AI translation does not convert words from one language to another. It predicts what the target language text should look like given the source. When that prediction is wrong, the result is a fluent, grammatically correct translation of something the source document never said. The user who cannot read the source language has no way to catch it.
Addition hallucination. The model adds information that is not in the source text. A simple sentence gains qualifiers, context, or details that the original does not contain. The translation is longer than the source and the extra content is fabricated.
Omission hallucination. The model drops information from the source. A sentence with three clauses becomes two. A critical qualifier disappears. The translation reads well but is incomplete - and incompleteness in legal, medical, or contractual translation is dangerous.
Substitution hallucination. The model replaces a concept in the source with a different concept in the translation. A specific medication name becomes a different medication. A legal term is translated as a related but legally distinct term. The substitution is plausible enough that a non-expert would not notice.
Detached hallucination. The most severe form. The model generates a translation that is entirely unrelated to the source. This happens most often with low-resource languages, very short inputs, or inputs that confuse the model's language detection. The output is fluent text in the target language that has no connection to the source.
The whole point of translation is that the user cannot read the source. That is exactly what makes translation hallucination undetectable to the person who needs the translation.
The user cannot verify. Translation is the purest delegation task. The user asks for translation because they cannot read the source language. They have zero ability to evaluate accuracy. They can only evaluate fluency - and hallucinated translations are fluent by design.
Fluency masks errors. Unlike older machine translation that produced obviously broken grammar, modern AI translation is fluent. A hallucinated translation reads as naturally as a correct one. There is no signal in the output quality that indicates something is wrong. The same calibration problem - no distinction between correct and fabricated output.
Low-resource languages suffer most. Models trained on less data for a language pair hallucinate more. Languages with limited parallel training data - many African, Southeast Asian, and indigenous languages - see significantly higher hallucination rates. The people with fewest alternative translation options get the worst AI translation quality.
Context window limitations. Translation of long documents suffers from the same attention degradation as summarization. The model loses track of context, terminology choices, and specific details over long passages. Consistency degrades. Names are translated differently in different paragraphs. Technical terms drift.
Legal translation. Contracts, court documents, immigration papers. A hallucinated clause in a translated contract is a fabricated legal obligation. A mistranslated immigration document can result in a denied application based on information the applicant never provided.
Medical translation. Patient records, medication instructions, informed consent documents. A substitution hallucination that changes one medication name to another in translated instructions is a direct patient safety risk.
Diplomatic and intelligence translation. A fabricated nuance in a translated diplomatic communication changes the meaning of a negotiating position. An addition hallucination that inserts a qualifier the source did not contain changes the interpretation of intelligence.
The fix follows the same principle: verify output against the source. For translation, this means automated back-translation checks, semantic similarity scoring between source and target, and flagging of additions, omissions, and substitutions before the translation reaches the user.
Grounding in translation means the source text is the reality. The translation must be verified against it. Check applies this principle - validate AI output against the source of truth before delivery. The language pair changes. The verification principle does not.
120 verifications a day free. No card, no signup.