Tax is the hardest legal AI test
Tax law makes legal AI uncomfortable because it rewards determinacy. The answer is not persuasive because it sounds right; it is persuasive because a statute, regulation, ruling or case actually says what the lawyer claims it says. Bloomberg Law commentary this week put the point sharply: tax law still needs humans because AI systems are probabilistic while legal authority is verifiable.
That is not an anti-AI position. It is an anti-delegation position. AI can summarize, draft, compare, flag and accelerate. It cannot discharge the lawyer's duty to verify authority.
The hallucination problem is really an authority problem
Hallucinated citations are the obvious failure mode, but they are not the only one. A model may cite a real case for the wrong proposition, omit a limiting regulation, confuse superseded guidance with current law, or flatten a jurisdictional distinction that matters in tax practice.
The practical risk is seduction. The answer appears complete, professionally phrased and internally coherent. That is precisely why it must be checked. The more polished the output, the easier it is to forget that legal validity comes from sources, not style.
What a verification workflow looks like
A serious tax AI workflow should separate drafting from authority checking. First, use the tool to create an issue map or draft. Second, independently verify every cited authority in primary-law databases. Third, log what was checked, by whom, and what changed after verification.
Firms should not treat this as optional hygiene. In a sanctions environment, the audit trail is part of the defense. It shows the lawyer used AI as assistance, not as an unexamined substitute for professional judgment.
The bottom line
Tax lawyers may become the profession's AI realists. They will use the tools because speed matters. But they will insist the final answer must be traceable to authority. That is the standard the rest of legal AI will eventually adopt.