Defining truth is a question impossible for mere mortals. Defining facts is a good deal easier.
Determining whether a statement is provable is still a subjective exercise, but one familiar to journalists and juries. In other words, the distinction has economic value. To give just a few examples, it plays a role in determining insurance premiums and whether medical research does or does not get funded. Distinguishing between positive and normative statements is essential for figuring out budgets: where schools and hospitals will be built, and the funds allocated to police and first responders. In other words, the ability to understand and recognize facts affects people’s lives and health directly.
Can an AI be taught to recognize the same types of distinctions? I would argue that it most certainly can. It probably doesn’t even require an LLM. Just a simple text classification model.
Here’s how I would do it.
We have a hypothetical function:
is_provable(x)
Input: String of up to 1000 characters in length.
Output: Boolean
Goal: Evaluate whether this string fits the definition of a factual statement (e.g. one that can be verified as either true or false).
It should be possible to train a text classification model to evaluate different statements and determine whether they are likely to be provable or impossible to prove.
Examples of provable statements:
Quantitative Expressions – Ex. “The population of the United States is 330 million people.”
Comparative Statements – Ex. “The Nile is the longest river in Africa.”
Direct Quotations – Ex. “John F. Kennedy told the people of Berlin, ‘Ich bin ein Berliner.'”
Descriptions of Past or Present Events – “On June 6, 1944, Allied forces landed on the beaches of Normandy.”
In general, data that can be cited or attributed may be considered factual. However, this depends on trust in the methods and judgment of those compiling the information source.
I need to stress that the goal here is not to determine whether the statement itself is true or false. It is only to predict whether the statement is possible or impossible to verify (e.g. “a fact”).
Not every important statement can be proven. For example, scientific hypotheses are not provable, because future evidence may call them into question — but summaries of experimental results certainly qualify as factual statements.
Of course, plenty of rabbit holes and pitfalls with this approach. I should emphasize that I am proposing something more akin to sentiment analysis than a precise epistemology. The reason this exercise might be at all worthwhile is that at the end of the day, it could be used to build a repository with some very interesting and relevant applications.
Wikipedia would be the obvious place to go for training data. A pool of human undergraduates (pre-law, economics, psychology, and philosophy) could provide a secondary source of validation data. Members of Debate Clubs would be your ideal candidates.
Yes, we live in the era of fake news. Stock valuations are difficult to pin down. The concept of calling a politician or corporate leader out for telling a lie seems quaint and old-fashioned. In practice, we are overwhelmed with information. Trust is a vanishing commodity. Training an AI model to distinguish between factual and non-factual statements will not restore that trust. What it will do is allow collection of a dataset that may inform the basis of a worldview. One defined based on human standards, but accessible and interpretable by machines. (And by the way, we are certainly talking about an LLM here — one trained with a massive amount of parameters.)
To some, that prospect may seem frightening. But I would argue that if unchecked, reliance on AI systems incapable of distinguishing between fact and fantasy could result in far greater harm.