We all live in grotesque meat cages and, as we get older, we become increasingly sure of what is right and wrong about the world. We do not enjoy the idea that we can be duped by a robot in any sphere of our lives. And this arrogance, I think, has forced us to believe that Large Language Models (LLMs) will never have the power to convince us of a misleading fact or a dangerous narrative online.

However, a new study suggests that LLMs like ChatGPT can be used to craft convincing arguments… as long as it knows personal information about you, which, of course, it can find out thanks to social media.

Researchers found that when GPT-4 was given a person’s basic demographic information it was considerably better than a human at persuading a participant to change their opinion in a debate. A web-based platform was created for the study and participants either engaged with a human or LLM opponent.

Read more in Daily Maverick: After the Bell: Quelling questionable quandaries with GPT-4

When the arguments were personalised based on background info, GPT-4 increased the odds of shifting opinions in its favour by 81.7% compared to humans. Without personalisation, GPT-4 still outperformed humans but to a far lesser degree.

But the kicker was when humans were given the same info on who they were debating with they did worse than when they knew nothing about them. This suggests that when humans were given advantageous information it clouded their judgement and they weren’t able to quickly use it in a way that benefited them. That is a pretty bleak result.

And there is a deeper, crazier level to this. LLMs are largely kept in their place at the moment and seem to struggle with tasks that require exploration or strategy. This paper from Cornell University which was updated in December last year reckons it has cracked the code to change this with a “Tree of Thoughts” approach. This uses a different way of using an LLM that allows it to consider multiple different reasoning paths and make a decision based on that information (as well as looking ahead and backtracking if it thinks it has made a mistake).

Researchers tested the theory by asking an LLM to play Game of 24 (a maths game where you need to calculate 24 with an assigned set of numbers) and got it to jump from a 4% success rate to 74%. This reveals that the models we are using for rather basic tasks at the moment have huge potential, we just need to figure out how to unlock it.

I covered AI and elections a few weeks ago and there is a nervous energy this year around how AI chatbots are being weaponised to spread propaganda and disinformation at scale. We have to take note that when we use ChatGPT we are seeing the shiny face of the model, but they can be tuned and adjusted (with Python code that you can ask ChatGPT itself to write) to create bots that can spread all kinds of chaos. And we are probably at a point where we need to think about making counter bots of our own.

How AI can improve your life

GPT-5 is coming but there is a good argument for why you shouldn’t care. We are caught in an AI hype cycle of the major players endlessly releasing new models that will keep promising to improve your life. Certain corners of the Internet will drip-feed you tidbits and “leaks” that can make you feel like you need to keep up to date.

But the best way to judge the competency of AI (and how it can help you) is to start keeping a spreadsheet of all the work tasks you do in a day. This may take some time to build. Then systematically start asking ChatGPT to perform these tasks for you, one at a time. And then in your spreadsheet give the results a rating. When the rating hits a 10, move the task over to being ChatGPT’s responsibility (before that you will need to keep doing it). This is the best way to keep track of where AI can help you in your work and where it can’t.

This week’s AI tool for people to use

I’ve been cynical of “prompt engineering” in the past (where people go on courses to learn how to ask an LLM questions), because it feels like OpenAI are on a mission to improve their service so anyone can get what they want without much effort. However, even the AI giant is admitting people need help. The OpenAI Cookbook is an incredible resource of papers, videos and guides to help you write killer prompts.

This will be a rather dry hour of YouTube, but is worth your time to learn how to understand why certain prompts work better than others.

What AI was used in creating this piece?

I used ChatGPT for the picture and to read and summarise this PDF for the main article. I asked it to give me the top 10 facts of the paper. It’s interesting because I wouldn’t usually trust an AI to do summaries in my work, but it feels like I need to start testing and displaying AI’s capacity in this letter. It did a pretty good job, but I still needed to go and check if the results were correct.

What’s new at Develop AI?

Plenty of exciting training is happening at Develop AI. During the upcoming week, I’ll be doing two AI online training sessions for The British University in Egypt. The first is on ‘Data Verification’ where I will equip students to select the best tools to check the accuracy and completeness of their datasets. The next one I’ll be teaching is on ‘Data Visualisation and AI’ to make sure students visually represent their stories in an engaging way and use AI tools responsibly. If you are interested in Develop AI conducting a workshop for your team or institution, get in touch. DM

Develop Al is an innovative company that reports on AI, builds AI-focused projects and provides training on how to use AI responsibly.

Subscribe to Develop AI’s newsletter here.