AI has a television set problem

One of the problems of AI-generated content is that it has modelled its content on the basis of content it has read. How do you measure which content has what value? Content can be written casually by a website, by a blogger, or by a PhD scientist. Even if you determine that the person who has written it seems like an expert, how do you know that the person has done it thoughtfully or that it was their best piece of research? Every line that is written has to be assessed as to who has written it and what weight can be assigned to it. That is literally impossible to build. You can build some algorithm, but there is every chance that you could then make a mistake.

When you make a judgment over judgment, that is a problem. When you judge that a line is coming from a credible source, that itself is a judgment. If another perspective comes from another source, how do you judge which perspective is worthy of being good? What is the score value? You cannot apply a mathematical equation to all these things.

You can say, ‘Oh, it’s coming from this source, and this was a bestseller book, therefore it might resonate more with people.’ But it could simply be that people liked other sections of the book and this particular perspective was not as exciting. How do you know that? A blogger may have only five articles and may not have written for many years, but that does not mean the perspective that person shared in one or two lines is missing.

If it is an online website source, you might look at PageRank and backlinks to get an idea of the credibility of the source. Once you have the credibility of that source, how do you measure it against another source which is a book? How do you measure a book version versus an online version? Perhaps you make another judgment that PageRank or backlinks are equivalent to the number of copies sold. Then you are making these equivalent comparisons.

Perhaps you simply say, ‘We are just looking at the essence of the content, and we are judging whether this is a universal way of saying it.’ But how do you judge whether it is universal? You may see how many times people have presented the same perspective, but you do not have the volume of all perspectives. Maybe you miss some old books which are no longer there. Maybe a better perspective was there, or perhaps a perspective which was different.

It is a highly complicated affair to get a paragraph together which is reflecting someone’s opinion. It becomes a paragraph of jumbled-up opinions without a proper weighting exercise because you cannot weigh perspectives. Otherwise, you have to take up a lot of assumptions. I don’t understand how such a paragraph is built. It is the common notions on a limited set of data, but then common notions become a generalised version.

I think AI chat content may turn out to be a very general information tool. Once the wow factor of ‘Oh, instant information’ goes away, when people look at the essence of the information, they may find that it is mainly general information.

There has to be a separation between common procedures and something that requires understanding, weighing and perspective. Common procedures, steps and checklists are what AI is wonderful at. It is excellent for commonly asked questions such as, ‘What should I do? I’m having a bit of a sore throat.’ AI can also be useful if I ask, ‘What other steps do I need to take if I’m planning for a hike?’ The common notion is that you buy a good pair of boots and check all of that. AI is good for it.

The problem begins when something requires a bit of understanding, weighing and perspective. That is a purpose-related judgment. Someone might want a perspective, not from an AI summary, but from someone they can relate to. Perhaps they want a perspective from someone they aspire to be. They may even need a perspective from someone they dislike because they want to sharpen their own reasoning.

I might want to listen to someone who is always against the political party I like so I can understand whether they are being reasonable or not. The perspectives are unshown by AI, and they can never be shown by AI unless the personality behind the perspective is visible. Otherwise, it is just words.

If I ask whether I should go hiking on Mount Agung in Bali or another mountain, that becomes a weighing exercise. People might have written about it, but maybe some old book mentioned that the views are not so amazing at a particular time because of the weather. Someone might have given a perspective saying it is overrated for particular reasons. If AI copies the overrating of an Agung mountain trek, AI cannot reduce that because AI would go by the common notion.

AI can quote a person, but if it does not show the person, even if it shows a small text link corresponding to the article, there is still a problem. One line can be drawn from various perspectives. Different bloggers might have a similar view, such as, ‘I dislike this mountain for a thousand reasons.’ You may have a thousand bloggers, a thousand articles, a thousand pieces of content and a thousand books which have said it. You now have a generalised version of a perspective, but who do you attribute that to?

The attribution problem is a problem because AI does not really know where that generalised perspective has come from exactly. Who do you attribute a disagreement that Agung Mountain is not a good mountain to climb to? It is not that AI models perhaps do not want to attribute. Maybe they do. Who do they attribute it to?

If it is a common-notion perspective, even if it is controversial, it becomes like a perspective without the person. If a perspective is coming from someone from New Zealand, their way of looking at things might be different from someone coming from South America. The perspective is linked to the person.

You watch television, and if you watch television without looking at the screen, at least you get the perspective linked to the person because there is a voice there. Now imagine watching television where you are not looking at the screen and the content being shown to you is AI-generated. The voices are taken away. The actual words are taken away. AI is summarising what is being said on television.

It is like watching television without the characters and without their own chosen words. AI says, ‘In these two minutes of this television series, this happened, this happened, this happened and this happened.’ Is anyone going to enjoy it?

I think that is the problem with perspectives as well. Once you take the person away, take their voice away and take their chosen words away, you are left with a generalised summary of what people appear to have said. The AI model has a television set problem.