I had the great privilege of speaking to a class of 11 year olds last week on the subject of AI. I had some slides prepared with some questions, as I wanted to make it as interactive and fun as possible.
Interestingly nearly all of the students had used AI in some form, ChatGPT, Google Summaries and Gemini were mentioned, this surprised me, but on reflection I’m not sure why — I think I sometimes think that those of us in the data and AI space live in our own echo chamber, and AI has clearly spread far beyond.
When I asked how AI works there was an interesting discussion, with most describing AI as a rules based or expert system (although not actually using those terms), we then discussed how it actually works and when it would be appropriate for to use.
An example I talked through was based on some homework that I knew they the class had received. They had been tasked with a comprehension exercise about the sinking of the Titanic, along the lines of ‘read these 10 pages and answer 5 questions based on the text’. I downloaded a small model with little knowledge of the titanic, configured a simple RAG setup on my laptop and fed the titanic pdfs into the machine. For those interested you can do something similar following the instructions at this rather cool repo.
I then asked the AI the homework questions, and the model gave decent answers — interestingly about the same quality as an 11 year old student (which I put down to the quality of the comprehension material rather than anything else). But as soon as I asked questions beyond the titanic story, the model did a wonderful job of making up plausible nonsense. Questioning the model about iceberg lettuces was entertaining given how the model attempted to weave a leafy vegetable into the answers.
Get Andrew Mott’s stories in your inbox
Join Medium for free to get updates from this writer.
Essentially we all agreed that we do think carefully about the responses especially if we are using AI for homework!
Then we discussed some of the videos that AI is able to generate, and how these look very very believable, and again can we trust them? Lastly we moved into what I would describe as a broad discussion about ethics.
This very quickly went down a route of what information should we give AIs and also more broadly social media and the internet in general, with some back and forth we concluded that the most important thing was ‘kindness’. Whatever we ask of AI, whatever we write about ourselves or others publicly or privately needs to written with kindness. Everyone broadly agreed that words and actions taken with kindness are best and would be the least embarrassing for everyone (author and subject) in most scenarios.
These 11 years olds are a clever bunch!
My favourite question of the day, was this: “Is AI male of female?”
My initial thought was, that AI is based on the content that it has learnt from, so I asked the class if the content on the internet has been mostly written by males, females or about 50/50. Interestingly the large majority suggested 50/50. I suggested that it was likely that the content used to train most models was likely to be more written by males, and so therefore we might consider AI to be ‘more male’.
I have since performed literally no research at all — I asked Perplexity and ChatGPT questions, the consensus seems to be that male authored content dominates traditional forms of media and especially academic literature, whereas female authored content dominates in some of the newer social media platforms especially influencers.
I’m genuinely not sure what this means, or even if it matters, but it did make me think is this something that folks who are focused on aspects of data governance for AI in large organisations are considering?
