docs(crowdin-ai): document AI translation limitations and model outpu…

src/content/docs/crowdin/translation-process/crowdin-ai.mdx

@@ -485,6 +485,24 @@ Advisor String Context Review uses AI to evaluate the quality of the context pro
   href="/advisors/"
 />
 
+## AI Translation Limitations
+
+AI-powered translations rely on large language models, so their output is probabilistic and its accuracy can vary depending on the language pair and content complexity. Keep the following behaviors in mind when using auto-translation and AI suggestions.
+
+### Wrong-Language Output and Hallucinations
+
+Occasionally, an AI model may return a translation in the wrong language or script, or produce text that doesn't accurately reflect the source (a hallucination). This is a general limitation of large language models rather than a Crowdin-specific issue, and it's more likely for less common language pairs or when little context is provided. For example, when translating into a target language that uses a non-Latin script, a model might occasionally return characters from a different script.
+
+**Best Practices:**
+
+- Select a different model or AI provider for the prompt.
+- [Give the prompt more context](#configuring-ai-prompts), such as file context, project context, previous and next strings, glossary terms, TM suggestions, and style guides.
+- Always review AI-generated translations before approving them, and use [QA checks](/project-settings/qa-checks/) (including [AI QA Check](#ai-qa-check)) to catch issues.
+
+### Identical Output Across Different Models
+
+Different models don't always produce different translations. When you use the same prompt configuration and context, different models may return identical output, especially for short, unambiguous, or well-defined strings. When comparing models to choose the best fit, test them on representative content from your project so that identical results aren't mistaken for a misconfiguration.
+
 ## See Also
 
 <LinkCard

src/content/docs/enterprise/translation-process/crowdin-ai.mdx

@@ -522,6 +522,24 @@ Advisor String Context Review uses AI to evaluate the quality of the context pro
   href="/enterprise/advisors/"
 />
 
+## AI Translation Limitations
+
+AI-powered translations rely on large language models, so their output is probabilistic and its accuracy can vary depending on the language pair and content complexity. Keep the following behaviors in mind when using auto-translation and AI suggestions.
+
+### Wrong-Language Output and Hallucinations
+
+Occasionally, an AI model may return a translation in the wrong language or script, or produce text that doesn't accurately reflect the source (a hallucination). This is a general limitation of large language models rather than a Crowdin-specific issue, and it's more likely for less common language pairs or when little context is provided. For example, when translating into a target language that uses a non-Latin script, a model might occasionally return characters from a different script.
+
+**Best Practices:**
+
+- Select a different model or AI provider for the prompt.
+- [Give the prompt more context](#configuring-ai-prompts), such as file context, project context, previous and next strings, glossary terms, TM suggestions, and style guides.
+- Always review AI-generated translations before approving them, and use [QA checks](/enterprise/project-settings/qa-checks/) (including [AI QA Check](#ai-qa-check)) to catch issues.
+
+### Identical Output Across Different Models
+
+Different models don't always produce different translations. When you use the same prompt configuration and context, different models may return identical output, especially for short, unambiguous, or well-defined strings. When comparing models to choose the best fit, test them on representative content from your project so that identical results aren't mistaken for a misconfiguration.
+
 ## See Also
 
 <LinkCard