Article
BeginnerWhere AI Translation of Member Communications Slips, and When to Call a Translator
An AI tool can turn a renewal notice into fluent Spanish in seconds, and the errors are hardest to see for the person sending it. Research on negation and a translators' association statement show where it slips.
AI Literacy Communications Translation Member Services
A chapter president asks for the renewal email in Spanish by Friday. Nobody on staff reads Spanish, an AI tool produces a fluent version in seconds, and it looks finished. The trouble is that it would look just as finished if it told members the opposite of the policy, and research on one small kind of word shows how easily that can happen.
A small word with a large effect
Researchers at the University of North Texas and George Mason University studied what happens to translation quality when a sentence contains a negation, a “not,” “no” or “never.” Their paper, It’s not a Non-Issue: Negation as a Source of Error in Machine Translation, calls negation “a universal, core property of human language that significantly affects the semantics of an utterance.” Their headline result: “the presence of negation can significantly impact downstream quality, in some cases resulting in quality reductions of more than 60%.”
That figure needs its context, and it is an important one. The study looked at translation systems tested in 2018 and 2019 research competitions, comparing human ratings of sentences with and without negation across 17 language directions. It did not test the AI chat tools staff use today, and we would not claim it tells you how often a current tool drops a “not.” What it does tell you is where to look. Negation is a place where translation systems have been measured going wrong, and badly.
The authors named the ways it goes wrong. They label them “Negation Omission,” “Negation Reversal” and “Incorrect Negation Scope.” In plain terms, the negation can disappear, it can flip so the sentence says the opposite, or it can attach to the wrong part of the sentence. Picture a renewal notice that says members who have not renewed by the deadline cannot vote at the annual meeting. Drop one negation and the translated notice may tell those members they can. This is a teaching example, not a case study, but it is the kind of sentence associations send every renewal season.
Why the person sending cannot see it
If a translation error announced itself, none of this would matter much. It does not. The American Translators Association’s statement on AI, approved by its board in May 2025, puts the problem in one sentence: “One of the greatest dangers of AI-generated translations and interpretations is that they may appear accurate to the general observer, making errors harder to detect for those without linguistic expertise.”
That describes most association staff asked to send a translated notice. The text is fluent. The formatting is right. The member’s name is in the greeting. Nothing on the screen signals that a deadline moved or a “not” vanished. The usual safeguard in a communications office, a second pair of eyes before send, does not help when the second pair of eyes cannot read the language either. A colleague can confirm the email looks professional. Only someone who reads Spanish can confirm it says what the English says.
You might hope an automatic score would catch it. The North Texas and George Mason team checked that as well, and found that “automatic metrics are bad estimators of machine translation quality when negation is present.” In other words, the measures researchers use to grade translations at scale were least reliable exactly where the meaning was most at risk.
The translators’ association adds a second warning. It notes that “AI translation performance varies widely across language pairs, and users may be unaware of these inconsistencies, mistakenly assuming AI provides uniform quality across all languages.” An association that checked a tool’s Spanish once may assume its Vietnamese, Haitian Creole or Korean is just as good. That assumption has no support, and the members least likely to catch an error are the ones who most need the translation.
Where we draw the line between a draft and a send
None of this means AI translation is useless for an association. The translators’ association itself says AI-assisted translation “has proven to be very useful for automating predictable, low-risk content, for drafting preliminary versions for expert revision, or in very specific scenarios.” In the same passage it adds that it “may introduce inaccuracies, inconsistencies, and misinterpretations that require expert human intervention to discern and correct.”
Read together, those two sentences describe a workable split. Low-stakes material, such as a social post announcing a chapter mixer, can start from an AI draft and get a quick read from a fluent colleague. Anything that changes what a member must do, pay or decide is different. We think the following deserve a qualified human translator, with the AI version used only as a starting draft:
- dues, fees and refund terms
- deadlines members must act on, including renewal and exam dates
- eligibility, voting and certification rules
- health, safety and accommodation notices
- anything quoting bylaws or legal language
That list is our recommendation, not a rule any source sets for associations. It lines up with where the research and the translators’ association locate the risk: sentences whose meaning turns on a small word, and readers who cannot check the result.
Tone deserves a sentence too. The translators’ association warns that businesses “may mistakenly believe they’re reaching their audience through AI-translated content while their message is lost or distorted, wasting resources and opportunities.” For an association, a membership campaign that reads stiffly or oddly in translation is a quieter failure than a wrong date, but it is still a failure, and no automated check will flag it.
What to do before the next translated notice
Three moves will cover most desks. First, decide in advance which kinds of communication always go to a qualified translator, using a list like the one above, so the decision is not made under Friday pressure. Second, keep a short glossary of your own terms of art, program names and credential titles, with the approved translation for each, so a tool’s guess never replaces an agreed term. If your certification is called a Certified Member in English, the Spanish version should use the same approved phrase in every notice, not whatever wording the tool picks that day. Members who read both versions notice when a credential seems to change its name between emails. Third, before any AI-translated text goes out, run a quick check that does not require reading the language: confirm the numbers and names survived, and translate the result back to English to see whether every negation is still there.
That last step, back-translation, is our recommended practice, not a professional standard, and it has a blind spot: a tool can quietly repair a mistake on the way back, so a clean back-translation is a screen, not proof. Our same-day tutorial, check an AI translation with a term list and a back-translation, walks through the check step by step, with a lab that runs it on invented practice text.
Sources (2)
- Hossain, Anastasopoulos, Blanco and Palmer (University of North Texas; George Mason University): It's not a Non-Issue: Negation as a Source of Error in Machine Translation (Findings of the Association for Computational Linguistics: EMNLP 2020, pp. 3869 to 3885; peer reviewed)
- American Translators Association: ATA Statement on Artificial Intelligence (approved by ATA's Board of Directors May 3, 2025; posted May 20, 2025)