INDUSTRY DEVELOPMENTS
There’s a lot being said about ChatGPT, and AUSIT Senior Member Sam Berner has – not surprisingly – found it hard to condense all she’s learned about the ‘chatbot’ into a few pages of In Touch – but she’s made a good start, and this will doubtless not be the last article we publish on this trending topic.
ChatGPT will … teach your kids … walk the dog … [and] eventually replace humans on earth.
There is a wealth of information about this new world that needs to be discussed and pondered upon, so with my very foolish optimism I set out to research the topic and write this article … only to discover it to be a much harder task than I’d predicted. I explored both the free version, ChatGPT-3, and the paid ChatGPT-4 extensively; attended AI (artificial intelligence) webinars; read articles and books on the chatbot’s technology and ethics; and interviewed GenZ users applying ChatGPT in various fields.
As I type my final draft, OpenAI (the company that created ChatGPT) has just announced a new version of the software,[1] one predicted to have an even bigger impact on language services (including T&I) – so by the time you read this article, parts of it may already be outdated.
What is CHATGPT?
It could be assumed that writing objectively about a scientifically constrained subject is quite straightforward – unfortunately, this is not always the case. Whenever a lot of money and a potential watershed in everyday activity coincide, hype is generated. AI in general has been on the receiving end of dystopic hype for decades, from Lem’s 1961 novel Solaris to the 1999 and 2004 films The Matrix and I, Robot. With the advent of ChatGPT, the hype has become more polarised: marketeers, tech-evangelists and software geeks on one end, and doomsayers from all fields on the other. ChatGPT will set your dinner menu, tell you how best to work, write your essays for you, teach your kids, and walk the dog. It will also steal your job, further embolden power grabbers, speed up climate change, and eventually replicate in unpredictable ways and replace humans on earth.
The actual facts are slightly more nuanced. It’s true that ChatGPT can streamline workflow, produce excellent graphics for your presentation, and tailor a recipe to what you have in the fridge (within the obvious limits). It’s also true that it’s causing major environmental damage (one of the ‘most significant concerns’ being ‘the amount of energy required to train and operate AI algorithms’)[2] and that it will eventually replace many jobs.[3] However, each of these statements can be qualified further. So let us first explain what ChatGPT is, and how it functions. Let me quote ChatGPT introducing itself:
Imagine your brain is like a super library filled with billions of books. When someone asks you a question, you quickly skim through these books and come up with an answer. ChatGPT-4 is like a digital version of this. It’s been trained on a huge amount of text, like reading billions of books. When you ask it a question, it quickly checks its ‘knowledge’ and gives you an answer, trying its best to sound like a human. In essence, it’s a big computer program that’s really good at understanding and generating text based on what it’s learned!
To be able to perform this tailored knowledge transfer, ChatGPT was trained on a mixture of licensed and publicly available data, as well as data created by human trainers. This data wasn’t just randomly scraped from the internet without any filtering. The aim was to train the model on a diverse range of information, to make it as knowledgeable and general purpose as possible, but there were certainly considerations about data quality and representation. The data was multilingual, so ChatGPT can understand and generate text in a variety of languages, and it can handle conversations in many of the world’s major languages. It can perform translations between multiple languages, however by its own admission it is:
… more of a generalist in many tasks, including translation, rather than a specialist in any single one. If you need a quick translation or an understanding of text in another language, it can help, but for critical or nuanced translations, dedicated translation services might be more reliable.
– proof that it wasn’t created to take away our jobs and destroy our profession (assuming we’re included in ‘dedicated translation services’).
How Can ChatGPT help translators?
- When compared to neural machine translation (NMT) applications – designed to translate text as accurately as possible, but often falling short on cultural nuances or idiomatic expressions – ChatGPT can sometimes provide more context or nuance in the target text, because it’s been trained on diverse content.
- It also allows users to ask questions about the translated text and request clarifications.
- Because it has a ‘randomness’ factor built in to make it more ‘human like’, users can ask it to retranslate the text in various ways.
- It can paraphrase, gist, shorten the text, ensure inclusive language, find resources (paid version with access to the internet only), fix grammar and syntax, and provide synonyms or related vocabulary, as well as information about cultural nuances, idioms, or historical contexts that might be relevant to the translation. ChatGPT might be aware of new slang, idioms, or language trends that a translator hasn’t yet encountered.
- The paid version with internet access is quite good for researching the subject matter of a text.
Therefore, while ChatGPT is a powerful tool, human intuition, cultural understanding, and expertise remain crucial in translation work. ChatGPT, in the words of a presenter at one of the webinars I attended, is like ‘a young, smart and eager assistant that doesn’t have the practical experience’.
ChatGPT’s flaws
- For technical reasons the chatbot occasionally ‘hallucinates’ – that is, it provides factually incorrect answers, including creating references that are almost always nonexistent when it isn’t accessing the internet.[4] This is the main flaw.
- It struggles with texts that are context-heavy, nuanced, or specialised – for example, in medical, legal or literary fields.
- The current (as I write) ChatGPT training dataset was compiled in 2021, so it is unaware of anything beyond then.
What It Takes
To be able to utilise ChatGPT for your optimal benefit in translation, you will need to:
- be a good translator without any tools – this is the baseline – and learn how to use the chatbot as a complementary tool (as with any AI) – this involves understanding when to trust the bot’s translation, and when to rely on your own professional expertise
- have some basic technical proficiency which, with ChatGPT, is mainly about writing excellent ‘prompts’ (the questions/directions you type into ChatGPT in order to obtain an answer from it, or have it perform a task), as the style of your prompts may affect the quality of the resulting translation outputs[5] – this isn’t hard to learn, and there’s a lot of help online
- be open to learning, change and continual evolution
- understand that while ChatGPT ‘performs competitively with commercial translation products on high-resource European languages’, it ‘lags behind significantly on low-resource or distant languages’[6] – and know the way around this: have the text translated into the low-resource language via a high-resource language first and then into the low-resource language (ChatGPT can do this simultaneously)
- (if you are serious about utilisation) invest in the ChatGPT-4 version, and in a few plugins that assist its performance – even with this paid version, the bot sometimes times out, and has a limit on the number of interactions per hour
- be very clear about your professional ethical standards, because you will need to recognise the ethical implications of AI translations. Our AUSIT Code of Conduct needs to be updated to cover the new tools.
This last point warrants expansion. Here are a few ethical issues to look out for:
- As ChatGPT learns from vast amounts of data, it might perpetuate or amplify any bias contained in that data. Be prepared to identify and address any biases in output.
- It is, needless to say, unethical to misrepresent ChatGPT or NMT translations as human, so be transparent.
- Be mindful of confidentiality and privacy concerns. ChatGPT learns from every line you type, and that data becomes part of what stays on its servers.
- Watch for the occasional inappropriate or offensive ‘hallucination’ – you are ultimately responsible for the final product, so maintain due diligence and quality assurance.
- Ensure that you are using the technology responsibly and in a way that aligns with the broader values and needs of your audience.
What Does the Future Portend?
AI is reshaping translation, offering both opportunities and challenges. While it won’t replace human translators yet, those who effectively harness tools such as ChatGPT to handle routine tasks can focus on cultural nuances and intricate elements of translation. This synergy boosts their efficiency and accuracy and accelerates outcomes, especially for bulky projects and/or tight timelines. AI thus becomes an essential tool for translators, heightening their marketability.
However, collaboration with ChatGPT presents ethical dilemmas, including job displacement, bias, and data privacy concerns. The line between human and AI-generated text is blurring, affecting professions such as writing, translation, and even engineering and medicine. Overreliance on AI might standardise language, but on the other hand ChatGPT is being utilised in the preservation of endangered languages.
Concerns arise about intellectual property, data access and age verification. Despite OpenAI’s CEO emphasising AI’s risks, legal battles have emerged.[7] And while training AI is costly, some firms are threatening ChatGPT’s viability by profiting from its API (application programming interface) without compensating OpenAI. What the future holds for ChatGPT is anyone’s guess. It is not the only such application, nor the last. AI, in short, is here to stay.
Sam Berner has four decades of experience working as a legal translator both in Australia and overseas. Prior to settling in Australia, she worked for UNICEF and UNHCR. She is a past AUSIT National President, and has served on various AUSIT committees since 2003. Sam’s research interests lie in intersectionality of ethics, technology, language politics and translation. Earlier this year, she was one of the first AUSIT members to be granted senior membership.
[1] Mehta T (2023, June 25). GPT-5: Everything We Know So Far About OpenAI’s Next Chat-GPT Release. SlashGear.
[2] Li R (2023, May 8). The Environmental Impact of AI. Global Research and Consulting Group Insights.
[3] Blake A (2023, March 30). ChatGPT could threaten 300 million jobs around the world. Digital Trends.
[4] Marr B (2023, March 22). ChatGPT: What Are Hallucinations And Why Are They A Problem For AI Systems. bernardmarr.com
[5] Jiao W, Wang W, Huang J, Wang X & Tu Z (2023). Is ChatGPT a Good Translator? Yes With GPT-4 As The Engine (Version 3). arXiv.
[6] As above.
[7] Hines K (2023, April 11). ChatGPT And Generative AI Tools Face Legal Woes Worldwide. Search Engine Journal.
