Troubleshooting AI Translation Quality Issues

Overview

AI translation quality issues typically fall into four main categories: register and tone inconsistency, literal or context-free translations, hallucinations, and prompt optimization failures. Each has a distinct cause and resolution path. This article helps you diagnose and resolve those problems — it does not cover how to run AI translation or how to write custom prompts.

When to use it

Use this troubleshooting guide when:

• AI translations are using the wrong register (e.g., formal vs. informal in German, Spanish, French)

• Translations are too literal and miss contextual meaning

• AI is adding content not present in the source (hallucinations)

• Custom prompts are not producing the expected output

• Quality was good previously but has degraded after a configuration change

• A specific language pair is consistently underperforming

Requirements and limitations

• Engine compatibility: Register and tone instructions via custom prompts are only supported by LLM-based engines (GPT-4o, Claude, Gemini) — not traditional MT engines

• Prompt scope: Custom prompts cannot override system-level translation principles

• TQS scope: Translation Quality Score currently only assesses AI translation quality; TM matches and human-reviewed translations receive maximum scores automatically

How it works

Using TQS as a diagnostic starting point

Before troubleshooting manually, use Translation Quality Score (TQS) to identify which segments are scoring low. TQS provides a score from 0 to 100 for each AI-translated segment. Low-scoring segments are your starting point. They pinpoint where the AI is struggling without requiring a full manual review.

To view TQS scores:

1. Open your project in Smartcat

2. Look for the TQS widget on the Project Overview page

3. Segments below your threshold are flagged as "needs review"

4. Focus your troubleshooting on these flagged segments first

For more information, see Understanding Translation Quality Score.

How Smartcat processes AI translations

Understanding the pipeline helps you identify where quality issues originate:

1. Document Segmentation: Files are broken into segments (usually one sentence each)

2. Translation Memory Lookup: Each segment is checked against your TM for exact matches

3. AI Translation: Segments without TM matches are translated by the selected AI engine

4. QA Checks: Automated checks flag errors in spelling, punctuation, terminology, and formatting

5. Glossary-Term Fix: For non-LLM engines, Smartcat triggers corrections when glossary terms aren't used correctly

6. Fallback Translation: If a segment still has critical errors, Smartcat automatically reruns translation using Google NMT as a backup. This is expected behavior — a safety net to ensure deliverable quality — not a failure.

For example, if TM matches are poor quality, they'll be applied before AI translation even runs. For more details, see AI Translation Overview.

Issue 1 — Register and tone inconsistency

Symptoms: German translations inconsistently use "du" (informal) and "Sie" (formal). Spanish translations mix "tú" and "usted." French translations alternate between formal and informal register.

Cause: No register instruction in the custom prompt, or the AI model is defaulting to informal register for certain language pairs. This is one of the most common issues reported by enterprise clients, particularly for German translations.

Resolution:

1. Add an explicit register instruction to your custom prompt:

2. For formal German: "Always use formal register (Sie) when translating to German."

3. For informal German: "Always use informal register (du) when translating to German."

4. For Spanish: "Use formal register (usted) for all Spanish translations."

5. Verify your glossary does not contain informal term variants that conflict with your register preference

6. If using a non-LLM engine, switch to an LLM-based engine (GPT-4o, Claude, Gemini) — these respond more reliably to register instructions

7. Test your prompt across multiple segments before running full translation

For detailed guidance on writing custom prompts, see Customizing AI Translation with Custom Prompts.

Issue 2 — Literal translations lacking context

Symptoms: Translations are word-for-word and miss the intended meaning. Marketing copy sounds flat or awkward. Industry-specific terms are translated incorrectly.

Cause: The AI model has no context about the content type, industry, or intent. Without context, AI defaults to word-for-word translation.

Resolution:

1. Add a content description to your custom prompt:

2. For marketing: "This is marketing copy for a sportswear brand. Translate with a dynamic, energetic tone rather than literally."

3. For technical: "This is software documentation. Maintain technical accuracy and use standard industry terminology."

4. For e-commerce: "This is a product description for an e-commerce catalog. Focus on benefits and readability."

5. Check whether your translation memory contains literal translations that the AI is matching against — poor TM quality feeds poor output

6. Use TQS to identify which segments are scoring low and review those specifically

7. Ensure your glossary contains industry-specific terms with correct translations

For more information on TQS, see Understanding Translation Quality Score.

Issue 3 — Hallucinations (AI adding content not in the source)

Symptoms: Translated segments contain information, explanations, or phrases that do not appear in the source text. AI "fills in" ambiguous content with plausible but fabricated details.

Cause: LLM-based engines can generate plausible but fabricated content, especially for short or ambiguous segments. This risk is higher for very short segments, ambiguous content, and language pairs with less training data.

Resolution:

1. Add a hallucination prevention instruction to your custom prompt: "Translate only what appears in the source. Do not add explanations, interpretations, or content not present in the original."

2. Review short or highly ambiguous segments manually — these are most prone to hallucination

3. Consider switching to a more conservative AI engine for content where accuracy is critical

4. Set critical QA checks for segments that are significantly longer than the source — this can flag potential hallucinations

5. For segments with numbers, dates, or specific data, add: "Preserve all numbers, dates, and proper nouns exactly as they appear in the source."

For detailed guidance on hallucination prevention prompts, see Customizing AI Translation with Custom Prompts.

Issue 4 — Prompt optimization failures (prompt not producing expected output)

Symptoms: Custom prompt instructions are ignored or inconsistently applied. Output quality varies unpredictably. Prompt works for some language pairs but not others.

Cause: Prompt is too vague, too complex, conflicts with system-level instructions, or is not compatible with the selected engine.

Resolution:

1. Start simple: Reduce to one or two clear instructions and add complexity incrementally

2. Verify engine compatibility: Custom prompts only apply to LLM-based engines (GPT-4o, Claude, Gemini). Traditional MT engines ignore custom prompts entirely.

3. Test across language pairs: A prompt that works for Spanish may fail for Japanese. Test each target language separately.

4. Check for conflicts: Avoid instructions that contradict basic translation principles

5. Use specific, actionable language: Instead of "make it sound natural," use "use conversational tone with contractions where appropriate."

6. Document and escalate persistent issues: For problems that persist after optimization, document specific examples (source text, prompt, output) for SME or Support escalation

For detailed guidance on prompt writing, see Customizing AI Translation with Custom Prompts.

FAQs

My AI translations were good last week but quality has dropped. What changed?

Check whether your AI engine, custom prompt, glossary, or TM has changed recently. Any of these can affect output quality. Also check whether new document types have been added to the project that weren't tested with your current prompt.

Can I fix quality issues without a custom prompt?

Yes. Switching AI engines, improving your TM quality, and ensuring your glossary is complete all improve output without prompts. Prompts are the most powerful lever but not the only one.

The AI keeps translating a brand name even though I've added it to my glossary. Why?

Verify the glossary is connected to the project and the term is set to "Do Not Translate." For LLM engines, also add an explicit instruction in your custom prompt: "Do not translate [brand name]."

How do I know if my quality issues are caused by TM matches vs. AI translation?

Check the segment source in the editor. Segments matched from TM are labeled differently than AI-translated segments. If poor-quality segments are TM matches, the issue is TM quality, not AI translation. Consider cleaning your TM or adjusting match thresholds.

Related articles

• Running AI Translation — How to configure and run AI translation on your projects

• Customizing AI Translation with Custom Prompts — How to write effective custom prompts

• Understanding Translation Quality Score — How to configure and interpret TQS

• AI Translation Overview — The six-step AI translation pipeline

• Quality Assurance for Accurate Translations — QA check configuration and error types

• Post-edit AI Translation — The post-editing workflow for AI-translated content