GPT-4 Released: Another Step Up for LLM Translation

On March 14, 2023, OpenAI released GPT-4. Compared with the ChatGPT phenomenon (based on GPT-3.5), the new model improves markedly in reasoning, instruction following and multilingual understanding. Intensive industry testing followed immediately, and preliminary conclusions converged: GPT-4's translation ability is "another level up".

Standout areas in testing: literary and stylized text — noticeably finer control of tone, register and rhetoric, with much-improved "voice" in poetry and prose translation; specialist text — better than its predecessor on legal and medical benchmarks, with more stable terminology; low-resource languages — broad quality gains; and instructed translation — higher fidelity to complex instructions like "translate in a formal business register" or "preserve the humor of the original".

But industry sobriety deepened in parallel: hallucinations have not disappeared, long-document consistency still needs human control, and the data-privacy red line still stands. GPT-4 makes the "AI draft + expert review" workflow more usable; it does not change the fundamental structure in which humans answer for quality.

From GPT-3.5 to GPT-4 took four months. The pace of model iteration itself is the new normal the industry must adapt to.