AI Translation Hallucinations: How to Keep Translations Faithful with SelectTranslate

An AI translation can be fluent, grammatically correct, and still say something the original text never said.

A software announcement mentions a feature that may become available next month. The translation makes it sound guaranteed. A technical document recommends checking a setting, but the translated version turns that recommendation into a requirement.

These differences can be easy to overlook. Nothing in the translated sentence necessarily sounds wrong—until you compare it with the source.

When an AI translation introduces information that isn’t supported by the original, it’s often described as a translation hallucination. Other changes, such as turning may into will, may be better classified as meaning or modality errors rather than hallucinations in the strict sense. Both can mislead readers.

For people using AI to read technical documentation, product announcements, and other foreign-language webpages, the goal isn’t simply to get smoother translations. It’s to preserve what the author actually said.

SelectTranslate lets users configure translation models and, where supported, customize their prompts. With more specific instructions, you can ask an AI model to prioritize faithfulness over unnecessary rewriting.

Here’s how to approach those settings without disrupting the translation process.

1. What Does an AI Translation Hallucination Look Like?

Consider this sentence from SelectTranslate’s official model configuration documentation:

The extension includes built-in, optimized general translation prompts that cover most common translation scenarios.

The important qualification is most common translation scenarios.

The sentence does not say that the prompts work in every situation. Nor does it guarantee perfect translations across every language or subject.

A translation that changes most common to all would misrepresent the original scope. A translation that adds “no customization is ever needed” would go further, introducing a claim that the source never makes.

These are examples of possible translation errors, not actual outputs recorded from a particular model.

The distinction matters beyond product documentation.

Source statementMeaning that must be preserved
“Available in selected regions”Availability is geographically limited.
“Expected to launch next month”The launch is planned, not completed.
“May improve performance”An improvement is possible, not guaranteed.
“We recommend reviewing the settings”The action is recommended, not mandatory.

AI models can generate natural-sounding translations while losing some of these distinctions.

Research has identified hallucinations as a genuine challenge in LLM-based machine translation. In their 2025 NAACL paper, Mitigating Hallucinated Translations in Large Language Models with Hallucination-focused Preference Optimization, Zilu Tang and colleagues investigated ways to reduce unsupported translation output through model training.

That research addresses changes to the model itself. It doesn’t establish how much a user-written prompt can improve translation quality.

Still, custom prompts offer a practical way to make your expectations clearer: translate the source, preserve its qualifications, and avoid adding information.

2. Customize the System Prompt in SelectTranslate

A language model can perform many tasks besides translation, including summarizing, explaining, and rewriting. A translation prompt tells it which behavior you want.

In SelectTranslate, supported LLM configurations expose two types of instructions:

System Prompt defines the model’s overall role and translation behavior.

User Prompt specifies the translation request, including the input and the format of the response.

If your translations contain unnecessary explanations or changes in meaning, the System Prompt is the best place to start.

Find the translation model you’re using

Open SelectTranslate’s settings and go to the translation model section. Select your current LLM and open its configuration options.

If that configuration supports prompt customization, you’ll find a System Prompt field.

The SelectTranslate model configuration documentation explains the available settings, including custom prompts and performance parameters.

You don’t need to change every option. Start with the prompt that controls the model’s general behavior.

Make the instructions specific

For example, the GPT-5 mini configuration has been shown with this default System Prompt:

You are a professional, authentic translation engine.

It establishes the model’s role, but it doesn’t explicitly address the types of changes you may want to avoid.

For translation tasks where preserving the source matters, try the following:

You are a professional translation engine.

Translate the source text faithfully into the requested target language.

Preserve the original meaning, tone, uncertainty, numbers, and conditions.

Do not invent facts, add explanations, strengthen claims, omit information, or rewrite the author's intent.

Use natural and readable language without changing the source meaning.

This prompt focuses on four things that matter when reading translated content.

Preserve uncertainty. Words such as may, might, and could should retain their intended meaning rather than becoming promises.

Keep conditions intact. Requirements, exceptions, version numbers, dates, and availability restrictions should not disappear.

Avoid unsupported additions. The model should not introduce explanations, conclusions, or facts that the source does not contain.

Allow natural wording. A faithful translation can still be idiomatic and easy to read. It doesn’t have to copy the original sentence structure word for word.

Save the configuration, then retranslate a passage where you previously noticed a problem.

For the first comparison, keep the model and other settings unchanged. That way, you’re testing the revised instructions rather than several variables at once.

3. Do You Need to Change the User Prompt or Temperature?

Usually, not right away.

The System Prompt is a reasonable starting point for controlling translation behavior. The other options have more specific purposes.

Keep the User Prompt’s required structure

The User Prompt may contain information that SelectTranslate needs to process webpage translations correctly.

For example, a template used in an LLM configuration can look like this:

Translate the text in the following JSON array
into idiomatic and fluent {{target_lang}}.

Return a JSON array with the same structure and order
as the input, without any additional explanation.

Input: {{original_text_json}}
Output:

The placeholders are important:

  • {{target_lang}} identifies the target language.
  • {{original_text_json}} supplies the text to be translated.

The JSON formatting instruction also matters. It tells the model to return the translated text in a structure that the extension can process.

Replacing the entire template with a generic instruction could remove those requirements, potentially causing missing translations or incorrectly ordered content.

If your goal is simply to stop the model from adding information, leave the User Prompt as it is initially.

If you later decide to refine it, preserve the placeholders and output structure. You can add a brief instruction such as:

Preserve the source meaning and all qualifications. Do not add, invent, or omit information.

Then check that a webpage with multiple paragraphs still translates correctly.

Temperature doesn’t automatically improve accuracy

Temperature affects how varied a model’s responses can be.

Where the parameter is supported, lower values generally produce less variation. Higher values may allow more diverse wording.

But less variation is not the same as greater accuracy.

A model can consistently produce the same incorrect interpretation. Lowering Temperature won’t necessarily fix that.

Support for the parameter also depends on the model and API. If your current configuration offers it, begin with the default value and change it only when you have a reason to test output consistency.

Text length matters when context is missing

Other settings, such as maximum text length and maximum paragraphs per request, affect how translation requests are divided.

These can become relevant when translating longer articles where one paragraph depends on information introduced earlier.

For example, if a product is named in one paragraph and referred to as it in the next, insufficient context may make that reference harder to interpret.

Adjusting how much text is processed together may help in some situations, but increasing request size is not a general solution to hallucinations.

The simplest rule is to change the setting that matches the problem you’re seeing, rather than adjusting everything in search of a perfect configuration.

4. Test Whether Your Translation Actually Improved

A carefully written prompt doesn’t guarantee a better result.

The useful test is whether it corrects the problem you originally noticed.

Start with a real passage from a webpage you’re reading. Product announcements, technical instructions, and policy updates are useful because their meaning often depends on exact conditions.

For example, use a passage that contains a qualification such as:

The feature is currently available only on desktop devices.

The word only is essential. A translation that omits it could lead readers to believe the feature is available on mobile devices too.

To evaluate your prompt change:

  1. Save the original text and its current translation.
  2. Modify only the System Prompt.
  3. Translate the same passage using the same model and target language.
  4. Compare both translations with the source.

Pay attention to the meaning, not just how polished the sentence sounds.

What to checkWhat you’re looking for
Added informationClaims or explanations absent from the source
QualificationsWords such as only, unless, may, and except
Factual detailsCorrect numbers, dates, names, and version requirements
ReadabilityNatural language that doesn’t change the author’s intent

SelectTranslate’s bilingual webpage translation mode can make this easier by allowing you to read the original and translated text together.

If the revised prompt fixes the issue, try it on several other passages before treating it as your preferred configuration.

If the error persists, the source may be ambiguous, the model may lack sufficient context, or the model itself may struggle with that language pair.

In those cases, testing another model may be more useful than adding increasingly complicated prompt instructions.

For users who specifically want to connect their own model provider through an API key, SelectTranslate also provides a separate custom API configuration walkthrough. That setup is distinct from the general task of refining translation prompts.

Final Thoughts

AI translation errors aren’t always obvious. Sometimes a single omitted condition or unsupported detail is enough to change the meaning of an otherwise fluent sentence.

Custom prompts can help you communicate what matters most: preserving the source rather than expanding, explaining, or rewriting it.

With SelectTranslate, start by reviewing the System Prompt for your chosen LLM. Keep the User Prompt’s required formatting intact, and adjust other model parameters only when a specific issue calls for it.

Then compare the result against the original text.

A better translation isn’t necessarily the one that sounds more polished. It’s the one that helps you understand what the author actually meant.

Frequently Asked Questions

Can custom prompts eliminate AI translation hallucinations?

No. Custom prompts can instruct a model to preserve the source and avoid unsupported additions, but they cannot guarantee error-free output. The results depend on the model, the source text, and the available context.

What’s the difference between a System Prompt and a User Prompt?

The System Prompt defines the model’s general behavior. The User Prompt supplies the specific task and may include variables and formatting requirements used to process the translation.

Should I lower Temperature for more accurate translations?

Not necessarily. Lower Temperature can reduce variation where supported, but it doesn’t guarantee better understanding or more faithful translation. Start with the model’s default setting.