As Chatbots Show Hints of Real Understanding

BY KIT EATON@KITEATON

It's hard to ignore the artificial intelligence chatbot wave washing over the world. Tech giants from Microsoft to Quora are leveraging systems like OpenAI's ChatGPT to bring chatbots to everyone from consumers to small business users, with implications for the future of the workforce.

Researchers from Princeton and Google's DeepMind announced they have found hints some large chatbots may be a tiny bit more like the sinister HAL 9000 computer from the movie 2001: A Space Odyssey than a dully functional text chat system that helps answer a customer IT helpline. 

Complex legal issues are already emerging around the Large Language Models (LLM)  that power chatbots like ChatGPT. These include arguments that LLM training data incorporates stolen content that was found online. Some artists even sued generative AI company Midjourney over allegations their art was used to train its systems to generate art in the styles of the original creators. Voice cloning is now also a thing and, though it's not AI in its own right, could be combined with a text model like a chatbot that could then generate audio output that seems and sounds like it came from a real person-with implications for the acting profession. 

The question of how your company can use these generative AI models as part of daily business operations without exposing it to legal risks is already being asked.

So the big, tricky, legal question emerging from the new research is: Are chatbots sort of alive now? 

When you order a chatbot like ChatGPT to do something like, "help me write an ad for my boutique smartphone repair company?" its algorithms consider your request and generate text based on the vast array of text sources to which it has already been exposed. The chatbot's output is influenced by millions of slight changes in its model that were caused when it was trained with text about writing ads, about smartphones and what a company is and so on. 

Since LLMs are statistical at heart, that's partly why when you ask a chatbot to answer your questions its outputs can vary, particularly over time. Sometimes the replies can be totally unreal hallucinations or even factual mistakes.

Sanjeev Arora of Princeton University and Anirudh Goyal from DeepMind suggest that in the largest of LLMs the ever-increasing amount of training data is actually giving the chatbots new skills. LLMs' language skills are getting better as their training continues, and importantly they're also combining skills in a way that is "unlikely" to have existed in the training data. This means LLMs aren't just repeating phrases and data they've already seen. New skills are instead emerging from the AI's systems, in a way that can be considered the very first glimmering of understanding.

This is all, of course, hugely theoretical, even philosophical. Right now it really will not impact your use of chatbot systems-particularly if you're using one of the new generation of custom GPTs, trained on a smaller dataset unique to the particular requirements of your company. You can happily go ahead and connect up, say, your website's interactive FAQ to a chatbot without worrying you're enslaving a basic form of computer intelligence-though as Chevrolet found out this isn't a trick totally free of risk!

The news has more serious implications for the future of LLM AIs, though, and how they'll be used in business.

In the U.K. a landmark legal ruling decided that an AI cannot be a patent holder. But the former founder of GoogleDeepmind recently said he thinks AIs will soon advance to the point they'll be able to act as CEOs, dreaming up a concept for a company and even running it. That sounds like real creativity and understanding. It's also worth remembering that Facebook's Mark Zuckerberg is directing his company's efforts into developing real AIs-proper Artificial General Intelligence systems-that will far surpass the chatbot tech your PR team may already be using.

The notion that CEOs of companies large and small should be incorporating AIs into their businesses is already well known. The implication that AIs may already be getting clever enough to understand your queries and that the tech is evolving really fast, means CEOs should also have a plan to keep their finger on the pulse of this incredibly fast-moving tech, and possibly refresh business continuity plans.

Photo Credit: Getty Images.