AI 翻译产业全景图 · 巴别塔倒塌之后 · 2026-07 · 全系列的时间机器Translation × AI · After Babel · Jul 2026 · The Series’ Time Machine
你的十年后,是翻译的今天
Translation today is your tomorrow
想知道你的白领行业十年后什么样?看翻译业的今天。机器翻译已经跑了七十年——这个行业与「被 AI 替代」共处的时间比任何行业都长,它是 AI 冲击最长的一场临床试验;它走完的路(译员→译后编辑员→哑铃化),是所有靠信息中介吃饭者的预告片
Want to know your white-collar trade in ten years? Look at translation today. Machine translation has run for seventy years — no industry has lived with «AI replacement» longer; it is the longest clinical trial of the AI shock, and its road (translator → post-editor → the barbell) is the trailer for every information-brokering trade
本图相信:价格坍塌与岗位停滞是可量化的事实——美国口笔译岗位十年预测仅 +2%(慢于全职业平均且被下调,A);行业实值规模较 2019 峰值缩水 16.3%(B);36% 译者自报已因生成式 AI 失去工作(B)。三个指标互相印证。This map believes: price collapse and job stagnation are quantified facts — US interpreter/translator jobs projected at just +2% over a decade (below average, revised down, A); real industry revenue 16.3% under its 2019 peak (B); 36% of surveyed translators report work already lost to generative AI (B). Three gauges, one story.
本图不相信:「翻译行业将被消灭」——七十年里它经历四代技术、四轮「这次真要失业了」,既没被消灭也没回到旧丰饶;口译市场甚至逆势扩张(约 8% 年增)。真实过程是反复的「技术跃迁—岗位重估」循环,这次最陡,但仍是循环。It does not believe «translation will be abolished» — four technical generations and four panics in seventy years abolished nothing and restored nothing; interpreting even grows against the tide (~8% a year). The true process is a recurring loop of leap and repricing — steepest now, still a loop.
最小单元:一个意思的一次跨语言搬运。主脊:一次翻译任务的生命周期七环节(接入→译前→机器首译→MTPE→审校→交付→语料沉淀),③④标红——「机器首译」成默认第一步、「译后编辑」这个 2010 年代还不存在的工种已是全行业默认形态;三条结构带:七十年代际史、费率坍塌的临床数据、巴别塔倒塌的外部性;判断层:领域纵深 ＞ 语言本身,现场责任 ＞ 可编辑性。姊妹分工:外语教育坍塌→edu,跨境语言红利→trade,导游陪同→travel,字幕影视→media。
The atom: one meaning, moved once across a language. The spine: a job’s seven-step lifecycle (intake → prep → machine first draft → MTPE → review → delivery → corpus), ③ and ④ flagged — the machine draft is now the default first step, and post-editing, a trade that did not exist in the 2010s, is the industry’s default form; three bands: seventy years of generations, the clinical data of rate collapse, Babel’s externalities; the judgment: domain depth ＞ language, live accountability ＞ editability. Siblings: language education → edu, cross-border arbitrage → trade, guides → travel, subtitles → media.
70 年
机器翻译的年龄:1954 乔治城-IBM 实验(60 余句俄语、作者放言「3–5 年攻克」)→规则→统计→神经→大模型四代技术——每代都伴随一次质量跃迁与一次失业恐慌Machine translation’s age: from the 1954 Georgetown-IBM demo («solved in 3–5 years», its authors said) through rules → statistics → neural → LLM, four generations — each with a quality leap and a redundancy panic
−16.3%
行业实值规模较 2019 峰值的缩水(按通胀调整,B);同期美国口笔译岗位十年预测仅 +2%、被下调且慢于全职业平均(A)——量稳、单价降、岗位停Real industry revenue vs its 2019 peak (inflation-adjusted, B); US translator/interpreter jobs projected at just +2% a decade, revised down, below the all-occupation average (A) — volume steady, prices down, jobs flat
36%
英国作家协会 2024 调查:已因生成式 AI 失去工作的译者比例(787 份回复);43% 收入下降、77% 预期未来收入被拉低;另一行业学会调查:49% 工作量显著下降(B)Translators reporting work already lost to generative AI (Society of Authors 2024, 787 replies); 43% saw income fall, 77% expect worse; a sister institute found 49% with volumes sharply down (B)
+8%/年
反直觉硬数据:口译市场逆势扩张——2024 约 117 亿美元→2029 预计 171 亿(B);「现场、一次性、可担责」是笔译坍塌中最硬的人类据点The counter-intuitive line: interpreting grows against the tide — ~$11.7B (2024) → $17.1B projected (2029, B); «live, one-shot, accountable» is the hardest human redoubt amid the collapse of written work
⚠️ 口径裁判(先读):① 市场规模三口径并列勿混用——同一年「语言服务市场」三家权威估值相差近一倍(497 亿/727 亿/270 亿美元),口径不同(服务+技术/行业/可寻址),方向一致:传统按字计价被压缩;② 「千字价格二十年不涨反跌」方向可信但缺权威长期序列——实值 −16.3%、MTPE 折扣、每词 $0.05→$0.01 是方向证据,完整二十年中文语对序列未检索到,引用请标「业内广泛流传」;③ 厂商自测与第三方盲测直接冲突(一专用引擎自报胜通用大模型 1.7 倍 vs 独立成对盲测结论相反)——本页并列不采信单方;④ 部分中国数据(外语就业率 76.8%、译者收入 −70% 等)为单一转述无一手出处,按估算降权;⑤ 一 A 股概念勘误:「传神语联」挂牌新三板(NEEQ 835737),并非深市 300594;⑥ 四份深度研究交叉整理,一份全程带可核 URL 为主力;渗透%为编辑估值。
⚠️ Basis rulings (read first): ① three market sizes coexist — the same year’s «language services market» spans nearly 2× ($49.7B / $72.7B / $27.0B) on different scopes (services+tech / industry / addressable); direction agrees: per-word pricing compresses; ② «rates flat-to-down for twenty years» is directionally solid but lacks an authoritative series — real −16.3%, MTPE discounts and $0.05→$0.01 per word carry the direction; no full 20-year Chinese-pair series was found; ③ vendor self-tests clash with third-party blind tests head-on (one engine claims a 1.7× win over general LLMs; independent pairwise tests conclude the reverse) — this page juxtaposes, never adopts one side; ④ some China figures (76.8% employment rate, −70% income) are single-relay without primary sources, down-weighted; ⑤ a ticker correction: the «Transn» concept trades on NEEQ 835737, not Shenzhen 300594; ⑥ cross-compiled from four deep-research reports, one fully URL-verified as backbone; penetration %s are editorial.