What Does Bird Watching Have to do with AI?

What Does Bird Watching Have to do with AI?

By Matt SMITH

July 2024

 

I was speaking to a Manager at a UK bank today (let’s call her Jen), about the wonders of AI. The conversation took an unexpected turn though when Jen told me that the bank’s considering implementing a new policy which essentially bans employees from using Generative AI technologies at work. 

So, the current situation is this: employees need help writing a document. They sketch it out, send it to something like ChatGPT or Copilot and ask the AI to finish the job. 

There’s a lot of potential in the use of AI in this way. Who wouldn’t want a (pretty much) free way of spending less time writing documentation? 

The issue though is that AI tends to be pretty confident – why’s that a bad thing though?  

Well, I think it all comes down to what’s known as the Dunning-Kruger effect. A very interesting Wikipedia article on the subject gives a good definition:  

 

“The Dunning–Kruger effect is a cognitive bias in which people with limited competence in a particular domain overestimate their abilities.” 

 

We’ve all been there… we start learning a new skill and think we’re great at it… until we realise we’re not! 

Let’s say I want to get into bird watching – I start learning about birds and in just a day I have loads of new bird facts I want to share with my friends. I’m now pretty knowledgeable about the top 10 birds in England. Blue Tits, Sparrows etc. I feel like Bill Oddie! And after only a day I have a high level of confidence about my new hobby.  

But, the next day I go out with my binoculars and I start seeing a load of birds I’ve never seen before! And after a bit of research I now find out there are 634 species of birds in the UK.  

Don’t believe me? List of birds of Great Britain - Wikipedia 

I was really confident about my bird knowledge. But, I only actually know 1.5% of my newly found hobby! My confidence therefore drops.  

Finally, over the next few weeks I do a load of research, get to know hundreds of species of birds and my confidence level, quite rightly, increases. 

And just for fun, here’s Microsoft Copilot’s attempt at creating a chart based on the above data! 

image (6).png

This isn’t a new thing that’s come along with AI – the concept’s been around for a couple of decades. And it describes people, not AI. So, why does it apply to Jen and her colleagues at the bank? And if AI isn’t self-aware then how can it have a ‘brain state’ resembling overconfidence? I have a couple of degrees in Philosophy so I could talk about this topic all day... But, for now, let’s not worry about consciousness and instead let’s talk about… hallucinations? I feel like I’m opening a can of worms here, replacing one difficult concept with another. But bear with me! 

So, what are AI hallucinations? Sometimes, when an AI doesn’t know the answer to a question, it just makes one up! For further reading on the topic, I’d recommend this great article by IBM: What Are AI Hallucinations? | IBM. If you don’t want to read the whole article, here are three examples that IBM uses: 

 

  • Google’s Bard chatbot incorrectly claiming that the James Webb Space Telescope had captured the world’s first images of a planet outside our solar system. 

  • Microsoft’s chat AI, Sydney, admitting to falling in love with users and spying on Bing employees. 

  • Meta pulling its Galactica LLM demo in 2022, after it provided users inaccurate information, sometimes rooted in prejudice. 

 

So, is AI just acting like a new hire that doesn’t want to admit they lied on their CV? Or Matt on day one of his bird watching hobby? Who knows – that’s too complex a discussion for this blog (as it involves assigning brain states to technologies that don’t have a brain). But whether AI ‘feels’ over-confident or not isn’t really important – either way, AI can act in a way that appears overconfident when it provides an incorrect answer. 

Don’t get me wrong, I’ve also seen many examples of AI acting reasonably. For example, Copilot suffixes its responses with the following: “AI-generated content may be incorrect” (which I think is a bit lazy, but that’s a topic for another blog…). 

 

Back to Jen then – she tells me her colleagues don’t just use it to spruce up their documents, but they also ask it policy questions that are unique to the bank. For example: 

 

  • How much paternity leave am I entitled to? 

  • What day of the month do I get paid on? 

  • Where’s the template for onboarding new IT vendors?

 

These are all questions that could cause havoc and confusion if the AI provides an incorrect answer. And worst still what about the bank’s customers that also have access to the same AI through the bank’s online chat function? 

 

  • Am I entitled to a loan? 

  • What exchange rate will I get for a £5,000 loan? 

  • Can you forgive my loan please? I can’t afford it?

 

I had good news for Jen though. I asked her “what if I could tell you that you could keep the AI and ensure that it no longer hallucinates?” Obviously she bit my hand off and asked me how much it would cost and when she could have it! 

The answer is relatively simple actually. And there are two steps to it: 

  • Step 1: “Train” the AI on the bank’s policies (in reality it isn’t “training” in the traditional sense, it’s more like providing the AI access to policy documents and the like) 

  • Step 2: Control the “prompt” sent to the AI 

 

I’m the product owner of an amazing bit of tech called dataServe® and it was designed a few years ago to perform the above actions and give colleagues and customers a safe place to interact with any AI that has an API (Application Programming Interface). dataServe® ingests an organisation’s data from any specified location – for example, technical documentation from Confluence, HR documents on OneDrive, and company announcements on SharePoint. The data is then indexed and made available to an organisation’s chosen AI tool safely. That’s step 1. 

Step 2 is really easy – when dataServe asks the user’s question to the AI, it specifically instructs the AI to not answer the question if the answer isn’t present somewhere within the indexed data lake. 

This is called Retrieval Augmented Generation and we’ve created a video to help explain the concept to our customers. We were really ahead of the curve when it came to this type of technology – and we were in fact performing these functions before the concept had a name! 

 

If like Jen you’re worried about using AI technologies, please get in touch to see how we could help you to provide a better customer experience and improve the working lives of your colleagues. 

 

So, what does bird watching have to do with AI? Well, according to Copilot, quite a lot! 

“AI enhances bird watching by identifying individual birds and species, collecting data, and analyzing bird behavior. This technology helps both bird watchers and researchers better understand and enjoy bird life. Are you interested in bird watching?”