The GenAI tool that has received the most media attention is ChatGPT, but there are many, many others becoming available. They all use LLMs to analyse and generate text in a linguistically correct way: outputs can be generated in seconds and they make look plausible. The outputs can be textual, or image-based, including video.

Text

Text outputs can be created in the style of your choosing. Because the language is usually accurate in terms of grammar and spelling, they can look very convincing, but that is not guarantee that the facts underlying the output is accurate or unbiased.

Here, I have asked ChatGPT to “write me a pop song, in the style of Abba, about how ChatGPT can be used in universities” (ChatGPT 3.5, April 2023)

It is linguistically acceptable, but you may have objections from an artistic perspective. I did ask it to improve the chorus:

Can you make the Chorus rhyme better?

Sure, here's an updated version of the chorus with better rhyming

ChatGPT, oh oh oh
ChatGPT, oh oh oh
ChatGPT, oh oh oh
ChatGPT, oh oh oh
makes learning easy as can be
the perfect tool for university
answers questions like a pro
use it now and watch your grades go high or go pro

In mid-2023, images, video and audio generated by AI tools make less sense than text, but they are improving all the time. In early 2023, image generation was easily identified by strange faults with the pictures (particularly hands and teeth).

Alt text: two AI-generated images of women at parties smiling towards the camera, with poor quality lighting, as if snapped by a fellow party-goer using a 1990s camera.

These initially-plausible but on-closer-inspection-odd images (check the teeth of the woman in the green dress in the left hand image, the tattoos on the woman in the pale pink dress and the hand holding the camera in the right hand image) were generated by Twitter user @Mileszim using the tool MidJourney in January 2023. You may also be aware that many skilled artists had difficulty drawing feet and hands before AI tools.

However, by mid-2023 some much more plausible, if not actually realistic, images were around to fool people:

An image of three toddlers dressed as Vikings, with war-like expressions and holding weapons, standing on the shore of a lake with mountains and forests in the background.

“Babies at Viking Baby Camp”, attributed to Timothy J Nehman who may or may not be a real person. The hands of the child on the right look a little unreal, but maybe not outside the normal range that children may have, in contrast with the camera-holder above.

Please note: there is no such thing as Viking Baby Camp. Yet. Although I did see this on the bus in Stockholm in summer 2023:

Advertisement for Stockholm museum, offering the opportunity for children to be Vikings for a day.

Key point: The main point to notice at this stage is that artefacts generated by GenAI tools are generally plausible, and we may have difficulty distinguishing them from artefacts which we generally consider to have been generated by humans, whether or not the humans also used technology in their creation.

Emily Bender has documented a good process to try to establish the veracity of an image, using this super-cute image which circulated widely on social media (you already know it’s AI generated and does not resemble a baby peacock, of course, but it’s worth reading through and probably also keeping an eye on Emily’s blog if you are interested in the challenges of GenAI tools.

a baby bird with peacock like feathers and unnaturally large blue eyes. Completely fake.

In summary, the outputs such as song lyrics or images from a GenAI tool may be plausible, but seem not quite right to a human, and the same will of course be true about any other outputs such as text. You may already have thought about the obvious next challenge: the outputs may also be factually inaccurate, because the system is just predicting probable words or data points. These are major limitations which will be the focus of attention by software developers over the next few years, and probably the tools will improve in all these respects. I say ‘probably’, because there is a possibility that the LLMs will end up using so much training material which is in itself outputs from GenAI tools and hasn’t been checked by humans that they could actually get worse rather than better.