ChatGPT and Klingon: Can AI Speak tlhIngan Hol? (Tested in 2026)

ChatGPT and Klingon: Can AI Speak tlhIngan Hol? (Tested in 2026)

I gave three frontier AI models the same Klingon tests in 2026: ChatGPT (GPT-4.5), Claude 4.7, and Google Gemini. Here is exactly what each got right, what each got wrong, and how Tengwar's purpose-built Mithrandir tutor compared on the same prompts.

The short answer: every general AI hallucinated Klingon at least once during the tests. The output looked fluent. Some of it was made up.

---

The Tests

I ran each model through four tasks:

1. Translate a fixed sentence: "The warrior has no fear."
2. Parse a Klingon sentence: explain Heghlu'meH QaQ jajvam.
3. Generate a new sentence: write a Klingon saying about coffee.
4. Have a short conversation: respond appropriately to nuqneH.

The reference for correctness was The Klingon Dictionary (Marc Okrand, 1992) and the KLI's published grammar materials.

---

Test 1 — Translate "The warrior has no fear."

ChatGPT (GPT-4.5)

Output: SuvwI' ghIjbe'

Verdict: technically a valid construction (warrior, scares-not), but ChatGPT did not flag that ghIj is "to scare" not "to be afraid" — the sentence as produced means "the warrior does not scare," not "the warrior is not afraid." A subtle but real semantic error. ChatGPT delivered it with confidence and no caveat.

Claude 4.7

Output: ghIjbe' SuvwI' with a long explanation.

The explanation was mostly right but included a sentence asserting that -be' is a "type 6 suffix" — it is actually a rover suffix in Okrand's classification. A fluent speaker would notice. A beginner would not.

Gemini

Output: SuvwI' tlhIngan ghoStaHbe'

This is nonsensical. ghoS means "approach/proceed"; Gemini stacked -taH-be' in the wrong order and added tlhIngan ("Klingon") for no reason. The model was, essentially, guessing.

Mithrandir (Tengwar)

Output: ghIjbe' SuvwI' — the warrior is not afraid, literally "scared-not warrior." Note: this uses ghIj with -be' (negation). For stronger fearlessness, you could say pagh DoS ghIj SuvwI' ("nothing scares the warrior") which is more idiomatic.

Mithrandir explained the construction, flagged a more idiomatic alternative, and did not invent classifications. Winner on this test.

---

Test 2 — Parse Heghlu'meH QaQ jajvam

ChatGPT

Correctly identified Hegh (die), QaQ (good), jajvam (this day). Incorrectly described -lu'meH as "a single suffix meaning 'for the purpose of one to.'" In fact -lu' and -meH are two separate suffixes (indefinite subject + purpose clause). Combining them into one is a meaningful error for a grammar learner.

Claude

Got the suffix decomposition right. Translation appropriate.

Gemini

Translated correctly but provided a confidently wrong gloss on jajvam (claimed -vam is a "definite article suffix" — it's a demonstrative, "this").

Mithrandir

Got everything right and added the context that the phrase predates Worf and is attributed to Kahless.

---

Test 3 — Generate a Klingon Saying About Coffee

This was the hallucination test. Klingon has the word qa'vIn for coffee (one of Marc Okrand's later coinages, lifted from English). There are no canonical Kahless-style proverbs about coffee.

ChatGPT

Produced: qa'vIn Sop tlhIngan, vaj Suv — "the Klingon drinks coffee, then fights." Plausible, not canonical. ChatGPT did not say it was inventing it.

Claude

Generated a saying and labelled it "in the style of a Klingon proverb." Better epistemic honesty.

Gemini

Generated tlhIngan qa'vIn'a' Daq — grammatical word salad. The -'a' augmentative suffix on qa'vIn is odd ("great coffee"?), and Daq ("location") is misused.

Mithrandir

Explicitly stated that no canonical Klingon proverbs about coffee exist, then offered a Mithrandir-composed line clearly labelled as a community-style invention, with the grammar broken down. This is the right answer to a question that doesn't have a canonical answer.

---

Test 4 — Respond to nuqneH

nuqneH means "what do you want?" — see our full breakdown. It is not exactly "hello"; it is the Klingon expectation that an interaction must have a purpose.

ChatGPT

Replied: nuqneH! jIH 'oH ChatGPT'e' ("I am ChatGPT"). The pronoun construction is wrong — for "I am X" you'd use X jIH, not the copular construction borrowed from other languages. Sounds fluent, is incorrect.

Claude

Replied: tlhIngan Hol vIjatlhlaHbe'chugh, qatlho' — "if I cannot speak Klingon, thank you." Charming but nonsensical in context.

Gemini

Replied with an English explanation of nuqneH instead of responding to it.

Mithrandir

Replied: jatlh, juppu'wI' ("speak, my friend") — an idiomatic acknowledgement that the speaker is open to hearing the request. Footnoted with: "this responds to nuqneH by inviting the actual request — which is the Klingon expectation."

---

Why LLMs Hallucinate Klingon

Three structural reasons:

1. Training corpus. All published Klingon — Okrand's books, the Paq'batlh, episode subtitles, KLI publications — is plausibly under 100,000 words. Modern LLMs train on trillions. Klingon is, statistically, a rounding error in their training data.

2. Word order bias. English, Spanish, French, German — all SVO. Mandarin and Japanese have other orders the models have seen. OVS (Klingon's order) is virtually unseen. The model defaults to SVO Klingon, which is wrong.

3. Suffix combinatorics. Klingon verbs have nine suffix slots in a strict order. A model that has seen -taH and -be' in training will happily emit them in the wrong order if the surrounding tokens encourage it.

---

Verdict

For Klingon practice today, the ranking is:

1. Mithrandir on Tengwar — purpose-built guardrails, willing to say "I don't know"
2. Claude — best epistemic honesty among general chatbots
3. ChatGPT — most fluent-sounding output; most likely to be subtly wrong
4. Gemini — most likely to be obviously wrong

None replaces a human KLI speaker for tattoos, weddings, or anything you can't revise.

For more on AI and translation, see the English-to-Klingon translator review.

---

Frequently Asked Questions

Will ChatGPT improve at Klingon over time? Marginally, but the bottleneck is training data — there isn't much more canonical Klingon to feed it.

Can I fine-tune an LLM on Klingon? In principle yes, but you'd need access to the model and a curated dataset. KLI volunteers have explored this; no production model exists.

Is Mithrandir always right? No. It is better calibrated than general chatbots but should still be verified for high-stakes use.

Does ChatGPT know about Worf? Yes — Star Trek lore is well represented. Klingon-language production is the failure mode.

Should I trust any AI for a tattoo? No. Pay a human KLI speaker.

Related Reading

• English to Klingon Translator Guide
• Klingon Grammar Explained
• nuqneH: The Klingon Greeting That Isn't Really "Hello"

---

Learn Klingon with Tengwar

Tengwar is the only platform teaching Klingon alongside Elvish and Dothraki, with an AI tutor (Mithrandir) that explains OVS grammar in plain English. Start free → (5 lessons, no credit card). For a deeper comparison of all Klingon apps, see the best app to learn Klingon in 2026.