TL:DR Generative AI will build its response to your prompt from a pool of probable additions but it is unlikely to always pick the same ones.
ChatGPT is a form of artificial intelligence known as generative AI. This means it generates new content rather than another form of AI which might recognise faces in an image.
I put the following prompt to Microsoft Copilot (formerly known as Bing Chat) which is a generative AI application from Microsoft.
Write a 50 word summary on the latest episode of Eastenders.
For readers not aware of Eastenders, it’s a very popular TV soap opera in the UK. The answer came back as:
In the latest episode of EastEnders, the fallout from a dramatic New Year’s Day wedding was revealed. Nish died after falling from a balcony, but Suki and Eve survived. The episode ended with a touching moment between Suki and Eve, marking the start of their new chapter together.
I cleared the chat and used the same prompt again. And got this new response:
In the latest episode of EastEnders, Nish’s revenge plot ends in tragedy as he falls to his death, while Suki and Eve survive. The episode features a touching moment with “Julia’s Theme” playing, marking the end of Nish’s torment and the beginning of a new chapter for Suki and Eve.
Why did they differ?
Most generative AI models work by selecting the next part of the output from a range of possible candidates.
For example, if we take an easy example of adding to the sentence “The cat sat on the “, typical words we might choose might include rug, mat, seat, chair, fence, soft, high, sunny, etc.. ChatGPT will select one of these but there’s no guarantee that it will always select the same one.
A comparison which might sit well with you is the gameshow “Family Fortunes” where researchers ask one hundred people questions such as “Name something you can eat”. There’ll be a top answer, probably bread, and then more obscure answers.
That’s a simplistic way of how generative AI will choose the next part of its response to your prompt.
Most generative AI models allow you to adjust how wide the pool of next words is which in turn would either create a more conservative response (a good choice for generating emails, summaries and documents) or a more creative response (good for generating poems, songs or stories).
A narrow pool might include the examples above but a wide pool could include myriad other words resulting in an output such as “The cat sat on the perfectly positioned viewpoint” or “… yellow patch of daisies”.
Imagine asking ten strangers to complete the “The cat sat on the “ sentence. You’re likely to get different responses. Perhaps five people will offer “The cat sat on the rug”. One will offer “…chair”. One person (the creative one) will offer a random but strange “…pineapple”. And the other three people will probably stare at you like you’re an alien and quietly walk away.
This is one reason generative AI is not such a good choice for use cases where you need a repeatable, deterministic response. There will be more suitable forms of AI or logic for those cases.