Why General AI Chatbots Can Get Indian Tax Law Wrong

An AI answer can be polished, confident and wrong. In Indian tax practice that matters: a missed amendment, a misread proviso or a citation that does not support the stated proposition can undermine an otherwise careful piece of work.

Imagine an associate preparing a tax note with the help of a chatbot. The draft refers to a statutory section and two court decisions. Before the note reaches the client, a reviewer finds that one reference is outdated and the other cannot be located. The example is hypothetical, but the failure pattern is real: fluent language can conceal weak or unverified research.

General-purpose AI is not useless for tax work. It can help frame questions, organise facts and produce a first draft. The problem starts when a plausible answer is treated as an authoritative statement of current law.

1. Language fluency is not the same as legal research

Large language models generate responses from patterns learned during training and, depending on the product, from information retrieved through search or connected sources. Products differ in their training data, legal databases and live-retrieval features. A model may explain a familiar rule well yet struggle with an exception, a conflicting judgment or a provision that has changed.

Claims about a model’s legal knowledge should therefore be tested against the specific version and configuration in use. A response that sounds like a tax opinion is not evidence that the statute or judgment was actually checked.

2. Legislative change makes old knowledge risky

The Income-tax Act, 2025 came into force on 1 April 2026 and reorganised the structure and numbering of many provisions. A researcher who keeps using a familiar 1961 Act reference without checking the relevant tax year may cite the wrong provision or overlook a transition rule.

For example, section 80C of the Income-tax Act, 1961 broadly corresponds to section 123 of the 2025 Act. The life-insurance exemption formerly in section 10(10D) is now set out in section 11 read with Schedule II, Table item 2. These examples show why mappings must be verified; they do not replace checking the conditions and transitional provisions that apply to a particular case.

The same caution applies to Finance Act amendments, rules, CBDT and CBIC circulars, notifications and judicial interpretations. A section number may be valid yet irrelevant to the tax year, transaction or question being analysed.

3. Access to current sources must be tested

Some AI products search the web or retrieve from a curated legal corpus. Others answer mainly from model knowledge. Even with live search, a system can retrieve an outdated page, miss a corrigendum or summarise a rule without accounting for its effective date.

A useful test is to give the tool a recent, verifiable notification and ask it to identify the issuing authority, date, operative change and effective date. Then open the official source and compare. Ask the vendor how updates are ingested and how corrections are handled; do not rely on a broad claim that the product is “up to date.”

4. Citations can look real without supporting the answer

Checking a citation means more than confirming that a case name or neutral citation exists. The cited judgment must be the right decision, the quoted paragraph must appear in it, and the judgment must actually support the proposition attributed to it. A real case can be misquoted or used for a point it never decided.

The Buckeye Trust matter before the Bengaluru Bench of the Income Tax Appellate Tribunal is a cautionary example. The Tribunal recalled its order of 30 December 2024 on 7 January 2025, citing “certain inadvertent errors”; the recall order itself does not mention AI. The AI concern came from press reports pointing to apparently non-existent authorities in the original order, and from the assessee’s writ petition before the Karnataka High Court, which said the order appeared to have been driven by artificial intelligence. No tribunal finding identifies an AI tool as the source of the errors, so the episode is best read as a warning rather than as proof about any product.

The Supreme Court has since dealt squarely with fabricated AI-generated precedents. In Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (2026 INSC 668, decided 2 July 2026), the Court set aside the NCLT and NCLAT orders in an insolvency matter after they relied on non-existent or misattributed authorities, restored the Section 7 application to the NCLT, and directed the Bar Council of India to constitute a committee on safeguards and disciplinary norms. In Vijay Ghanshyam Gadiya v. Union of India (2026 INSC 947, decided 2 September 2026), the Court set aside a customs penalty of about ₹425 crore and the High Court order upholding it, because the adjudicating officer had relied on case laws that were non-existent, carried fake citations or did not support the propositions attributed to them. The matter went back for fresh adjudication by a different officer.

Both rulings arose in adjudicatory proceedings, not in a CA disciplinary matter, and they should not be read as rulings on a chartered accountant’s professional liability. Their lesson for tax researchers is narrower but important: legal authorities must be independently verified before anyone relies on them.

5. What does a reliable verification workflow look like?

Use AI to speed up research, not to take the researcher out of it. Before a tax note, appeal submission or opinion is finalised:

• Open every cited statute, rule, notification and circular in an official source. Confirm the exact text and effective date.
• Check every case citation in an authoritative reporter or recognised legal database. Read the relevant paragraphs rather than relying on a generated summary.
• Confirm that the decision has not been reversed, stayed, overruled or materially distinguished, and note the level of the court or tribunal.
• Apply the authority to the client’s facts independently. Flag assumptions, missing facts and competing interpretations.
• Record who verified the authorities and who reviewed the conclusion before the document is issued.

6. What the benchmark evidence does—and does not—say

The Taxmann.AI–IIT Kharagpur LE-BTL study reports material differences between models and subject areas. Its published methodology describes more than 100 expert-validated questions spanning direct taxes, GST, FEMA, accounting standards and judicial precedents. The highest direct-tax score reported was 72.84%, and the study noted that models were weaker at applying law and justifying conclusions than at identifying issues and rules.

This is a useful warning about the limits of the models tested, not a permanent accuracy ceiling for every AI product. The models, prompts, questions and scoring approach all affect the results, and model capabilities keep changing.

The practical takeaway

Judge an AI answer by the authority behind it, the quality of its reasoning and its fit with the applicable law and facts—not by its tone. Curated legal sources and pinpoint citations can make checking faster, but they do not remove the need for independent verification and professional judgement.

Related reading

• AI for Indian Tax Research: Common Failure Modes
• How to Verify a Tax Answer Against Primary Sources
• Income-tax Act 2025 Section Mapping: Old vs New Reference