17-04-2026, 11:23 PM
A quick note:
I know I brought up the cross-check between two LLMs myself. However, I mainly use it as a "numerical" plausibility check – by running the same count or statistical analysis through Claude and ChatGPT to detect hallucinations. It is unlikely that both LLMs would produce the same hallucinations – i.e. the same values – during a count or statistical analysis. This has allowed me to spot many minor errors, with ChatGPT hallucinating significantly more and constantly refusing to actually count, simply spitting out random numbers, compared to Claude. This is also because I have internally prohibited Claude from hallucinating.
It is far more critical to use LLMs for translation.
When it comes to languages, LLMs sometimes have the same training data and will therefore hallucinate similar patterns. So as soon as two LLMs agree on the phrasing based on their training data, I can no longer distinguish between what is a hallucination and what is the truth. Gemini and Google Translate in particular get very creative when they don’t know what they’re dealing with.
That’s why I try to keep this part manual, or check against real dictionaries as soon as a decent level of accuracy is required. Since I actually do this, I’ve also noticed just how much hallucination appears in the translations, particularly with Gemini, but also with ChatGPT and even Claude.
As I said, just a note.
I know I brought up the cross-check between two LLMs myself. However, I mainly use it as a "numerical" plausibility check – by running the same count or statistical analysis through Claude and ChatGPT to detect hallucinations. It is unlikely that both LLMs would produce the same hallucinations – i.e. the same values – during a count or statistical analysis. This has allowed me to spot many minor errors, with ChatGPT hallucinating significantly more and constantly refusing to actually count, simply spitting out random numbers, compared to Claude. This is also because I have internally prohibited Claude from hallucinating.
It is far more critical to use LLMs for translation.
When it comes to languages, LLMs sometimes have the same training data and will therefore hallucinate similar patterns. So as soon as two LLMs agree on the phrasing based on their training data, I can no longer distinguish between what is a hallucination and what is the truth. Gemini and Google Translate in particular get very creative when they don’t know what they’re dealing with.
That’s why I try to keep this part manual, or check against real dictionaries as soon as a decent level of accuracy is required. Since I actually do this, I’ve also noticed just how much hallucination appears in the translations, particularly with Gemini, but also with ChatGPT and even Claude.
As I said, just a note.
![[+]](https://www.voynich.ninja/images/collapse_collapsed.png)