Article 14(3)(f) of the International Covenant on Civil and Political Rights (ICCPR) guarantees free interpreter assistance to defendants who cannot understand or speak the language used in court. When a state fails to provide it, the obligation remains, but the practical burden falls on the defendant. Having faced criminal charges in Taiwan, a country whose language I was unable to read, I can attest to the harm caused by the failure to fulfill this guarantee. To that end, my argument is simple: machine translation is already the interpreter of record for unrepresented non-citizens, not by any policy decision but by default, in the space the guarantee was supposed to occupy. That makes enforcement of an existing procedural right the most effective AI-policy intervention available to any court system, and one that costs nothing new to enact.

I know this because I did it for four years. Two things belong up front: I was the defendant, and I am not asking anyone to retry the facts.

The Case and the Language Barrier

The courts found that in 2021, during a dispute with my landlord, I posted online documents that contained her personal data, including our lease, and that some of the material stayed online for months. Taiwan enforces its Personal Data Protection Act criminally. I was sentenced to six months in prison; the conviction became final at the Supreme Court on Oct. 16 2024 (Taichung District Court 111 Su 607; Taiwan High Court Taichung Branch 112 Shang Su 3114; Supreme Court 113 Tai Shang 4256). My own account is that the lease was online for about two days, and the witnesses I applied to call on that point were never summoned. I could not read the judgments and no one translated them for me; I learned what they say only in September 2026, which is itself part of the story this piece tells.

Taiwan is not a party to the ICCPR, but in 2009 it gave the Covenant the force of domestic law, and in 2020 it amended Article 99 of its Code of Criminal Procedure so that a defendant who does not understand the language “shall” be assisted by an interpreter, a change the Judicial Yuan announced as giving effect to the Covenant. I cannot read Chinese. Across four years of proceedings, no document was ever translated for me. Some hearings were held with no interpreter at all; the court’s own file records an evening hearing that went ahead after the judge asked me, in Chinese, whether I understood Chinese. At others the person sworn as interpreter conveyed, by my account, a fraction of what was said; the file shows interpreters sworn at two early interrogations. On Dec. 22, 2025, the Supreme Prosecutors Office declined to bring an extraordinary appeal; it did not dispute the absence of interpretation or translation, reasoning instead that legal-aid counsel’s presence at first instance was enough. That counsel did not interpret, and I had no counsel on appeal. On 18 September 2026 the same office declined a second petition. It set the earlier reasoning aside and wrote that whether or not an interpreter was used had “obviously” not affected the findings, because the defendant’s own appeals, filed with neither an interpreter nor counsel, had not argued that it did. The summons for execution of my sentence carries a printed line in Chinese: if you do not read Chinese, bring your own translator. On March 31, 2026, Taiwan’s National Human Rights Commission (NHRC) reported that defendants do not always receive timely, appropriate interpretation or case materials in a language they understand, and that effective mechanisms for checking interpretation accuracy during and after proceedings are lacking. The burden, it found, falls mainly on non-fluent defendants.

An Invisible Reliance on AI

Throughout those four years I read the state’s documents about my own case through consumer AI. Every summons, ruling, and prosecutor’s letter went into a chatbot, and what came back was the only version I had.

In 2025 I ran a letter from a Taiwanese oversight body through two systems. They disagreed, not on nuance but on the central holding. I believed the one that told me what I had spent four years wanting to hear. I published that reading, cited it to journalists, and began building a legal submission on it. In October 2025 the agency confirmed in writing that its letter did not say what I had understood it to say. I withdrew the claim, killed the submission before filing, and rewrote my public materials so the correction leads rather than hides.

Here is the point that matters for policy rather than for me. Nowhere in the court file does it say that the defendant’s understanding of the proceedings came from a language model. The substitution is invisible to the system in which it happens. A court that never appoints an interpreter never learns what took the interpreter’s place.

The Risks of Chatbot Translation

First, fluency is decoupled from accuracy. A wrong translation of a legal document reads exactly as authoritative as a right one. A human interpreter who is guessing usually sounds like it.

Second, there is no uncertainty signal. A court interpreter unsure of a term says so, on the record, and the court can respond. A chatbot does not tell the defendant which sentence it was least sure of. It presents the whole document at a uniform level of confidence.

Third, divergence between systems, which in my case was the only real evidence of unreliability, is visible only if you happen to run two. The error lands on the person least equipped to detect it, which is the structural problem in a sentence.

Enforcing the Right to an Interpreter

Regulating translation models for legal use would be slow and contested. Enforcing Article 14(3)(f), and its domestic equivalents such as Taiwan’s Article 99, requires no new law at all. The untranslated documents raise a separate question, under the right to be told the charge in a language one understands and under fairness generally, rather than under the interpreter guarantee itself. Where a competent interpreter is actually appointed, the chatbot returns to being a study aid rather than the trial itself. That is why I think interpreter provision should be treated as an AI-policy question and not only as a fair-trial one.

Two smaller measures follow. Where a machine translation will support a legal decision, whether by a court, a legal-aid office, or a defendant filing a document, divergence signalling should be required: run two systems and flag disagreement. And courts should be asked a question I have never seen put: what proportion of unrepresented non-citizen defendants are reading their filings through consumer AI? I have been unable to find any published figure for how often interpreters are appointed in Taiwan’s courts; the Judicial Yuan’s statistics on foreign parties record nationality and outcome, not language assistance, and the Judicial Reform Foundation has documented untranslated summonses, indictments, and judgments for years. On Sep. 3, 2026 I filed a freedom-of-information request for the number.

Anticipated Objections

I anticipate three objections to this argument. What follows is how I would respond.

1. The defendant should have brought his own interpreter; the summons said so. A mandatory guarantee that must be claimed in the language the defendant cannot read is not a guarantee. The Human Rights Committee’s General Comment 32 places the right “at all stages of the oral proceedings” and makes it free of charge; it does not condition it on the defendant’s initiative. And the people least able to bring their own interpreter are precisely those the NHRC found bearing the failure.
2. Machine translation is improving; the problem will solve itself. A 2018 pilot study found that a fluent but misleading translation did not significantly reduce users’ trust in the system, while a disfluent translation did. As fluency rises, the last property above gets worse: divergence shrinks toward the cases that matter most. I know of no court that audits the product, and the vendors’ own terms say their output may not be accurate and should not be relied on as a sole source of truth.
3. This is just one man’s grievance. It is one man’s file, which is why it can be checked. The pattern is the NHRC’s finding, not mine.

Three Proposals for Courts

I am not asking anyone to say a court was wrong, and I make no allegation of misconduct against any judge or prosecutor. I am asking for three modest things: treat interpreter provision as AI policy, because enforcing the existing right is the most effective intervention against machine mistranslation in legal proceedings; require divergence signaling where a translation supports a legal decision; and count the defendants who are reading their cases through a chatbot. The interpreter is how a trial reaches the defendant at all. When the interpreter is missing, something fills the gap. Right now it is a product with a strong tendency to sound confident and downplay uncertainties. We should at least know that, and then we should stop letting it happen.

Ross Cline is a Canadian who lived in Taichung, Taiwan, for 15 years and ran an English school there. He was convicted in Taiwanese criminal proceedings in which, he says, no document was translated for him, some hearings were held with no interpreter at all, and at others the sworn interpreter conveyed only a fraction of what was said. He says Taiwan’s Supreme Prosecutors Office did not dispute either point in the aftermath of the case. He now lives in Phnom Penh, Cambodia. He is not a lawyer.

Disclosure: the author is the subject of the proceedings described and thus has a personal interest in how they are discussed. The 2025 correction referred to above, and the underlying documents, are available for review here. A summary and short video can be found here.