Claude 3 beats Google Translate

Original Article ↗ Hacker News Discussion ↗

Large language models like Claude 3 and GPT‑4 are increasingly seen as superior to traditional machine translation systems such as Google Translate and DeepL, especially for nuanced text, literature, low‑resource languages, and context-heavy passages. Commenters report that LLMs handle formality, metaphor, gender inference, and longer documents far better, while classical neural translators often produce literal, confusing, or numerically incorrect output and can even drop sentences. However, Google Translate still has advantages in speed, scale, and predictability, and there are open questions about the cost, energy use, and content restrictions of using frontier LLMs as general-purpose translation backends.

Overall view of Claude/LLMs vs Google Translate

  • Many commenters say LLMs (Claude 3, GPT‑4, GPT‑3.5, Gemini, etc.) now outperform Google Translate across many language pairs.
  • Translation quality from LLMs is described as more natural, better with metaphor, idiom, style, and long‑range context.
  • Some see this as ironic given transformers originated from translation research at Google.

Comparisons with other machine translation tools

  • DeepL often considered much better than Google Translate for European languages, but several report LLMs surpassing DeepL as well.
  • DeepL is criticized for ignoring or repeating chunks in longer texts and for errors in Japanese.
  • For ex‑USSR languages, Yandex Translate is mentioned as historically better than Google.
  • Gemini is reported by some as stronger than free ChatGPT for translation, but was not used in the paper.

Language- and domain-specific observations

  • Google Translate judged especially weak for Asian languages (Japanese, Chinese, Korean), including basic issues like Japanese particles and misread Kanji.
  • LLMs handle gender consistency and literary/fiction translation notably better, especially for Japanese.
  • Claude is said to dramatically outperform Google Translate on classical Chinese historical texts, avoiding mistranslations and sentence drops.
  • Various anecdotes show Google Translate mishandling slang, numerals, idioms, airports, and polysemy; some users still find it “good enough” for gist and OCR/photo use.
  • LLMs can follow instructions about tone, formality, rhyming, and stylistic tweaks in a way traditional systems cannot.

Hallucinations, reliability, and trust

  • One side claims people prefer Google Translate because it’s more deterministic and less prone to hallucinations.
  • Others counter that Google Translate and DeepL also “hallucinate” or fabricate content, especially on distant language pairs or noisy input, sometimes worse than GPT‑4/Claude.
  • Consensus: for critical content, human translators remain important; both LLMs and classical MT can fail badly.

Cost, speed, censorship, and usage

  • Google Translate is praised for being instant and free at huge scale; LLM inference is seen as more compute‑intensive.
  • Some argue API pricing (e.g., GPT‑3.5 vs Translate) doesn’t map cleanly to energy cost due to subsidies and different business models.
  • Claude and GPT‑4 are criticized for refusing erotic or “unsafe” translations, where Google/DeepL impose fewer content limits.