Hallucinating AI vs intuitive AI – which will win?

Today’s AI fabricates facts and fumbles meaning. A new study proposes a model that actually interacts with the world.

“Bon boo bee,” said William, three years old in a Kindergarten classroom I used to work in. William was attempting to count to three in a very playful way, and seemed unconcerned about the teacher's linguistic corrections, and why should he have been?

When you type those words into Chat GPT with no context, it doesn’t have any idea what’s going on. Even in “reasoning mode” with the characters 1,2,3 placed beside the rhyme, GPT deciphered:

“I'm noting that Bon, Boo, and Bee appear to be listed sequentially. This might be part of a phonetic pattern or a creative sequence.”

A recent survey by Computational Linguistics professors at the AI lab of Vrije University, Brussels, raised the question of whether AI can learn a language as a child does.

It’s important to be clear that with LLM in AI, meaning “large language model,” this AI is trained more on massive datasets, rather than grasping the meaning or simply experiencing the sensation.

A computer can’t intuitively interact with its environment and make second guesses, so it’s not as prevalent in the abstract as a human, especially a toddler.

Absurd and offensive

AI models tend to have hallucinations and biases, too, often producing fabricated facts that don’t exist. A few times, while setting out to make a fun yet accurate quiz for my teenage class, AI took the request to mean I wanted a playful, therefore nonsensical style.

When making a multiple choice quiz on the topic of culture, it created untrue and meaningless content about Turkish folklore, which could have offended that region, especially as a country's legacy and identity is not to be goofed around with.

That’s when the strict commandeering for accuracy comes in from the user, which wouldn’t have sat so well with William in learning to count.

The study compared the techniques used by children and computers. Children naturally play and experiment with language through meaningful interaction. Linguistic constructions are learned through interpretation and imitation.

LLM’s meanwhile, observe gargantuan amounts of texts, and in turn cobble together content. They are “extremely powerful in many forms of text generation,” said the report, especially in things like “summarizing, translating, or answering questions.”

The study proposes a new type of model in which the agent plays a more interactive role in their environment, making it more immersive.

This direct interaction with the world would allow a more nurturing “trial and error”-based effort rather than having to use fake content simply to satisfy the request.

Fake it till you make it

As it stands, it seems that Chat GPT and Gemini are notorious for making up citations and occasionally responding out-of-sync with what you want.

The report explains that these alternative AI models would “be more deeply rooted in meaning and intention, enabling them to understand language and context in a more human-like manner.”

Surely a more responsive chatbot-style AI would work better when responding to voice commands, as currently, Gemini is quite buggy when it comes to that, mixing up various layers and answering questions that you might have asked last week.

Such a proposed model would be built from an intention perspective rather than a pattern-based one. Crucially, the environmental benefits would also be huge, as the data would be used more efficiently, with a smaller ecological footprint.

The question is – would companies be able to scale this approach? And, would it even work?

Can AI grow up?

One of the biggest challenges for a proposed instinctive digital language learner is that current models are trained to process billions of words. With such a large vocabulary, how can an LLM be expected to grow up like a child?

Where would the point of entry begin? Would the agent already know just 20 words like William? And should a model be a quiet observer like a baby would be? It’s hard to imagine it not knowing a few hundred words, at least for starters.

Factor in other language elements that might be hard to gauge, such as the phrase “I love you,” which can be uttered aimlessly in a bar in Western culture but may be expressed very carefully but with a deeper emotional impact, like in Japan or Korea.

“Sorry” is often uttered in many Asian cultures in order to avoid conflict, whereas, in many Western cultures, it's used for taking ownership and truly accepting one's wrongdoing. If Chat GPT says sorry, does it actually care? At this stage, probably not – but it would be interesting to see such a model that would demonstrate such emotional intelligence.

The real-world engagement needed for a different type of language learning model would also be an issue. Navigating space with propositions like “up,” “down,” or “near” would require a robotic body to plot a route.

When asked about one's well-being, factor in the tricky emotional dynamics of the sentence “I’m fine,” which could be interpreted with different dynamics such as indifference, politeness, sarcasm, frustration, seeking validation, or even a genuine response.

More cultural nuances like bowing to people in Japan would have to be learned and practiced as well, as language is not just about the spoken, but involves ritual and grace.

This alternative, more empathic, language learning model would need to navigate these subtleties in a more cultivated development, as opposed to bulldozing one’s way through mistakes, and confusion and be largely associated with churning out content with mixed results.