Implementing an AI Translation Workflow: A Practical Guide to ChatGPT Success

Cathy Castling
Sep 11
9 min read

After about a year of experimenting with ChatGPT for translation, I discovered something unexpected – asking it to "rewrite" text produces dramatically better results than asking it to "translate". This small change transformed our French translations from robotic failures to human-approved content that needed little to no editing.

ChatGPT isn’t a translation too and, honestly, it’s not even really a writing tool. Yet when I was tasked at work with figuring out how we could implement it into our translation workflow, I knew I had to find a way to make it work. The challenge wasn’t just getting decent translations; it was producing content good enough to publish across multiple languages without embarrassing the brand.

There’s plenty of generic advice online about using AI and ChatGPT, but almost nothing dives into the specifics of translation. This post shares some insights from our months-long experiment in making ChatGPT produce translations that don’t make native speakers cringe.

TL;DR After 6 months of testing ChatGPT for professional translation:

Use custom GPTs (like TranslateExpert) instead of basic ChatGPT
Ask it to "rewrite" instead of "translate" for more natural-sounding results
Provide maximum context—translate entire newsletters, not single sentences
Always have native speakers review output before publishing
Focus on iterative feedback loops with linguists to continuously improve results
ChatGPT works best as a first-draft tool, not a replacement for human translators

Can ChatGPT translate?

Yes… and no.

Machine translation is the process in which computer software translates a text from one language into another without human involvement, tools like DeepL. ChatGPT is a generative artificial intelligence chatbot based on large language models (LLMs), so is essentially a chatbot that generates new output based on its existing dataset.

If we use part of the definition of machine translation above, "software that translates a text from one language into another", then yes, ChatGPT can translate. But how well can it translate? Can it replace human translators within the translation workflow? At the moment, I’d say no, as ChatGPT still struggles with the basics – like counting – and really has issues with producing translations that are fit for purpose.

Problems with ChatGPT when translating

Over the last year or so, I’ve learned a lot about ChatGPT, prompting, and how to refine it so that the output is decent. I still have a couple of issues with it, including:

Inconsistent output: ChatGPT produces inconsistent translations which means that you (or whoever is checking the output) need to be extra careful when checking the text as small errors and inconsistencies could be introduced into your translation and published on your website without anyone noticing – except your customers. One of the biggest issues I’ve experienced is brand names when they are localized for foreign markets, ChatGPT doesn’t "remember" the localized names and always uses the German brand name, no matter how many times I "remind" it.
Awkward output that lacks nuance: ChatGPT produces translations that are awkwardly phrased and generally lack cultural nuance. From applying English grammatical rules to Finnish to using literal translations of words and phrases that native speakers would never use, there are a whole host of weird things that could end up in your translation. Your Finnish customers will spot errors and weird phrasing a mile away and may, as a result, not buy anything from your website if there is raw (i.e. unedited) translated output on there.
Short memory: ChatGPT’s memory – in my experience – is not good. There is a limit of how many tokens it can remember and once you reach that limit, it forgets all of the instructions and translations that it’s done prior to this reset of its tokens. This means that you have to spend a lot of time training it, feeding it with examples and instructions, for it to produce anything that’s even remotely close to the quality of the output that you’d want to publish on your website.

Translation output created by machines and software always has issues and, generally speaking, should always be checked by a human to ensure it’s fit for purpose. But I wanted to share the differences between issues that crop up with tools like DeepL and in ChatGPT.

DeepL is a well-known and widely-used automatic translation tool and claims to be the best on the market. You paste your text into the box on the screen and DeepL produces a translation based on a general database of texts, if you’re using the free version. Issues with DeepL output include mistranslations, awkward phrasing, and gender bias (e.g. always using the generic masculine), to name a few.

ChatGPT throws up more nuanced issues. In my experience, it has issues with sentence structure and grammatical rules in different languages. It has a habit of applying the source sentence structure (in my case, English or German) to the target language (e.g. Finnish, Spanish, Swedish, etc.). It does the same with punctuation and capitalisation rules (we had lots of issues with Title Case being applied to languages that don’t use it). It also makes up words. You probably knew it makes up facts and statistics, but it makes up words too (we had particular issues with the translation of the word “fashionista” as many of the languages we work with just use the English word but ChatGPT came up with some creative but non-existent words for it in Swedish and Finnish).

Both tools can be used as part of the translation workflow but I’d always recommend that you have a human review the texts so that they can pick up on the issues mentioned above and you can use the feedback to improve the output.

The challenges of using AI for professional translation

There’s been chat on the internet for years that AI, machine translation, and robots will replace human translators, which isn’t true. Human translators should adapt the way they work and embrace the right AI tools that support them in the work, not make their lives harder.

Personally, I’ve had mixed feelings about using ChatGPT and other AI tools for translation. There are better tools on the market that have been developed specifically for translation and translators that can help users manage their terminology and are much easier to use. These tools require big time and monetary investments; licenses are expensive and users have to invest time in learning how to build and manage glossaries while getting to grips with the software ecosystem.

I use Trados Studio 2019 in my freelance translation work, which is a Computer Aided Translation (CAT) tool. It doesn’t automatically translate texts for you but uses glossaries that I’ve built myself over the last five or so years. I’ve invested a lot of time and energy into building glossaries for my clients but it pays off massively when translating repetitive texts, like product descriptions. My CAT-tool-aided workflow uses technology and means that I create high-quality translations that are free of errors and fit in with a brand’s tone of voice and style.

Challenge: Improving ChatGPT’s output for more natural-sounding translations

I was asked to improve the translated output we were getting from ChatGPT. We started off using basic prompts with GPT-4, like:

“The tone of voice in the translations should be friendly, warm, yet informative”
“Use these one-to-one translations in all translations [list of words]”

The results were inconsistent, robotic, and often completely wrong. We constantly fed feedback from our team of linguists into ChatGPT but we couldn’t fix the weird phrasing in any language.

Through my research, I came across the idea of using custom GPTs to create output that better aligned with our brand. I didn’t have the time to learn how to build my own custom GPT from scratch, so I decided to use TranslateExpert GPT, a pre-existing tool recommend by another linguist. The key feature that appealed to me was that TranslateExpert is able to mimic the style and tone of a specific webpage using just the URL for that webpage and align the output with the brand’s voice.

Designing the workflow

The workflow that we’re currently using is a result of trial and error and constant feedback for the GPT. I’ll detail our current workflow below.

Train the GPT in separate chats using longer texts to teach it our tone of voice in every language. We use one chat per language, so that all of our Swedish translations, for example, are done in one chat and stored in one place.
Provide the GPT with an existing piece of content and clear brand-specific instructions, e.g. capitalise “Du” in German, use the formal “you” in French, etc., so that the output aligns with the existing brand.
Provide the GPT with as much context as possible when translating. Words and sentences don’t exist in a vacuum, so feeding the machine as much context as possible was key, e.g. feed an entire newsletter in for translation not just single sentences or phrases.
Give the translations to our human linguists and ask them for specific feedback on errors, inconsistencies, and general issues.
Feed these notes back into the GPT to fine-tune future results.

Feedback loop: Refining translations with native speakers

Throughout the process, we focused a lot on working with our in-house native speakers and getting their feedback on all translations. We used their feedback to improve things like awkward sentence structures, which helped us develop clearer prompts. We found that a prompt like "Rewrite" produced more natural-sounding translations that "Translate" prompts.

The first noticeable improvements came after we had switched from GPT-4 to TranslateExpert. We still had issues to deal with but generally the translations were getting better. The issue of the GPT following the source text (German) sentence structure in the target texts (all other languages we translate into) helped us adjust and fine-tune our prompts so that the output continued to improve.

Working with a GPT is an iterative process, you need to provide the GPT with regular feedback so that you can refine and improve your translated output. We saw that we were able to fine-tune our prompts to get closer to the results we wanted with every new translation and round of feedback.

Breakthrough: “Good” translations in French and Swedish

A few weeks into the process, we got the feedback that the latest round of French and Swedish translations were good and didn’t need any editing by our human linguists.

Proofreading Swedish texts: I told the GPT to proofread the Swedish texts. I asked it to review the same text multiple times and tell me each time what had been changed and why. The final version that was reviewed by a human had been through 6 round of checking, editing, comparing the text with existing texts for TOV. The whole process took me about 20 minutes to review 100 words. This amount of reviewing highlighted where the source text (German) wasn’t particularly clear, caused confusion in Swedish, and resulted in a weird-sounding translation. And, as a result, we reviewed the German text and all of the other translations to ensure the same issue hadn’t been missed in other languages.

Feedback on French translations: In the same week, we received feedback from our French team that the translations were spot on and didn’t need any changes. These are the steps we followed to get this result:

We prepped the GPT with our brand guidelines and preferred keywords immediately before feeding in the text.
We gave the machine a webpage URL as a reminder of our tone of voice in French.
We gave it specific instructions for French, e.g. use of formal “you”, localized brand name, etc.
We fed previously corrected texts into the GPT and asked it to apply these changes to future translations.
We were working within a character limit, so we asked the GPT to “Rewrite the text”.
We fed the text in to be translated and shared the output with our linguist for feedback.

I think step number 5 (rewriting the text) is what really pushed the GPT to create translations that sounded natural in French.

Lessons learned: The importance of the human element

The work I’ve done with GPTs has confirmed what I already knew – humans need to be involved in the translation process. We got some decent results and good feedback but I wouldn’t publish output from the GPT without having it checked by a linguist.

Working with the GPT is an ongoing and ever-evolving process. I’m pleased with the progress we’ve made (from terrible to decent translations) but recognise that I’ll have to continue giving the GPT feedback so that it maintains the quality of the translations. It’s a process that will keep going as long as we’re using the GPT to translate.

Final thoughts

One year ago, ChatGPT was producing translations so awkward that our Finnish customers would have cringed. Today, our French and Swedish translations regularly pass human review without edits. This transformation didn’t happen because ChatGPT suddenly got better – it happened because we learned to work with its strengths rather than its limitations.

The journey taught me that ChatGPT definitely isn’t replacing human translators; it’s becoming a sophisticated first-draft tool that amplifies human expertise rather than replacing it. The most successful approach combines AI efficiency with the cultural nuance and linguistic precision that only humans can provide.

My key recommendations for anyone considering this path:

Use custom GPTs to align output with your brand voice and existing content. Generic ChatGPT simply won’t cut it for professional-quality content and translations.
Invest in human linguists for reviewing your most critical content. AI can handle the heavy lifting, but human expertise ensures quality and cultural appropriateness.
Embrace the iterative process. Working with AI translation requires ongoing refinement and feedback loops. It’s not a “set it and forget it” solution.

As we come to the final stretch of 2025, predictions suggest that consumers are increasingly craving authentic, human-crafted content. The most impactful translation workflows will be those that harness AI’s speed and consistency while preserving the human touch that creates genuine connections with audiences.

ChatGPT is evolving, and while it’s a long way off from winning any translation awards, it’s becoming a valuable tool in the right hands. I’m excited to see what purpose-built translation AI tools emerge next – and how human translators will continue to adapt and thrive alongside them.

I’m Cathy, language-nerd and expert in making German brands shine in the English market. By choosing the right keywords and optimising your content, I can help your products reach more potential customers and increase your sales.

Ready to see your brand’s visibility soar? Get in touch today for a chat and let’s make your content work harder for you!