Crikey, ChatGPT's gone bush! How AI is learning the art of Aussie slang
Ever tried to explain why a sausage would be referred to as a "snag" while overseas, or why the toilet is the "dunny"? If you found this challenging, spare a thought for large language models (LLMs) such as ChatGPT, which have to contend with slang terms from all over the world.
Is it possible for AI to decipher the strange "code" that is Australian slang, given all the nuance and loaded into it?
Cracking the code
LLMs don't "understand" language like we do. Rather, they are trained on massive quantities of online text data (including websites, news articles and books) to learn patterns between words. They can then mimic these patterns to produce human-like responses.
So it follows that unless AI systems can mingle with people in informal real-world settings—or can access TV shows such as Kath and Kim—they're unlikely to grasp the finer points of our real-world conversations.
Take words such as "cooked" and "random," which can have different meanings in different contexts. Or consider the phrase "flat out like a lizard drinking." What could it mean? Is the speaker comparing themselves to a thirsty reptile sprawled out under a dripping tap?
The phrase actually refers to being very busy, using the visual metaphor of a lizard's fast-moving tongue. While an AI may not make this connection, many people living in Australia will have a lifetime of experience that helps them understand the message being conveyed.
The art of understanding the Aussie lingo… @ballcaptravel
Get free science updates with Science X Daily and Weekly Newsletters — to customize your preferences!
To further complicate matters, Aussie slang continues to evolve, and doesn't always follow the rules of grammar and structure.
Slang phrases tend to follow a and are often filled with idioms, metaphors, abbreviations and culturally-specific humor. Australian language expert Roland Sussex estimates we use more than .
Slang also changes from one generation to the next. For instance, one older Australians are more likely to shorten words with an "ie" or "o" sound, such as "truckie" instead of "truck driver" and "ambo" instead of ambulance. Young Australians, meanwhile, are more likely to clip words or add an "s," such as "mobes" for mobile phone.
Are we there yet?
Can AI chatbots learn Aussie slang? There is evidence many are already developing a broad understanding of the most frequently used terms and their current interpretations.
For example, "give it a crack" and "mozzie" are both understood by .
In 2021, Alexa partnered with local celebrity and comedy duo The Inspired Unemployed to incorporate a large collection of Aussie slang into its vocabulary. The personal AI assistant even comes with an feature.
Keeping up-to-date with changing Aussie slang terms, interpretations and regional dialects is a resource-intensive undertaking. Nonetheless, ChatGPT and other LLMs have made progress on this front, as this example shows:
Some chatbots, such as , can scour the internet in real-time to try and find the best possible response to an input.
Trying to peek inside
LLMs continue to advance in their sophistication and capabilities. The most recent models such as GPT 4o, DeepSeek and Claude 3.7 even incorporate "thinking" to tackle more complex tasks by displaying an before revealing their answer.
However, has shown many AI models, when prompted, won't always reveal the full "chain-of-thought" they followed to arrive at a particular answer.
This makes it harder for us to understand the models' intentions and reasoning processes. So while they may be learning to adapt and respond to our niche slang and cultural references, in many ways they remain a .
Beyond that, AI models can only regurgitate our own slang back to us. They can't grasp why it is meaningful. Nor do they understand the important role slang .
Aussie slang is born out of millions of interactions and conversations—and LLMs can only ever respond to our use of it. To create it remains an entirely human endeavor.
Provided by The Conversation
This article is republished from under a Creative Commons license. Read the .