallucinated by a large language model, with counsel failing to correct the record after the IRS identified errors and repeating “Cacchillo” in a later table of authorities. The Tax Court stopped short of formal sanctions and acknowledged uncertainty about whether generative AI had actually been used, but issued a severe warning to the tax bar. The court described the briefing as “a bouillabaisse of case names, reporter citations, and legal propositions” suggesting “something cooked up by AI”.

Thomas v. Commissioner, Oct. 23, 2024, had previously flagged apparent AI-hallucinated citations in a Tax Court order. Whiting v. City of Athens in the Sixth Circuit involved fake citations suspected, but not confirmed by the court, to be AI-generated. Counsel in Whiting had to pay a large penalty. These outcomes differ, but responsibility remained with the professionals who submitted the work. An uncertain account of how the citations arose did not make the citations acceptable.

What OPR Alert 2026-19 and Circular 230 now require practitioners to understand about AI source quality

OPR Alert 2026-19, issued June 24, 2026, clarified that existing Circular 230 obligations apply to AI-assisted work. It gives guidance on due diligence under Section 10.22, competence under Section 10.35, fees under Section 10.27, and firm procedures under Section 10.36. The governing principle is that practitioners retain full responsibility for work produced with AI assistance. Delegating the production of text does not delegate the professional obligation to support it.

OPR Alert 2026-19 interprets Section 10.35 to include technological literacy within competence. Practitioners must understand how AI generates content, recognize its limitations, and identify where errors, bias, or unreliable outputs may arise. Source quality becomes a compliance issue because those limitations include the failure to distinguish legal authority from other material. To judge whether an answer can be trusted, a practitioner needs to know whether the system retrieved authority or just generated an explanation.

IRC Section 6694 imposes preparer penalties for understated tax liability due to unreasonable positions. A wrong position backed by a made-up or unsupported AI citation can trigger a penalty on the same grounds as any other unreasonable position. Circular 230 Section 10.51(a)(13) also covers false opinions given knowingly, recklessly, or through gross incompetence. These provisions attach consequences to the practitioner’s work, rather than creating an exception for machine-generated material.

OPR Alert 2026-19 also flags a potential unconscionable-fee violation under Section 10.27(a) when AI time savings are not reflected in fees and sets firm-level requirements for staff training, secure data handling, accuracy monitoring, and vetting third-party AI tools. Documentation must demonstrate adherence to those procedures.

How retrieval-augmented generation changes the risk profile when it is properly implemented against a primary-source corpus

Retrieval-augmented generation, or RAG, changes how an answer is made by feeding retrieved documents into the process. In a properly implemented primary-source system, the tool searches a licensed database, retrieves the controlling document, and constructs a response grounded in that text. Citations then connect the response to an inspectable source. This architecture gives the practitioner a basis for review that prediction over mixed training data alone does not provide.

However, retrieval adds context. An outdated or irrelevant document can still support a confident but incorrect response. When retrieved text does not explicitly answer the question, the model fills the gaps by blending it with prior training. The finished answer can sound fully grounded even though retrieval supplied only part of it. RAG lowers the risk of hallucination, but practitioners still need to check whether the evidence supports the conclusion.

Document chunking creates another failure mode. A system may retrieve a paragraph with the general rule but leave a jurisdictional exception or effective-date qualifier in the next chunk, outside the context window. The retrieved passage is genuine, but incomplete for the question being answered. Corpus quality and document handling therefore matter as much as the presence of a RAG layer.

The practical evaluation must go beyond whether a vendor uses RAG. Practitioners need to establish what the corpus contains, how primary and secondary material are labeled, and whether citations open to the actual controlling document. A primary-source library and an unlabeled mixed web corpus do not become equivalent merely because both feed a retrieval system. The architecture is useful only to the extent that it preserves the distinctions required for legal analysis.

Which AI tools give cited sources, and what distinguishes purpose-built tax platforms

Thomson Reuters Checkpoint, Wolters Kluwer CCH AnswerConnect, and Bloomberg Tax provide curated research libraries with editorial depth. Their traditional retrieval tools bring back source documents for practitioners to read and judge, and the added AI layers use those same checked collections. That separates access to research material from the additional question of how faithfully a generated answer uses it.

Thomson Reuters’ CoCounsel Tax makes source inspection explicit through its Citation Viewer. A practitioner can click a citation to review the source, retrieve metadata, search within the document, download it, or print it. These features let practitioners use an authoritative citation as evidence as well as see it in a footnote. The citation trail supports the review process without deciding the legal conclusion for the practitioner.

Blue J illustrates a different nuance through its blended corpus. A purpose-built tool can draw on primary authority and secondary material without making the two interchangeable. The practitioner still needs to identify which kind of source supports a particular proposition.

Marble combines AI-powered tax research with client intake, document analysis, and return preparation. Its research addresses complex federal and state questions with answers grounded in primary authority and cited claims, including retrieval from the IRC, Treasury Regulations, and binding case law. Its broader workflow premise is to automate routine backend work so practitioners can concentrate judgment on advisory work. That makes it an option for integrating research with the engagement, rather than a basis for claiming superior research performance.

General-purpose assistants applied to tax questions offer a different reliability profile because they can point to secondary commentary, reproduce prior-year thresholds, or omit jurisdictional exceptions without signaling uncertainty. Vendors must clarify whether the AI cites only documents in its own database and whether each citation opens to the controlling text. A vague answer should be treated as a negative answer. For federal and state research alike, the useful comparison is inspectable authority and coverage.

What a defensible AI-assisted tax workflow looks like under the current regulatory standard

OPR Alert 2026-19 leaves AI use open to practitioners but requires them to build responsible use into their work. The workflow begins at tool selection with source evaluation and a Section 7216 consent analysis before client data is shared. Firm leaders must also establish written policies for training, tool vetting, secure data handling, and accuracy monitoring. These controls belong before an engagement is underway.

Research should then use narrow, single-jurisdiction questions in place of a request for a complete multijurisdictional answer. Longer output creates more room for unsupported gap-filling. Each generated citation should be treated as a lead, with the practitioner opening the section, ruling, or case and testing the conclusion against its text. Particular attention belongs to thresholds, phase-outs, amended provisions, effective dates, and state or multistate rules.

The verification step also needs a record. A note-to-file should identify the tool, its output, the citations reviewed, the sources used for that review, and the judgment the practitioner applied. That record supplies evidence of due diligence. Where AI materially reduces research and drafting time, billing should reflect that efficiency under the alert’s guidance on Section 10.27(a).

The final standard applies regardless of which tool the firm selects. A citation is useful only when the cited authority supports the position, and a generated answer becomes professionally usable only through that assessment. Understanding where AI output sits outside that hierarchy, and knowing how to reconnect it to authority, is now part of practitioner competence.