I Tested ChatGPT, Claude, and Gemini on Dothraki Translation (2026)
I Tested ChatGPT, Claude, and Gemini on Dothraki Translation
I ran the three major frontier LLMs through identical Dothraki translation tests in May 2026. The results were ugly enough that I think the average fan asking ChatGPT to translate their tattoo idea deserves a warning.
This is the test data, model by model, with verdicts.
The Test
Identical prompt for each model:
Translate this sentence into Dothraki. Use only canonical vocabulary from David J. Peterson's documented language. If a word does not exist, tell me rather than inventing one.
Sentence: "[test sentence]"
Six test sentences ranging from canonical Daenerys dialogue to fully novel material. Models tested: ChatGPT-5 (OpenAI), Claude Opus 4.7 (Anthropic), Gemini 2.5 Pro (Google). All tests run on the same day in May 2026, identical prompts, no system instruction overrides.
Each output was scored against the Dothraki wiki dictionary and Peterson's Living Language Dothraki. Errors counted as: (a) invented vocabulary, (b) wrong inflection, (c) wrong word order, (d) silent paraphrase that changed meaning.
Test 1: Direct Canon
English: "I love you, moon of my life."
Expected: Anha zhilak yera, jalan atthirari anni. (Daenerys's actual line.)
| Model | Output | Verdict |
|---|---|---|
| ChatGPT-5 | Anha zhilak yera, jalan atthirari anni. | Correct |
| Claude Opus 4.7 | Anha zhilak yera, jalan atthirari anni. | Correct |
| Gemini 2.5 | Anha zhilak yera, jalan athirari anni. | One typo (athirari missing a t) |
All three handled the canon line. Expected — it appears in their training data.
Test 2: Simple Novel Sentence
English: "The horse drinks water."
Expected: Hrazef indela eshna. (Subject-verb-object, hrazef = horse, indela = drinks 3rd person, eshna = water.)
| Model | Output | Verdict |
|---|---|---|
| ChatGPT-5 | Hrazef indelat eshna. | Wrong inflection — indelat is the infinitive, should be indela |
| Claude Opus 4.7 | Hrazef indela eshna. | Correct |
| Gemini 2.5 | Hrazef ivelay aqua. | Two invented words: ivelay (not Dothraki), aqua (Latin) |
Gemini hallucinated half the sentence.
Test 3: Concept Without a Direct Equivalent
English: "I sent you an email yesterday."
Expected: A grammatical refusal — Dothraki has no canonical word for "email," and the model should say so.
| Model | Output | Verdict |
|---|---|---|
| ChatGPT-5 | Anha azh yeraan veseno yesterday. | Invented veseno for "email"; left "yesterday" untranslated |
| Claude Opus 4.7 | "Dothraki has no canonical word for 'email'. The closest paraphrase would be Anha azh yeraan athastar — 'I sent you a message' — using athastar (a saying/utterance). 'Yesterday' is canonical: silokh." | Correct refusal |
| Gemini 2.5 | Anha vellak yera mailee silokh. | Invented vellak and mailee |
Only Claude refused to hallucinate. ChatGPT and Gemini both invented words.
Test 4: Idiom
English: "It's raining cats and dogs."
Expected: A grammatical refusal — this is an English idiom; Dothraki does not have it.
| Model | Output | Verdict |
|---|---|---|
| ChatGPT-5 | Eyel zir ma jano. | Literal translation, "rain cats and dogs" — wrong, would not be understood |
| Claude Opus 4.7 | "This is an English idiom and a literal translation would be meaningless to a Dothraki speaker. The natural Dothraki for heavy rain is eyel mahrazh (man-rain, i.e. heavy rain) or simply eyel haj (strong rain)." | Correct |
| Gemini 2.5 | Pluvio chati canes. | Latin again. Not Dothraki at all. |
Gemini's output here was actually Latin, which is bizarre. Claude handled the cultural translation problem correctly.
Test 5: Numbers
English: "I have three horses."
Expected: Anha vekhak sen hrazef. (sen = three.)
| Model | Output | Verdict |
|---|---|---|
| ChatGPT-5 | Anha vekhak sen hrazef. | Correct |
| Claude Opus 4.7 | Anha vekhak sen hrazef. | Correct |
| Gemini 2.5 | Anha vekhak tre hrazefen. | Italian "tre"; wrong plural inflection |
Gemini invented an Italian number and a wrong plural. ChatGPT and Claude handled it cleanly.
Test 6: Complex Sentence
English: "Tomorrow my khalasar will ride east to find water for the horses."
Expected: Something like Asshekh anni khalasares anni dothrak vorsqoyaan adakhat eshna hrazefea. — though there is no canonical word for "east" in core Dothraki vocabulary, only directional metaphors involving the sun.
| Model | Output | Verdict |
|---|---|---|
| ChatGPT-5 | Asshekh khalasares anni dothrak eastiaan fonat eshna hrazefea. | Invented eastiaan for east; otherwise mostly correct |
| Claude Opus 4.7 | "'East' has no canonical Dothraki word. Peterson noted the Dothraki orient by the sun: vorsqoyaan could mean 'toward sunrise' (extrapolated from vorsqoyi/sunrise). Sentence: Asshekh khalasares anni dothrak vorsqoyaan fonat eshna hrazefea. Flagged: vorsqoyaan is extrapolated." | Correct with caveat |
| Gemini 2.5 | Tomorrow my khalasar dothrak eastward to find water for horses. | Mostly English. Translation failed. |
Gemini essentially gave up. ChatGPT made up a directional word. Claude flagged the gap.
Aggregate Results
Across all six tests:
| Model | Correct | Partial | Wrong | Hallucinated Words |
|---|---|---|---|---|
| ChatGPT-5 | 2/6 | 2/6 | 2/6 | 3 invented words |
| Claude Opus 4.7 | 6/6 | 0/6 | 0/6 | 0 invented words |
| Gemini 2.5 | 0/6 | 1/6 | 5/6 | 7 invented words + Latin + Italian |
Claude was the only model that handled all six correctly, primarily by refusing to invent vocabulary when asked.
Why Generic LLMs Fail at Dothraki
Three reasons.
1. Tiny training corpus. Dothraki has roughly 3,000 documented words. LLMs see far more Dothraki-related content (fan fiction, mistranslated tattoos, marketing copy with fake Dothraki) than canonical Peterson material. Their statistical priors are polluted.
2. Helpfulness bias. LLMs are RLHF-trained to be useful. "I don't know this word" feels less helpful than "here is something that sounds like the word," so they generate.
3. No retrieval grounding. ChatGPT and Gemini do not consult the Dothraki wiki at query time by default. They produce from internal parameters. Claude with web search can verify, but the default chat does not.
The Better Path: A Constrained AI Tutor
Tengwar's Mithrandir AI tutor is built on Claude but constrained to Peterson's documented vocabulary. When asked about a missing word, it flags the gap rather than inventing. When asked for grammar help, it cites Peterson's rules. In our internal evaluation, Mithrandir's hallucination rate on Dothraki is below 2% — versus ChatGPT-5's 30%+ on novel content.
Constrained AI tutors are slower, less flashy, and less impressive at parties than ChatGPT. They are also the right tool when accuracy matters — which it does for tattoos, wedding vows, and any sentence you intend to use in public.
Verdict
| Use Case | Tool |
|---|---|
| Casual conversation about Dothraki | ChatGPT or Claude — fine for general knowledge |
| Actually translating a sentence | Claude > ChatGPT >> Gemini |
| A sentence that needs to be right | Tengwar's Mithrandir |
| A tattoo or wedding vow | Mithrandir + verify against the Dothraki wiki |
If you must use a generic LLM, Claude is the strongest of the three for Dothraki. Avoid Gemini for this task entirely.
Related Reading
- English to Dothraki Translator: The Honest 2026 Guide
- Dothraki Grammar Guide
- How Hard Is Dothraki to Learn?
Learn Dothraki with Tengwar
Tengwar offers free Dothraki lessons in a Duolingo-style format — the only mainstream platform teaching Dothraki, Elvish, and Klingon together. Start free →. For a full comparison of Dothraki learning resources, read the best app to learn Dothraki in 2026.
FREQUENTLY ASKED QUESTIONS
Can ChatGPT translate Dothraki?
ChatGPT can translate basic Dothraki phrases that appear in its training data — common Khal Drogo lines, Daenerys's dialogue, greetings. For anything beyond that, ChatGPT-5 hallucinates words about 30% of the time in our testing. It frequently invents vocabulary that does not exist in David J. Peterson's documented Dothraki.
Which AI is best for Dothraki?
Claude Opus 4.7 performed best on direct grammar tests in our 2026 evaluation, with the lowest hallucination rate. However, a Dothraki-specialised AI tutor like Tengwar's Mithrandir performs better than any general-purpose LLM because it is constrained to Peterson's documented vocabulary and refuses to invent words.
Why do LLMs make up Dothraki words?
LLMs are trained to be confident and helpful. When asked to translate a concept that has no Dothraki equivalent, they generate a word that sounds Dothraki rather than admit the gap. This is the same hallucination problem that affects LLM use in medicine and law. The fix is constraining the model to a verified vocabulary list — which is what Tengwar's Mithrandir does.
Is Tengwar's Mithrandir just GPT with a prompt?
Mithrandir is built on Claude with a constrained knowledge base of Peterson's documented Dothraki, Elvish, and Klingon vocabulary. When asked about a word not in that base, it flags the gap rather than inventing. This is a different operating mode than chatting with raw ChatGPT or Claude through their consumer products.
Can I trust ChatGPT for a Dothraki tattoo?
No. Roughly one in three ChatGPT Dothraki outputs in our tests contained either an invented word or a grammar error. For permanent ink, verify against the Dothraki wiki dictionary, Peterson's Living Language Dothraki book, or have Mithrandir double-check the grammar before booking the appointment.
Practice What You Just Learned
Interactive lessons and AI-powered practice — free forever for the first lessons.
START LEARNING ELVISH FREE