Machine translation is now used almost universally. This raises the issue of assessing the quality of the resulting product. This is partly done with automated algorithms and partly by humans, in the traditional way. Here we will present the approaches we currently use. Language quality This is the traditional approach to assess translations. The best-known tool in this respect is LISA QA, where mistakes are graded by severity, and by type. A major problem with this kind of assessment is that it leads to spot far more errors in machine translations than in human translations. Edit distance Edit distance is interpreted and used differently from the original scope. Some see it as the amount of text that needs to be corrected; others see it as the amount of time to spend to bring the text up to a required quality level. We follow the second approach. In theory, a comparison of edit distances would help identify the best machine translation. It is essential to have a clear understanding of the reason for selecting a specific machine translation program, whether to provide a usable unedited text, or a translation that will then be edited. In the first case, the key is the amount of text to edit (i.e., the number and severity of errors), assuming that the translation would be assessed following the traditional approach. In the other case, machine translation should be seen as a tool and the main issue would be how much it increases the process effectiveness, i.e., how much time must be spent on editing. In any case, human assessment is prone to subjectiveness. Linguists have all their own strengths and weaknesses. For one, terminology may be a weak point and a lot of time would be spent looking for the right term. Another would take longer on grammar and fluency. Estimating the edit distance Edit distances can be reckoned precisely only downstream and is expensive and time-consuming. We therefore decided to estimate the amount of time to edit a text, rather than actually computing it. To this end, we identified three types of errors, based on how long it would take to correct them. Coefficients applies to errors based on the following logic: the time to correct a comma is assumed to be less than that to correct a grammatical error and even more an unedited fragment. Human error are the main source of inaccuracies in the assessment process. We found that despite changing the criteria used by editors, they continued to assess translations based on linguistic criteria. In fact, in all the batches analyzed, scores correlate with the results of assessment in terms of linguistic quality rather than the actual edit distance. Percentage of similarity Another widely used method involves comparing an unedited translation with an edited version. While this idea may at first seem attractive, as the fewer the corrections, the better the original translation, in practice we found this approach ineffective. The algorithms to assess similarity levels are often rather imprecise. For example, they frequently treat changing a capital letter to a lower-case one as equivalent to changing a whole word. Also, editors often use the filter options in CAT tools. For example, a text may contain a dozen examples of a term, but the editor changes it using the find and replace function. This takes just seconds to do, but the volume of edited text may be substantial. Automated quality assessment tools These days there is a great deal of talk about BLEU, hLepor, COMET and CHRF+. Indeed, these algorithms do not express the quality of a machine translation output. They do allow users to compare dozens of alternative machine translation platforms very quickly, but they are not by any means translation quality tools. Also, the results of comparisons can easily be misinterpreted. Most automatic scoring algorithms express the similarity of a sample output to a reference text, typically a human translation. This means that, if the same tool is given two different human translations of the same text, which do not use the phrases contained in the reference corpus, these translations may be rated quite differently, possibly worse than the sample machine translation output. This initially led us to false conclusions. The best way to use this kind of tools is to measure the increase in scoring by comparing the output of a trimmed engine with one from the old one. A comparison of many different machine translation systems may lead to the conclusion that the system that gives the best results is the one that was used for the reference translation. Another clear drawback is the need to have a reference translation to run a comparison. Conclusion Automation is getting unavoidable in our shared future. Possibly, it is only a matter of time until we have reliable automated assessment tools. Yet, it is still too early to rely on them entirely. The use of human-operated tools and assessment by human specialists will remain the standard methods for assessing the quality of machine translations for the next few years, the key being an intelligent and critical approach to the whole process.