What the evidence says
Thai sits low on real-content quality measurement — 75.0 on an index where French, Italian and Spanish sit around 80 and above — and the gap between the best engine and a reasonable alternative is 6.8 points, one of the widest measured. English to Thai enters the main international evaluation for the first time in 2026, which tells you how thin the independent evidence base still is.
Why it happens
Thai writing does not put spaces between words — only between sentences. Before an engine can translate anything, it has to guess where the words start and stop, and compound terms, product names and proper nouns are where it guesses wrong. Every error after that inherits the mistake.
On top of that: five tones marked with diacritics that are frequently mishandled; no capitalisation, so there is no signal marking a proper noun; no verb conjugation, no articles, no grammatical tense markers, and no gender — so information that English carries in its grammar has to be reconstructed from context; and a large set of pronouns and particles that encode familiarity, deference and the speaker's gender, chosen by social relationship rather than by grammar.
What it means for your workflow
Thai errors are structural and they are early. A segmentation error looks like a fluent Thai sentence about the wrong thing — which is the failure mode automatic scoring is least able to catch, because the output is well-formed.
What we do
We check segmentation of names, products and technical compounds first, because that is where the cascade starts. We check pronoun and particle choice against the audience. And we treat a passing automatic score in Thai as a starting point, not a result.