Machine Access to Primary Tax Law

SummaryThe finding in one paragraph

An AI can reach an official tax source in 79 of 133 jurisdictions. In eight it cannot reach one at all: Israel, Moldova, Cyprus, Mexico, Fiji, Samoa, Maldives and Chad. This does not make those countries' tax figures unknowable, and that is the more troubling result. The figures still circulate freely through commentary, advisory firms and payroll vendors. What is lost is not the number but the ability to check it, so an AI answers confidently from an intermediary while believing it consulted the government.

Section 1Why this was measured

Every argument for AI performing tax research rests on a single instruction: confirm this figure against the official source. That instruction assumes the source will answer. Where it does not, a language model does not stop and report failure. It answers from memory instead, and the result is indistinguishable from a verified answer.

The assumption had not been tested. This report tests it.

Section 2The eight closed jurisdictions

In each of these, every official source on record refuses or fails to answer an automated request. This is a property of those countries' infrastructure, not of any particular software.

Two major jurisdictions refuse at one door but not another and are therefore not on this list. France blocks at Legifrance while impots.gouv.fr answers. The United States blocks at congress.gov while ten other official sources answer, including the IRS, GovInfo and Cornell. Listing either as closed would have been the more striking claim and the wrong one.

Section 3The archive is not a substitute

Malta's blocked pages proved readable through the Internet Archive, which suggested a general workaround existed. It does not.

Coverage is page by page and depends on whether the crawler happened to get through on the day it visited. The same Malta host whose home page archived as a refusal has a social security page that archived intact, with the contribution table readable.

Where the archive does work, age erodes it. Tax rates are re-set annually, so a snapshot from 190 days ago (Samoa) or 108 days ago (Maldives) may straddle a change and be confidently wrong. Of the eight closed jurisdictions, one has an archived copy both substantial and recent enough to rely on.

Section 3bA blocked authority launders its own figures

A refusal is selective. Cloudflare's default configuration admits verified search-engine crawlers while turning away everything else, so the pages of a blocked authority are typically indexed in full even though no ordinary client can open them. Malta's contribution-rate pages are present in search results, correctly titled, while returning 403 to three independent fetchers tested for this report.

The consequence is not ignorance. It is unattributable knowledge. A query for Malta's 2026 contribution rates returns them readily — 10% from employee and employer, €6.62 weekly up to a €229.44 wage, a fixed €55.93 above €29,084, a maximum of €83.89 — sourced from commentary sites, advisory firms, payroll providers and a salary calculator. The figures are almost certainly right. Not one of them comes from the Maltese government.

This is measurable in our own corpus, where 65% of citations were found to be non-primary. Some of that is carelessness. In the eight closed jurisdictions it is structural: there is no primary citation available to make.

Section 4Who is holding the door shut

Twenty-eight sources actively refuse automated access. Cloudflare enforces eighteen of them. A single infrastructure provider therefore stands between AI systems and a material share of the world's published tax law, in all likelihood as an incidental consequence of default configuration rather than deliberate policy.

Nine of the refusing hosts are governments in their own right, including France's Legifrance, the United States Congress, Indonesia's regulations portal, Cayman Islands legislation, and the revenue authorities of Malta, Maldives, Samoa, Chad and Barbados.

Section 5What our own records got wrong

The probe was pointed at our own registry as well. It disagreed with reality 28 times in 356.

• Seventeen sources believed reachable now refuse. The world changed and nothing reported it.
• Eleven recorded as blocked in fact answer. This was a live defect: a source marked blocked with no recorded alternative is discarded when a Guide is served, so those jurisdictions were supplied with no official source at all while a working one sat unused.

Both have been corrected against measurement. The correction moves a flag only on conclusive evidence: a timeout is never recorded as a block being lifted, or the next automated visitor walks into a closed door with no warning.

Section 6Method, and two errors caught before publication

Sources were requested with a standard browser identity, following redirects, on a 25-second timeout. robots.txt was read separately from the page itself, so a request to stay away is never mistaken for an enforced refusal. Blocks are attributed to a provider only where response headers or body identify one.

A success code is not a source

Eleven sources return success and no content. Argentina's revenue service replies with 274 bytes; Chile's with 166. Any checker trusting the status code alone would record these as verified. They are recorded separately as insubstantial.

A government cannot be identified by its web address

An initial pass classified authorities by domain pattern and reported that the United Kingdom, Ireland, the Netherlands, Sweden, Denmark and Norway held no government source, because their revenue authorities do not use .gov addresses. This produced a confident and incorrect headline. Every figure in this report uses curated classification instead.

Readable sources responded in 558 milliseconds at the median and none exceeded ten seconds, so where access is granted, latency is never the constraint. The dataset and the probe are committed to the repository and the measurement can be reproduced in full.

Section 7What this means

For 79 jurisdictions, the instruction to verify against a primary source is achievable, and an AI equipped with a sound method can genuinely perform tax research. That covers the majority of the world and the large majority of economic activity.

For eight, an unregistered client is refused. The limit is not the model's reasoning, its context window or its instructions, and it is not solved by a better prompt. It is a server declining to answer this particular caller. Two remedies exist and they are of very different quality: register as a verified crawler and be let in under a declared identity, or accept the figures second-hand from whoever republished them.

Section 8Recommendations