What a translated transcript is good for, and what it is not

Machine translation of a meeting transcript has become good enough that the interesting question is no longer whether to use it. It is which decisions you are willing to make on top of it.

The honest answer is that it is excellent for some things and unsuitable for others, and the boundary is not where people usually assume.

What it does well

Telling you whether you need to care. The most common use of a translated transcript is triage: forty minutes of discussion, and you want to know whether anything in it affects you. Translation is more than good enough for this. An occasional clumsy phrase does not stop you identifying the two minutes worth reading closely.

Recovering the shape of a discussion. Who raised what, roughly when, and what the disagreement was about. These survive translation well because they are structural rather than lexical.

Finding the passage you half-remember. Searching a translated transcript to locate a moment, then reading the original at that point, is a workflow that plays to the strengths of both.

What it does badly

Terms of art. Every field has vocabulary where the everyday translation is wrong. A word that means one thing in general use and something specific in accounting, construction, or medicine will frequently arrive as the general meaning, and the sentence will read perfectly while saying something else.

Obligation and hedging. The distinction between "we will", "we should", and "we could probably" is exactly the distinction a commitment turns on, and it is carried by small words that translation handles inconsistently. A translated transcript will sometimes firm up a maybe and sometimes soften a promise. Neither error announces itself.

Register and intent. Whether somebody was being sardonic, deferential, or blunt is often the thing that tells you what they meant. Very little of it survives.

Anything a person will later dispute. Not because translation is especially bad at these passages, but because the argument will be about the words somebody used, and a translation is by definition not those words.

The rule that falls out of this

Use the translation to navigate and the original to conclude.

Concretely: it is fine for a translated transcript to be how you find the relevant part of a meeting, how you brief somebody who was not there, and how you keep track of what a team in another language is working on. It is not fine for it to be the only surviving record of a meeting where somebody committed to something, or where a decision was contested, or where the exact terms matter.

That is not a caution about machine translation specifically. A human interpreter's summary carries the same limitation; it is simply more obvious that a person was involved.

Which means the original has to survive

The failure mode worth designing against is not a bad translation. It is a good translation that becomes the only document, because it was the one everybody could read, and the original was discarded as an intermediate artifact.

Six months later the translation is the record. It reads fluently, nobody can check it, and the sentence somebody actually said is gone. This is a worse outcome than having no transcript at all, because a missing record produces caution and a confident wrong one does not.

Keep both, and make it obvious which is which.

Where DayMarshal fits

DayMarshal retains the transcript in the language the meeting was held in, and shows the English translation alongside it rather than in place of it. Summaries, action items, decisions, and commitments extracted afterwards link back to the passage of original transcript they were drawn from, so the reading and the record stay distinguishable — which is the whole of the discipline above, made structural rather than left to somebody's discretion at the end of a long day.