7 thoughts/questions to start 2025: use of raw MT output

It is New Year’s Day 2025, and I am finalising my “Who’s in/or the lead in 2025” post at the moment. I decided this year to also try to distill some of my comments on LinkedIn into mini blogposts. In this format, I’ll post seven thoughts/questions, throw them open to the hive mind and then try to draw the responses together in a response post.

In recent years, I have seen a lot of posts pointing out particular machine translation errors. Their tone can vary wildly from “considered” to “downright dismissive”. The approach of the former will be to explain the shortcomings of the use of MT (in particular its raw output), and how there is more to consider than fluency that convinces a lay audience. The latter will often attack the kind of output you expect to find on social media sites belittling signs found in English around the non-English-speaking world.

Here are my seven thoughts on the use of raw MT output:

1. To what extent do professionals (i.e. people in “white collar” positions) actually trust raw output from MT?
2. If such a raw MT translation does go to print/screen, who is accountable for it?
3. Imagine the outcome results in something with fatal/lethal consequences. Presuming that there are multiple levels of sign-off. Who takes the responsibility?
4. How far away are we from litigation over translation quality when premium machine translation solutions make bold claims about accuracy?
5. At which point will output get worse as synthetic data swamps training of MT engines?
6. What does it take for output to be good enough/fit for purpose?
7. Should we educate the users rather than blame the machine?

This list was originally posted as a comment to a post on LinkedIn. Feel free to share your thoughts here or on LinkedIn.