Every argument about automatic translation since 1949 has turned on a comparison with a human translator, and the comparison is usually made badly. Machine and manual translation are set against each other as if they were two ways of doing one job to one standard. They are not. The useful question is not which is better but which failure you can afford.
Machine translation
Its strengths are volume, speed and cost. A system translates millions of words in the time a human needs for a page, at a marginal cost close to zero, across dozens of language pairs at once. For understanding what a document says — gist translation — that is usually enough, and it is the use the 1966 committee judged uninteresting and the market later proved was the largest one.
Its weaknesses changed shape over the decades but never disappeared. Rule-based systems failed visibly: untranslated tokens, mangled word order, output nobody could mistake for finished work. Neural systems fail fluently instead, producing well-formed sentences that are confidently wrong. The error rate fell; the cost of detecting an error rose, because the output no longer announces its own failure.
What no system handles reliably is what Bar-Hillel identified in 1960: meaning that depends on knowing how the world works rather than on how the sentence is built. Ambiguity resolved by context outside the text is still the boundary.
The manual translation
A human translator brings the thing machines lack: a model of what the text is for and who will read it. That matters most where the stakes are not comprehension but consequence — legal instruments, medical documentation, marketing copy, literature. In each of these the right output is not the closest rendering of the source but the text that does the same job in the target language, which sometimes means departing from the source considerably.
The costs are the obvious ones: time, money, and availability in the pair you need. The 1966 committee measured exactly this and found, for Russian scientific material in the United States, that the supply of translators exceeded demand. That finding was specific to a language, a subject and a decade, and it has been generalised far beyond what it supports.
Post-editing, and why the comparison is a false choice
In practice most professional work is now neither one nor the other. Machine output is produced first and a translator edits it, at a rate somewhere between revision and retranslation depending on quality. This is the arrangement the ALPAC committee actually recommended — tools that assist the translator rather than replace them — arrived at sixty years later by a different route.
Post-editing also explains why the cost comparison keeps shifting. Whether machine-plus-editor beats a translator working alone depends entirely on how good the raw output is for that pair and that text type, and that varies more between domains than between the methods themselves.
Choosing between them
A rough division holds up well. High volume, low stakes, comprehension the goal: machine. Low volume, high stakes, the text itself is the product: human. Everything in between: machine output with human editing, priced by how much editing it needs.
The compendium of translation software records what the systems themselves could do in each era, and the turning points trace how the boundary between the two has moved.
Is machine translation accurate enough for professional use?
For gist and for internal comprehension, generally yes. For published work, it is a first draft. The distinction is not about the percentage of correct words but about whether an undetected error is tolerable.
Why does neural output feel better but still get things wrong?
Because fluency and accuracy are separate properties. A neural system is optimised to produce text that looks like the target language, and it will produce a plausible sentence rather than signal uncertainty.
Does post-editing take as long as translating?
It depends on the pair and the domain. For technical text in a well-resourced pair it is considerably faster; for literary or persuasive text, editing poor output can take longer than starting fresh.
Which texts should never go through a machine alone?
Anything where an error carries legal, medical or safety consequences, and anything where the effect on the reader is the point rather than the information content.