Dina Hrecak

For months now, there seems to be a new, somewhat disconcerting trend among my fellow translators, editors, and writers.

More and more, privately and on social media, I see them defending themselves against accusations they never actually expected to have to answer. They explain that they do not use AI to write. Or that they use it only as a sounding board. Or that they use it to check grammar, but not to generate text. Some are even beginning to defend perfectly ordinary features of good writing: an em dash, a particular sentence structure, correct punctuation, the use of articles in English.

The underlying anxiety seems to be the same: what if someone thinks I used AI?

This strikes me as a strange place for people who have spent years, sometimes decades, learning how to work with language to find themselves. We have been practising a craft, polishing it, making mistakes, learning from them, reading, studying, translating, editing, writing, and rewriting. And now some of us seem to feel that we have to prove that the result of all that work is still ours.

Part of the problem is the emergence of AI-writing detectors. They promise to tell us whether a text was produced by a human or a machine, but the research on their reliability is far from reassuring. Some widely used detectors produce substantial false positives, particularly for non-native English writing. More recent research continues to find serious problems with false positives and significant differences between detectors.

So we have arrived at the peculiar situation in which people who know how to write well are worrying that writing well might itself become evidence against them.

But what if we are asking the wrong question?

Instead of endlessly arguing over whether someone used AI, perhaps we should be asking a more useful question: now that these tools are here, what are we going to do with them?

There is an old story about a factory owner whose expensive machine broke down, bringing the entire assembly line to a halt. He called a repairman, who arrived, walked around the machine once, then twice, and announced that he knew how to fix it. The price would be $10,000.

The owner immediately agreed. The machine produced far more than $10,000 in a single day, and with the entire assembly line at a standstill, the price was worth paying.

The repairman took a hammer from his toolbox, stood beside the machine and tapped it once.

The machine started working. The assembly line moved again.

The owner stared at him. ‘What? That’s it? One swing of the hammer costs $10,000?’ 

‘No,’ said the repairman. ‘The swing of the hammer costs $10. Knowing where to hit, when to hit it, and with how much force costs $9,990.’

The story is almost certainly apocryphal, but I have always liked it because it captures something we tend to forget about skilled work: the visible action is not necessarily where the value lies.

And this is where I think the AI conversation becomes much more interesting.

Yes, AI can write. It can translate. It can edit. It can produce marketing copy, summarise documents, suggest alternatives, offer explanations and do many of these things much faster than a human. Pretending otherwise is pointless.

But AI is a tool.

A complex one, admittedly. Still, it requires someone to decide what to do with it.

Anyone can pick up a hammer.

Knowing where to hit is another matter.

We have been here before

I began my career as a translator, at a time when computer-assisted translation tools and machine translation were already part of the professional landscape but had certainly not earned the degree of trust they have today.

The early history of machine translation is a history of disappointment as much as technological progress. The first rule-based systems appeared in the 1950s, followed decades later by statistical and then neural approaches, with each generation improving on the last. Meanwhile, CAT tools and translation memories became more common, helping translators store previous work, maintain terminology and keep large projects consistent.

CAT – computer-assisted translation – is not the same thing as machine translation. A CAT tool is essentially a working environment for translators, combining features such as translation memories, terminology databases, and quality checks. Machine translation can be one of the tools within that environment, rather than the environment itself.

I remember working with these tools and, like many translators, learning what they were good for and what they were not.

Suppose you are translating documentation for a manufacturer of home appliances. The company launches a hundred new products a year. Each product comes with an instruction manual, and each manual contains pages of recurring language: safety warnings, operating instructions, function descriptions, maintenance procedures. The same phrases appear again and again.

This is precisely the kind of work in which translation technology can be extraordinarily useful.

If a particular technical term has been translated in a particular way, a translation memory can help you keep using it consistently. If the same instruction appears in fifty manuals, you do not need to reinvent it fifty times. The technology can save time, reduce repetition and help maintain consistency across a body of work far too large for a human translator to keep entirely in their head.

There is nothing shameful about that.

Quite the opposite: it is a good use of technology.

But now give the machine something else.

Give it Edgar Allan Poe. Or Dostoyevsky. 

Literary translation is a completely different beast. The fact that a word appears twice in a story does not mean it should necessarily be translated the same way twice. A choice that is technically equivalent in one sentence may destroy the rhythm, ambiguity, tone or emotional effect of another. A literary translator is not merely transferring information from one language into another. They are trying to recreate an experience.

The takeaway is not that machine translation is ‘good’ for manuals and ‘bad’ for literature. It is more subtle than that. Different kinds of language present different kinds of problems.

CAT tools were initially met with scepticism, but they have since become important components of professional translators’ and editors’ toolkits. Over time, the question stopped being whether technology was good or bad and became: what kind of technology was useful for what kind of work?

Translators did not disappear, as feared in the early days of MT. The profession survived – but the work changed. That is the part of this history I find most relevant to the current AI debate.

That distinction matters enormously.

The tool is only as good as the job you give it

There is no doubt – technology saves time.

It does not remove the need to make decisions.

And this is where I think the old argument about machine translation can teach us something about AI.

Having the machine do some part of the work is no ‘easy way out’, no ‘sit back while the computer translates’. The professional translator still had to be there. The work just became more nuanced.

You have to recognise when the machine has produced something useful. You have to recognise when it has produced something technically plausible but wrong. You have to decide whether a sentence needs correcting or rewriting. You have to know when the machine’s suggestion should be accepted, when it should be changed and when it should be ignored completely.

The machine gives you an output.

Your expertise tells you what to do with it.

This is also where tools such as Grammarly come in.

I find Grammarly useful. I use it as a final ‘idiot-check’: grammar, punctuation, consistency, perhaps British versus American English. It can catch things I have simply stopped seeing because I have stared at the same paragraph for too long.

But I do not regard its suggestions as instructions.

Quite often, I disagree with them.

Sometimes it suggests changing something that is perfectly correct because it would be ‘clearer’ or ‘better’ according to whatever linguistic preferences the system is applying. And sometimes that suggestion would make the sentence more generic. It would remove a little piece of its character.

There is an important distinction here: correct does not always mean better.

A grammatically impeccable sentence can still be dull. A deliberately unusual construction can be exactly right for the writer’s purpose. A sentence can be slightly awkward and still convey something that a smoother sentence would lose.

Grammar matters; understanding why a rule exists matters even more when deciding whether to follow it.

And then came generative AI

Generative AI changes the scale of the problem.

Machine translation primarily gave raw word-for-word swaps. CAT tools gave us memories, terminology, and assistance. Grammarly gives us corrections and polish.

Generative AI can do almost all of these things, while also producing new text from scratch.

It can draft an article.

It can rewrite the article.

It can suggest ten headlines.

It can translate the article.

It can explain why a sentence might not work.

It can propose a different structure.

It can tell you your argument has a hole in it.

It can then confidently invent something to fill that hole.

That last part is important.

The better these systems become, the more tempting it is to forget that their fluency is not the same thing as reliability. A beautifully written sentence can still be factually wrong. A plausible translation can still miss the meaning. A confident editorial suggestion can still be unnecessary. The machine can give you an answer before you have even decided what question you actually wanted to ask, or if you want to ask anything at all.

And this is where the role of the human professional becomes indispensable.

Refusing to touch the tool, or handing everything to it, is a childish debate. Usually, the useful position is somewhere in between. The new expertise, I suspect, will involve learning how to direct, interrogate, and evaluate the machine.

That means knowing what information it needs. Knowing how much context to give it. Knowing what sort of task is appropriate to delegate. Knowing when its answer is likely to be unreliable. Knowing how to compare alternatives. Knowing what must still be checked independently. Knowing when a stylistic improvement is actually a stylistic deterioration.

And, perhaps most importantly, knowing what not to give it.

I say ‘I suspect’ deliberately.

Researchers are already studying questions such as cognitive offloading, critical thinking, and the effects of AI-assisted work on how people engage with difficult tasks. The early research is interesting but far from providing a simple verdict. Some findings suggest that unstructured reliance on AI may encourage people to outsource parts of their thinking; other work suggests that deliberately structured use of AI can support critical and creative thinking. The research is young, and it’s too early to guess what the final answer will be. Or, frankly, if there will be one final answer.

But I do think language professionals should start paying attention. We need to ask ourselves a rather uncomfortable question: If a machine becomes really good at doing the things I have spent twenty years learning to do, what exactly is it that I still need to practise?

I don’t have a complete answer.

I don’t think anyone does now.

So what are we actually afraid of?

Perhaps this is why the anxiety around AI detection hits too close to home.

We are spending enormous energy trying to prove that competent human work is human, defending the fact that some of us simply know how to write. But good writing was never evidence of AI use.

Good writing is evidence that somebody learned how to write.

The rules of grammar, punctuation, and structure are not suddenly suspicious because a language model has learned them too. Of course it has. Language models were trained on human language. They have absorbed patterns that generations of writers, editors, translators, teachers, and readers have collectively produced. We have all learned from the same textbooks, both human and AI.

A pianist does not stop playing scales because a computer can reproduce the correct notes.

A carpenter does not stop knowing how wood behaves because power tools exist.

And a writer does not become less of a writer because a machine can produce a grammatically correct sentence.

The danger, as I see it, lies somewhere else. It lies in allowing the convenience of the machine to become a substitute for developing our own judgement.

If I ask AI to give me ten possible headlines, compare them, reject nine and choose one because I understand the audience and purpose of the article, I have used a tool. But, if I ask AI to write the headline because I cannot be bothered to think of one, then accept whatever it gives me because it sounds plausible, something else has happened.

The output might be identical.

The cognitive process is not.

And perhaps that distinction will become ever more important as the tools become better.

The $9,990 question

Which brings me back to the repairman.

The hammer swing costs $10.

The knowledge of where to hit, when to hit and with how much force costs $9,990.

For years, writers, translators, and editors have accumulated their own $9,990: vocabulary, grammar, rhythm, cultural knowledge, subject knowledge, pattern recognition, instinct, experience, judgement, the ability to hear when a sentence is wrong even when we cannot immediately explain why.

AI does not make any of that worthless. But it may change how we use it.

Perhaps some of the things that once took us hours will take minutes. Perhaps some of the things we once regarded as core professional skills will become less important. Perhaps entirely new skills will emerge. Perhaps some of the things we currently think are uniquely human will eventually become things machines can do surprisingly well.

No one can say for certain.

So I would rather we stopped building opposing camps of people who ‘use AI’ and people who ‘don’t use AI’. I would rather we stopped treating the use of a tool as a moral judgement on the person using it.

We have been through technological change before.

Machine translation did not destroy translation. CAT tools did not substitute translators. Automated grammar checking did not leave writers and editors out of jobs. They changed the work. AI will change it too.

We cannot keep the machine out. It is already here. What matters now is whether we can learn to use it without allowing it to flatten our voices, replace our curiosity or do the thinking we still need to do ourselves.

And that leaves me with the question I keep coming back to.

If AI is the hammer, and the swing costs $10, how do we build, once again, the $9,990 worth of expertise?

That, to me, is the conversation we should be having.