Beyond ‘Good Enough’: Asking the Harder Question About AI Translation
Much of the current discussion about AI translation and interpreting revolves around two words: good enough. The translation industry discusses whether clients will increasingly accept output that is good enough rather than excellent. Professional debates ask how this might change expectations, prices and purchasing behaviour. Researchers examine how users perceive AI-mediated communication, what kinds of errors they tolerate, and even whether the very definition of quality may change as people become accustomed to "good enough" machine translation and interpreting.
These are important questions, and the current focus on them is understandable. While AI systems have become remarkably good at many tasks, in many settings they still do not consistently perform at the level of the best professionals. Translation and interpreting are no exception. The gap may be small in some tasks and substantial in others, but it is still visible.
I increasingly wonder whether the enormous attention given to “good enough” risks becoming a form of technological myopia.
My point is that I increasingly wonder whether the enormous attention given to “good enough” risks becoming a form of technological myopia. Richard Susskind, among others, has argued for the importance of thinking about AI not simply in terms of what today's systems can do, but in terms of the trajectories along which the technology is developing. Only by taking those trajectories into account can predict the future unfolding before us. Avoiding this technological myopia is particularly important in translation and interpreting.
The issue: much of the “good enough” debate implicitly freezes the technology at its current level. When I hear translators and interpreters — and, more worryingly, scholars and enterpreneurs — discuss the future of these activities, the underlying assumption is that human experts will continue to occupy the upper end of the quality spectrum, while AI will continue to occupy a large and expanding space below them: cheaper, faster, ubiquitous and increasingly acceptable, but ultimately inferior. This assumption risks becoming dogma. We say it, because we keep repeating it. However, current performance gaps do not establish permanent artificial inferiority, and treating it as given, ignoring the measurable improvements in AI performance across domains, and the projections of the experts, has consequences. If debates are framed this way, efforts to anticipate change, adapt to it and shape its course are misplaced, because the object is wrong. And so companies cannot prepare for what technology will do to their businesses, professionals will not focus on the added value they might still offer, and universities will fail to make the reforms they need to remain relevant.
Today's systems are a starting point, not an endpoint
If we consider the development trajectory we are witnessing today, we must first of all acknowledge that the AI translation systems we are using today are likely to be among the least capable systems we will ever use. Over the past few years, improvements in machine translation, speech recognition, language modelling, speech synthesis and, more recently, end-to-end speech translation have been substantial. Systems have become more accurate, more context-aware, more natural, faster and increasingly capable of handling aspects of communication that until recently required substantial human intervention. Those building such systems see considerable room for further improvement.
There is no guarantee that this improvement will continue indefinitely.
There is, obviously, no guarantee that this improvement will continue indefinitely. A hard technological ceiling may eventually emerge. Progress may slow. Some aspects of expert translation or interpreting may prove extraordinarily difficult to reproduce. But we do not currently know where that ceiling is. To be honest, we do not see that ceiling at all.
The more consequential question we should ask, to my mind, is not: What happens now that AI translation is good enough? It is instead: What happens if it becomes very good? Or, more uncomfortably: What happens if, in a large number of communicative situations, its performance becomes comparable or even superior to that of human experts?
A different question to ask
That shift in perspective leads to a much more fundamental discussion than what "good enough" means for us. If AI remains permanently "good enough", hence below expert performance, the future is relatively easy to conceptualise. Automation absorbs the lower and middle segments of the market (a process already underway and likely to accelerate), while human professionals increasingly concentrate on complex, high-risk, sensitive or prestigious work. This is already a familiar pattern, so less interesting to discuss. But if technological performance continues to improve, as experts in AI expect, the consequences of AI moving beyond the "ggod enough" become much less obvious.
What distinctive role will human expertise play when high-quality linguistic output itself becomes abundant?
The interesting question becomes: what distinctive role will human expertise play when high-quality linguistic output itself becomes abundant? Perhaps expertise will shift towards responsibility, verification and accountability. Perhaps linguists and language companies will increasingly become communication specialists who supervise multilingual systems, intervene in exceptional situations and design communicative processes rather than produce every utterance or sentence themselves. Perhaps new premium forms of human labour will emerge precisely because human presence carries social, institutional or symbolic value. Or perhaps the field will simply disappear or change so profoundly that it will become something completely different.
We should start considering the possibility that “good enough” will describe human output rather than machine output.
These scenarios are much harder to discuss than whether clients will tolerate an occasional mistranslation, especially because the terms of that debate may soon be reversed. We should start considering the possibility that “good enough” will describe human output rather than machine output. This is happening apready in some contexts, and has happened trhoughout the entire history of humanity. As AI becomes increasingly capable in its use of language, its application of knowledge and its expression of empathy, to name just a few, we need to consider this scenario as a plausible one.
While people might immediately jump to the conclusion that this is the begiining of the end, I have a different position. This is the beginning of an extraordinary era of aboundance. While the risks and the challenges of this transformation are real, so are the opportunities that it raises. It is app to every stakeholder, whith their distinctive roles and interests, to analyse what this changes mean, and how they can use these changes for their own interests, and for the wellbeing of society at large.
I do not have the answers for any of them, but I think it is the time to shift the discussion from what is happening today to what will happen tomorrow.