Where AI Translation Stops Being Good Enough

Ask an operator in 2026 whether machine translation is good enough and you get an answer about quality. That was the right question five years ago. It is the wrong one now, because the output is fluent almost everywhere, and fluency has stopped being the thing that separates a usable Chinese page from a liability.

The question worth asking is narrower: which parts of what you publish can absorb being slightly wrong, and which cannot.

The short version: sort your Chinese text by what happens if a sentence comes out slightly off. Most of it — descriptions, scenery, background — absorbs that fine, and machine output can go straight out. A smaller set cannot: anything that creates an obligation, your own name, anything a platform will read as an advertising claim, and the conventions around numbers. Those need somebody who reads Chinese to look before publication, and the reason is not that the machine writes badly. It is that when it goes wrong here, the output still reads beautifully.

The failure mode changed, and that made it harder to spot

Bad translation used to announce itself. A stilted sentence, a word in the wrong register, a menu item rendered as something anatomical — the damage was visible to anyone, including people who spoke no Chinese, because the seams showed.

Current output has no seams. When it gets something wrong, it produces a confident, idiomatic Chinese sentence that means something you did not intend. Nobody scanning the page will pause at it. Your Chinese-speaking friend who "gave it a quick look" will not pause at it either, because nothing about the surface asks to be questioned.

That is the whole problem in one line: the cost of checking did not go away, it moved. It used to be proofreading, which anyone could do. Now it is verification, which needs somebody who can read the Chinese and knows what you actually promised.

Sort by consequence, not by length

The useful triage takes about ten minutes and applies to any page you were about to publish.

Most of a tourism page lives in the first two rows. The third row is usually a dozen sentences across an entire site, which is what makes this affordable: you are not buying a translated website, you are buying attention on the sentences that can cost you something.

The categories that stay in the third row

Anything that creates an obligation. Cancellation windows, deposit terms, what a price includes and excludes, weather policy, age and mobility limits. If the Chinese text says something different from the English, the Chinese text is the one your customer read. A model has no way of knowing that "we may cancel in adverse conditions" and "we will cancel in bad weather" are different commitments, because both are reasonable renderings of a vague original. The vagueness is yours, and translation is where it surfaces.

This is also the one place to be clear about what a translated document is not: we draft, your lawyer approves. Nobody should be publishing translated terms on the strength of anybody's translation alone, ours included.

Your own name, and the names of what you sell. Translation engines translate names that should be transliterated, and they do it inconsistently — a tour called "Golden Circle Classic" can come back three different ways across three documents, and now you are three faintly-supported businesses instead of one. The decision about your name belongs to you rather than to a model; naming your brand in Chinese works through how to make it.

Anything a platform will read as an advertising claim. Chinese advertising law restricts superlatives and claims you cannot prove, and translation manufactures both without being asked. "One of the best ways to see the coast" is a mild English hedge; rendered naturally into Chinese it can land as an absolute. Health, recovery and wellness wording carries a stricter regime again. The model is optimising for natural Chinese and has no opinion about what is enforceable.

Numbers and conventions, which are not really translation at all. Dates, times, currency, measurement, floor numbering, what a "double room" contains. The words come across correctly and the convention does not, which produces a page that is linguistically perfect and practically wrong. These are worth a separate read specifically for numerals, ignoring the prose.

Bad news, refusal and apology. Register is the hard part of Chinese for a machine, and a reader receiving bad news is judging one thing only — whether you mean it. We worked through this for review replies in answering a bad review in Chinese and the same applies to any message delivering something the reader did not want.

What it is genuinely good at, and you should use it

The argument above is not an argument against the tools. Three things they now do well enough to rely on:

Reading what arrives. An enquiry, a review, a supplier message, a platform notice. You need to know what it says and roughly how strongly it says it, and machine output gets you there. The care is needed on the way out, not the way in.

First drafts of descriptive material. Itinerary paragraphs, what you see at each stop, background about a region. Factual, low-consequence, and a human rewrite of a decent draft is much faster than writing from nothing.

Working out whether something needs attention. Running fifty Chinese reviews through a translator to find the three worth reading properly is a completely reasonable use of the tool.

Operators who refuse these on principle are paying for a caution that no longer buys them anything.

The part nobody can tell you

Where the boundary sits moves, and it has moved in one direction for several years. Some of what needed a person in 2024 does not now, and some of what still does will stop needing one.

What has not moved, and probably will not: a model cannot know what you meant when your English was ambiguous, cannot know which rules a platform enforces this quarter, and cannot be held to a commitment. Those three are the reason the third row of the table exists, and none of them is a quality problem that a better model solves.

Everything above comes from where these failures actually show up rather than from any benchmark, and it should be held loosely.

Where CN1X fits

We write and publish your Chinese material, and the drafting uses the same tools everybody else has. What you are paying for happens after that: somebody who reads both languages checking the sentences in the third row against what you actually offer, and telling you when your English original is the ambiguous part.

We are not your lawyer — translated terms are a draft for your own legal review — and we do not supply interpreters for the day itself. If you have a Chinese page already live and no way to tell whether it says what you think it says, send us the URL; a read of the obligation-carrying sentences is the fastest way to find out.