10 AI challenges that need to be addressed

With ChatGPT reaching 100m users in only 6 weeks after its launch (the World Wide Web took six years) there’s no doubt that we’re living through a technological watershed. But while AI is presenting truly ground breaking opportunities there are several issues causing concern.

At our recent Digital Summit, we were joined by the 英国牛津大学 Professor of Computer Science, Michael Wooldridge Wooldridge, to discuss the impact of Generative AI. As a fellow of the Association for Advancement of Artificial Intelligence and of the European Coordinating Committee for Artificial Intelligence, Michael’s finger is directly on AI’s pulse.

Here he looks at the other side of the AI coin, and briefly outlines ten challenges that have to be addressed in the near term.

1. Large language models lie. Frequently and convincingly

One of the biggest challenges we’re experiencing is that LLMs have no conception of truth or falsity. They are not trying to tell you the truth and they're not thinking about what they're doing. They're simply neural networks trying to make a best guess about what words should come next.

For example, when GPT3 was released, I tried the prompt: “Tell me about Michael Wooldridge”. It gave me a few sentences so there’s clearly data on me out there on the web. Part of the response said I had studied for my undergraduate degree at Cambridge. I didn’t. I've never studied at Cambridge so, why would it return that? It’s because GPT3 has read hundreds, thousands of biographies of Oxford professors and a very common feature of those is they studied at Oxbridge.

Maybe it had no information about where I studied for my undergraduate degree and essentially made its best guess. And it's very plausible, if you'd read that you wouldn't have thought there was anything remarkable about it.

The fact it gets things wrong but gets things wrong in very plausible ways is a fundamental limitation of the technology with enormous effort directed at trying to fix the problem. But it's possible this limitation may be one of the breakpoints for AI.

?2. Bias & toxicity

If your understanding of the world has been obtained through Reddit, then your understanding of the world includes every kind of obnoxiousness that's imaginable, and quite a lot that most of us couldn't imagine. It's all there in the neural networks.

Whilst networks are trying to build guard rails to prevent returns on certain types of query, they are quite flimsy and you don’t need a PhD in computer science to work around them.

In a similar vein, the AI network's training data is basically North American English, and as a consequence, those neural networks are acquiring the biases that are implicit within that data, raising some serious concerns.

3. Copyright

Another important AI challenge is copyright. The New York Times are busy suing OpenAI because they know that their copyrighted articles were used in training data. I know that some of my books have been used in training data, and other prominent authors know that their books have appeared too and are busy suing them as well.

AI companies are responding in a variety of ways and they have some of the deepest pockets in the world, so I expect to see fireworks as the legal arguments unfold.

?4. Intellectual property

The same can be said for Intellectual property. Imagine you're JK Rowling and spent years inventing the Hogwarts universe, you release your first book, and it's a smash hit. Then the next day, the internet is populated by thousands of Hogwarts clone books which capture your tone and your style perfectly - that's a world that we might be heading to.

What does that mean for creative professionals and their intellectual property? As well as unhappy authors the same applies to filmmakers, photographers, artists, anybody who creates content, and whose livelihood is based on their intellectual property.

?5. Missed employment

One big question is the extent to which AI is going to affect employment. I think for the vast majority of people, AI is going to be another technology that they use in their workplace, like Microsoft Copilot, and very often people won't realise that it's AI.

But there are signs that some sectors are being affected, for example, people who produce routine content in the gig economy such as SEO content writers. The content they produce is not high art, it's simple prose but it's a living for them. I think this year is going to be a year in which, if we see the effects of AI on employment, we’ll see them decisively by the end of the year.

?6. Giving power to bad actors

The government in the UK is very concerned about giving power to bad actors. What happens if you're a terrorist organisation and ask: "How do I build a pipe bomb? How do I build a chemical or a biological weapon?" The government are worried that this technology may be used in that way, which is why we now have the AI Safety Institute in the UK.

?7. We’ve used all the data

If you've already used all the data in the world, where are you going to get 10 times more data from to build ever larger models – we’re literally at the point where all the data has been used. Will less scrupulous companies use your data without permission in their training models?

?8. Will AI data pollute the internet?

I don't usually like making predictions, but I'm fairly confident that within a decade the overwhelming majority of content you’ll see on the internet will be AI-generated. So then we've got AI data polluting the internet, and that data is going to start being ingested when we train these models. We know that's not good for these models, so it’s going to be interesting.

9. Who controls the technology?

At the moment, from a UK perspective, it's a tiny number of foreign-owned companies that own this incredibly powerful technology. Governments across the globe are concerned about that.

10. The cost

Facebook announced recently their state-of-the-art large language model required about 60,000 AI supercomputers, with each of them costing between $50,000 and $100,000. About 60,000 supercomputers running for months. The estimated cost of training that model is something like $450 million.

But it's not just the dollar cost, but the cost in terms of the environment. You're generating a lot of CO2 to train a model.