Two questions frequently get conflated in the public debate. One is whether artificial intelligence can be useful in the tax field — short answer: yes, in ancillary tasks. The other is whether a conversational model can replace the review, classification and self-assessment process for an individual's Spanish tax return. There the answer is no, and we explain why with six operational and regulatory reasons.
1 · The model has none of your real tax data
A tax return is not a theoretical exercise. It is a calculation grounded in verifiable data: exact withholding amounts per payer (Form 190), acquisition and disposal values of investment funds (Form 296 and broker statements), capital income per institution (Form 193), Social Security contribution bases, cadastral values for imputed real-estate income (Article 85 LIRPF), energy-efficiency certificates if claiming home-improvement deductions. ChatGPT has none of this. It only has what the user types in. If the user types gross where net was required, the model does not detect the inconsistency and returns a plausible but incorrect figure. Responsibility for the wrong data lies with the taxpayer under Article 191 LGT.
2 · The model's knowledge cut-off lags previous campaigns
Large-language models have a known knowledge cut-off date. Tax law is updated each tax year. For the 2025 campaign in Spain, the new entries include: bracket and threshold adjustments to Article 20 LIRPF, regional deductions modified by the late-2025 fiscal-measures laws, new TEAC doctrine, updated DGT criteria in binding rulings, and the new reporting duty for cryptoassets in Form 721. The model does not know what was outside its training. Worse: it does not know that it does not know — it generates confident answers using outdated data.
3 · The downstream cost of error is long-tailed
An incorrect return can move in several directions, all of them costly:
- Overpayment from missed deductions — the taxpayer "donates" tax to the Treasury. Recovery requires rectification of the self-assessment within the four-year limitation period of Article 66 LGT, with the burden of proof resting on the taxpayer.
- Underpayment from omitted income — the AEAT can reassess with delay interest of 4.0625% per year and, where it identifies negligence or fault, apply the sanction regime of Article 191 LGT (between 50% and 150% of the under-declared tax).
- Failure to file information returns — Form 720 (assets abroad) or Form 721 (cryptoassets abroad). The specific sanction regime, even after CJEU judgment C-788/19 of 27 January 2022, remains material.
4 · No professional liability or coverage
A tax adviser registered with REAF carries professional indemnity insurance. If they make an error in the return, the cover responds for the economic harm caused to the client. The AI model has no such cover, nor does the platform offering it. The terms of service expressly relieve the user of any responsibility for decisions taken on the basis of the model's outputs. In practice the taxpayer is left alone facing the AEAT.
5 · Data privacy is not trivial
Pasting full personal and financial data into a conversation with an AI model in order to compute a tax return amounts to a processing of personal data subject to the GDPR and the Spanish Personal Data Protection Act (LOPDGDD). The legal basis for the processing on the model provider's side, retention periods, international transfers outside the EEA and the practical exercise of GDPR Article 15 rights remain open questions for many AI services. The AEAT, in turn, operates under the duty of secrecy of Article 95 LGT, with a known and bounded regime.
6 · What AI can do in the tax workflow
The conclusion is not anti-AI. Generative AI, fitted into the workflow, adds value in three ancillary tasks:
- Concept comprehension — what a regional deduction is, how imputed real-estate income works, what the Article 93 LIRPF regime is.
- Question preparation for the adviser — listing relevant doubts before a meeting and saving operational time.
- Summaries of long regulatory texts — TEAC rulings, DGT consultations, Supreme Court judgments.
Setting those cases aside, what AI does not yet do reliably is take concrete tax decisions on real data, with professional liability and updated knowledge of the current tax year.
The real-cost calculation
The viral thread's argument was savings on adviser fees. The figure was €150 in adviser fees against the model's free use. The calculation, however, must include the expected cost of error. A missed regional deduction — Madrid, Catalonia, Andalusia, Valencia — can cost between €500 and €1,237 of tax. Omission of the Article 7.p exemption for work performed abroad can reach €14,424 on the €60,100 exempt cap at the 24% rate. Omitting a foreign broker prevents loss carry-forward over four tax years. The Article 191 LGT sanction for inaccuracy, where the AEAT applies it, runs between 50% and 150% of the under-declared tax. The inflection point sits well below the cost of professional fees.
Our position
We use AI models internally on bounded tasks: drafting regulatory text, searching TEAC doctrine, first-read analysis of investment statements. We never use them as a substitute for tax-liability calculation or for professional judgement on income classification, residence determination or the application of special regimes. That decision we take in writing, with professional indemnity, with current-year knowledge and with documentary evidence on file. We therefore recommend taxpayers use AI to understand, not to file.