In a recent issue of The Economist I came across an article entitled The Babel wish. The subheading says: Machine translation is fast becoming a solved problem. Making it perfect will be a very hard problem. I would not disagree with the latter, but would add that making human translation perfect has always been a hard problem – and continues to be so. (And what is ‘perfect’, anyway?) As a linguist you will know how difficult it is to agree on a clear definition of what ‘quality’ is – and how divergent opinions can be.

The article starts out by dispelling the myth that we translators have been clinging on to, namely that to make a machine translation do a satisfactory job, there will always need to be a human in the loop.

The author recounts how a Lisbon-based startup, Unbabel, pitted its AI model against its own human translators. The competition led to the pretty damning conclusion “Humans are done in translation”. While currently it is estimated that some 95% of global translation requirements are handled by humans, they reckon that over the next 3 years this will be reduced to zero (no indication is given how much of the cited 95% are already relying on assistance by MT/AI.)

Reaching zero would mean leaving MT/AI to run in Autonomous Mode, with no human intervention required. And I think everyone – translators and clients alike – are reluctant to accept this mode of working, although their motives are different.  While many translators would happily go back to the traditional mode, where they take responsibility for their output, it is the in-between mode that we are currently saddled with. We might call it Collaborative mode. This involves the machine producing output (with more or less professional help, such as TBs, style guides, prompts …), followed by a Human going over it with a coarser or finer comb, to find the errors or faux-pas.

If we turn back to the “Babel” article, we find the author turning to Marco Trombetti (CEO of Translated, Rome) and his method for measuring the quality of MT output, TTE (Time to Edit). Trombetti claims that between 2017 and 2022 TTE dropped from 3 seconds per word to 2 (in the world’s most translated languages). He predicts that this will fall to just 1 second within the next 2 years.

Once that point is upon us, he says, the human contribution will be nothing more than to provide a “moral crumple zone” – a concept coined by Madeleine Clare Elish at Google Cloud, intended to be shorthand for “a face to take the blame when things go wrong, but with no reasonable expectation of improving outcomes”.

I was baffled by this concept. Yes, I know about the crumple zone, a German invention meant to protect passengers by dissipating the energies released by a crash. But ‘moral’?

Intrigued by this, I googled “moral crumple zone” and “Madeleine Clare Elish”. This took me straight to her paper “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction” (2019, see https://estsjournal.org/index.php/ests/article/view/260). Its keywords being  autonomous vehicles, responsibility, machine learning, human factors, accidents, social perceptions of technology, self-driving cars, robot, human-in-the-loop, human-robot interaction, it is clear that the focus is on who has to take responsibility if accidents occur when AI is used in automated systems – mainly with vehicles, but presumably also in operating theatres, or with technical equipment. Very interestingly, Dr. Elish describes a reversal that’s taking place: Whereas “… the crumple zone in a car is designed to absorb the force of impact in a crash, the human in a highly complex and automated system may become simply a component … that bears the brunt of the moral and legal responsibilities when the overall system malfunctions.”

So this is important: The crumple zone in the car protects the human driver (with car parts taking the hit), while the moral crumple zone protects the integrity of the technological system thus making the human operator the victim, absolving the machine! This clearly has lots of legal implications and is a real conundrum, as both Google and Tesla have experienced. It raises the question of responsibility: Who is at fault, who pays, who is taken to court, when an autonomous car crashes into bus?

It seems only logical that in the partnership between (wo)man and machine, the moral responsibility falls on the agent who is capable of taking a moral stance, has ethical standards at heart and can admit (and regret) errors – and therefore assume responsibility and blame.

In the case of translation/localization providers, that agent is no other than the translator.

Even though you may hear “oh, but I got this from the translation memory”, or “this came from DeepL” – no serious professional would take this as a legitimate excuse for a mistranslation. A segment that you wave through or confirm is a segment that you are responsible for.

I think the clients see it that way, never mind who provided the MT to be edited.

A very common scenario these days is for a client to provide raw, pre-translated target texts to their localization vendor. The idea is that Lite Post-Editing is all that is needed: “Just go over it to make sure there are no serious mistakes that would be misleading for the user or be embarrassing on a technical or cultural level.” Perhaps they have done tests on machine-translated output of some of their materials and are confident that a small amount of human validation will be sufficient to obtain a satisfactory outcome. They have been hugely impressed by the quality that MT can now achieve, be it through their own experience or by hearsay. They are keen therefore to reap the benefits that will arise in terms of cost and time savings.

If the outcome is indeed satisfactory, they will be quick to sing the praise of the automatic system. I am afraid I cannot report that post-editors of MT texts get applauded any more than the human translator who starts from scratch (I am talking about ‘commercial’ translation here, exempting literary ones, who are in a different league).

We have a situation therefore where the beneficiaries are the clients, who get their translations at a low price and with considerably shorter turnaround times, and the machines, who get praise for their almost-human – or better-than-human – translation quality. The one party who gets the blame when things go wrong, is – inevitably – the ‘linguist’ (translator, post-editor, reviewer …). It is us translators therefore that are lumbered with the responsibility. Given that we have no or very little control over the behaviour of the automated system, particularly when that system is under the client’s control and not ours, seems – shall we say – rather unfair.

Going back to Mr. Trombetti’s prediction that 1 sec. only will need to be spent on validating a machine-translated text, a simple calculation will tell us that we will have to brace ourselves for turning around 3600 words per hour and 28,800 per 8-hour day. A pretty daunting prospect if you ask me. Clearly, the faster you go, the higher the risk that you will overlook errors. So perhaps, alarm bells should be ringing?

With self-driving cars the jury still seems to be out as to who is the responsible agent – the human who is travelling in the vehicle which is in ‘autopilot’ mode and is meant to ‘supervise’ the car and take over in case that’s needed – or the designers and engineers who invented the thing and who are responsible for modifications and updates – the close cooperation between the MT/AI and the human translator or the post-editor – the case seems settled for the translation environment: It is always the humans on the job that will be blamed, never mind how little time they are allowed. And trying to get control over a system whose behaviour is barely understood by its developers, is of course extremely difficult.

I think that this is what we fear as professional translators: That it is getting more and more difficult, near impossible, to act in accordance with our long-cherished principles. If you are constantly under time and money pressure acting responsibly and ethically might slowly get compromised.

Accountability calls for a measure of control. We cannot allow the machine to take control and leave us humans to take the blame. It is worth mentioning that one of the fields where this is causing great concern is the medical sector, where clinicians are held responsible when things go wrong with AI-enabled operating systems.

Clearly, we need to find ways of not just providing the moral crumple zone – whether it is in situations of life and death or just the risk of an inadequate translation.

So is it all doom and gloom? Of course not.