Why general AI still answers with last year's tax law
Has a staff member ever pasted a chatbot's tax answer into a client email?
General AI tools are fluent and fast. They are also prone to one specific error that matters a lot in tax: answering with a rule that was correct for an earlier year. The answer sounds confident. Nothing in it is invented. It is simply anchored to the wrong year.
Why general AI gets the year wrong at a glance
| Cause | What happens | What you see |
|---|---|---|
| Training data has a cutoff | Recent changes are missing | An older limit or threshold |
| Older content outnumbers newer content | The model favors the most repeated version | The "famous" number, not the current one |
| The question does not state a year | The model picks one for you | An answer with no year attached |
| Sources are not shown | You cannot check what it relied on | No way to verify quickly |
Example: one question, two correct answers
Example (hypothetical). A deduction limit is $5,000 for returns through one year and rises to $8,000 starting the next. Ask a general AI "what is the limit?" and it may return $5,000, because most of what was written about the rule used that number. That answer was right for years. It is wrong for the return in front of you.
The pattern is the same for thresholds, phase-outs, filing requirements and effective dates. The more a rule changed recently, the more likely a general tool is to lag behind.
How to spot a wrong-year answer
- Look for a year. If the answer does not say which tax year it covers, treat it as unverified.
- Look for a source. No Code section, regulation or IRS link means no way to check.
- Check the source's date. Guidance written before a change can still rank high in search.
- Compare with your firm's last answer. If the number differs, find out why.
Rules for using AI at your firm
- Always put the tax year in the question.
- Never send an AI answer to a client without checking the source.
- Keep a record of what you verified, so the next person does not re-check from scratch.
- Don't paste client details into public AI tools. See IRC §7216 on disclosure of return information.
Frequently asked questions
Is general AI useless for tax work? No. It is useful for drafting and explaining. It is risky as the final word on a number, a threshold or a date.
Does adding "for this tax year" to the prompt fix it? It helps. It does not guarantee the tool knows the current rule, so the source still needs checking.
What is the biggest risk? A confident answer that was right for a past year, reused without anyone noticing.
How Factagora.tax helps
Factagora.tax is AI tax research for any tax year. It answers with the law that applied to the return you specify, not just today's law, and shows the primary authority and effective date behind each answer so your team can check it.
Want to see it on your own client questions? Get early access. Early access is free.