Observations

Generative AI makes it easy for us to translate writing into many different languages. It can’t help us understand how those translations come across to native speakers.

When I see texts translated into English by AI, I often wonder how the original writers feel. I imagine that some want to know how first-language English readers experience their writing.

To explore that thought, I translated one of my blog posts into Italian using ChatGPT and CoPilot. I wanted to know how I’d feel about seeing my writing in another language.

I asked ChatGPT and CoPilot to create two versions. The first had to retain my natural voice and the second had to sound “natural” to an Italian native speaker.

Here are the final paragraphs of each translation. 

CoPilot (prompt: retain my natural voice)

La consapevolezza di sé necessaria per evitare queste trappole può — e sarà — sviluppata nei contesti di lavoro. Nel frattempo, editor e altri professionisti dell’assicurazione qualità saranno lì a intercettare eventuali problemi.

CoPilot (prompt: sound natural to a native speaker)

La consapevolezza di sé necessaria per evitare queste trappole è qualcosa che può — e inevitabilmente verrà — sviluppato nei contesti di lavoro. Nel frattempo, editor e professionisti del controllo qualità svolgeranno un ruolo fondamentale nel intercettare eventuali problemi.

ChatGPT (prompt: retain my natural voice)

La consapevolezza di sé necessaria per evitare queste trappole può e verrà sviluppata nei luoghi di lavoro. Nel frattempo, editor e altri professionisti della qualità saranno presenti per intercettare eventuali problemi.

ChatGPT (prompt: sound natural to a native speaker)

La consapevolezza necessaria per evitare queste trappole si svilupperà — e in parte si sta già sviluppando — nei contesti lavorativi. Nel frattempo, editor e altri professionisti della qualità continueranno a vigilare, pronti a intercettare eventuali criticità.

My Italian is intermediate level, so I can’t tell whether any of these texts sound natural to a native speaker. I asked ChatGPT and CoPilot for an evaluation.

They both agreed that ChatGPT’s second translation sounded the most natural. CoPilot said it was more nuanced, made better use of idioms, and flowed better. ChatGPT said it sounded smoother, more polished, more idiomatic, and not as stiff as the other translations.

The problem here is that smoothness, polish, flow, and stiffness are not objective measures. For that matter, “natural” isn’t either. Who knows how they’re defined by AI?

The whole thing left me scratching my head. If I had to pick one of those paragraphs for real-world use, I’d choose the one preferred by ChatGPT and CoPilot. But I’d also wonder how much I could rely on their judgement, as it is built on their mysterious internal workings.

I also can’t tell if the two other translations retain my natural voice. I turned to CoPilot for advice. It said that ChatGPT’s translation sounded like the voice of an institutional authority. It added:

“A formal, institutional, and forward-looking voice that communicates ideas with measured authority, emphasizing systems, professional oversight, and gradual improvement.”

CoPilot said its own translation sounded more reflective and that it provided professional reassurance.

“A reflective, professionally grounded voice that combines analytical clarity with a reassuring, process-oriented perspective, using slightly elevated but controlled language to convey authority and calm confidence.”

I don’t think the original English blog post sounds institutional. Clearly, something has appeared in ChatGPT’s Italian translation that isn’t in the English.

I carried out this experiment because I wanted to explore how writers feel when they see their work translated into English by AI. I imagine that many end up feeling very frustrated. 

My key frustration points in this experiment were that I could not stress test or verify the AI outputs outside their own systems.

If this had been a real-world situation, my best solution would’ve been to ask an Italian human being!

You can read the original English blog post here.

People often talk about the skills needed to use AI successfully at work. One of these skills is learning how to write clear prompts that steer the AI in the right direction. Another is learning how to evaluate and adjust its output.

Editors are part of that evaluation and adjustment layer. We check AI-text in various ways, ensuring that it makes sense and does what it’s supposed to.

There’s another AI skill that is a lot harder to learn quickly because it’s about emotion and psychology. It’s a kind of self-management.

Users of AI must know when to trust themselves and their own knowledge. They also need to recognise the limits of their own abilities.

In the worst case, a person with imposter syndrome could defer too much to AI, allowing it to overwhelm them with the wrong advice. 

Similarly, someone who is too confident in themselves could fail to take the careful steps needed to evaluate what the AI has produced.

The self-knowledge required to avoid these traps can and will be developed across workplaces. Editors and other quality assurance professionals will be on hand to catch any problems that develop in the meantime. 

Generative AI is a writing tool, not a replacement for humans. People use it on a spectrum, with 100% AI wording in the final draft at one end and 0% at the other. For some, generative AI is just a sounding board. For others, it’s a text‑output machine steered by effective prompting.

These varied approaches have an impact on how writing comes across to readers. A 0% AI-wording approach is most likely to retain the writer’s voice. In other words, it will sound like an individual wrote it. With 100% AI wording, it’s more probable that the writing will feel “clean” and “standardised.”

For organisations that need to communicate with their stakeholders, the 0% and 100% approaches have different situation‑dependent advantages.

Quality‑assured AI‑generated text can be clear, well structured, and easy to read.

Writing with a human voice often sounds more authentically real. This is particularly important when the text is supposed to have been written by an individual known to the stakeholders.

The two approaches also have disadvantages.

Unedited AI‑generated wording isn’t always a good cultural fit for a particular group of readers. Anything could go wrong; for example, it could sound too businesslike, too energetic, or simply not appropriate for the community.

Purely human writing can be difficult to read when the writer’s thought habits get in the way. Common problems include being too tentative, over‑explaining, and changing ideas part way through.

Human editors can help in both situations. We often adjust the voice of AI‑generated writing so that it’s a better fit for readers. We also make sensitive structural edits to human writing so that its original voice is retained but ideas are communicated more clearly.

Of course, using AI to write doesn’t mean a binary choice between 0% or 100%. People move up and down the AI writing spectrum, using its wording more or less heavily depending on their needs and what they want to say.