ChatGPT’s ‘Snow White’ Problem: The Danger of Common Knowledge

ChatGPT is an incredible tool that has revolutionized the way we interact with technology. With no barrier to entry, remarkable knowledge, and apparent creativity, more and more people are relying on this powerful AI chatbot to answer their most pressing questions.

The latest in AI services, ChatGPT demonstrates a phenomenal leap forward in machine learning technology, but it’s not without its flaws — primarily the quality of its answers. To solve this problem, we must turn to the rising second branch of artificial intelligence: semantic reasoning.

Snow White and the Explainable AI

If you ask ChatGPT what the earliest feature-length animation movie was, it will, quite confidently, answer ‘Snow White and the Seven Dwarves (1937)’. This sounds about right. We’ve certainly heard it before and we’re sure many wouldn’t question its accuracy. However, the reality is that Snow White and the Seven Dwarves was not the first feature-length animated movie — it was in fact the second. Before it came a relatively unknown Argentine film called El Apóstol — released in 1917, now lost to time.

And that’s where the problem lies. ChatGPT ‘knows’ Snow White and the Seven Dwarves to be the original animated movie because, collectively, we already believed that it was. It’s a common misconception that has been floating around the internet for years, and as such, has been used to train and teach ChatGPT what it now feeds back to us. ChatGPT even knows about El Apóstol and can provide all the information that a human would need to answer the question correctly. The problem, therefore, lies not in the data but in the steps to understand a question and give the correct answer.

The real issue here though is not just that we’re given a false answer, it’s that there’s no way to differentiate it from the truth. As with all machine learning algorithms, ChatGPT can’t explain why it gave the answer that it did. Yes, you can ask for an explanation, but the same problem still applies — you can’t be sure of its truth. Suddenly we’re spiralling into an existential crisis over a fictional hypersomniac and her entourage.

This highlights one of machine learning’s key strengths and its greatest weakness — the answers it gives are based on statistical probability. For some applications, this is the very thing that makes it all possible, but when accurate results are required, or required with context, it becomes clear that we need another approach.

How to Improve Accuracy and Certainty with Explainable Artificial Intelligence

Semantic reasoning, the driving force of rules-based AI, provides a contrast to machine learning — applying a set of logical rules to infer new information from existing data, as opposed to applying a learned behaviour to a new scenario. The clear benefit of semantic reasoning is that any conclusion drawn can be traced back to exactly where it came from, step by step. Explaining the result no longer amounts to peering into a black box; now the workings can be shown in precise detail and you can be 100% confident that the results follow logically from the rules that were set.

To understand the importance of certainty, you need only imagine an autonomous vehicle as it makes a statistical-based decision that could mean the difference between avoiding or causing a collision. By instead relying on the logic of semantic reasoning, some automotive companies are already seeing the benefit of rules-based artificial intelligence.

This is where RDFox comes in. RDFox is a knowledge graph and semantic reasoning engine — a powerful AI software — developed at the University of Oxford.

Using RDFox, we can create a rule that simply states: find the oldest feature-length animated film and give it the tag ‘OldestAnimatedFilm’. By loading a comprehensive dataset into RDFox (we used Wikidata) we can ask the same question as before: ‘which was the first feature-length animated film?’ We ask this via a query instead of natural language but the meaning is the same. RDFox will of course give you the answer ‘El Apóstol’ and, if asked why, will direct you to the rule we just created plus the facts that matched the conditions — complete explainability.

Beyond Data Quality: How to get Accurate Answers with AI

As you can see, the challenge goes beyond the availability of accurate data — good data quality is vital, but so too are the interpretability and correctness of results. In the current context that might not seem as life-and-death as we’re making out to be but consider for a moment some of the questions that are being asked in the real world.

Has this individual committed fraud?

Does this patient require treatment?

Is this vehicle safe to turn?

Suddenly the need for accurate, auditable results is critical. There is no room for error when the stakes are this high. And even if they weren’t, wouldn’t you rather decisions were made based on answers that could be backed up by logic and reason?

Can rules-based AI like RDFox write you a poem? No. But it will give you correct answers and for some, that’s the difference between revolutionary success and complete catastrophe.