Can we all just calm down about AI please?
It has been observed that generative AI (i.e. AI that can generate text, images, audio and video based on user-supplied prompts) has been around for several years, but was not really featured in the media until journalists realised that their jobs were at stake.
That is not to say that comment was absent prior, eg the 2020 Guardian article headlined "A robot wrote this article. Are you scared yet, human?". The robot in question was GPT3, the same software that powers ChatGPT, only an earlier version. This was an example of hype, as it was clear to those familiar with generative AI at the time, and subsequently disclosed by The Guardian, that GPT3 output was edited, and several attempts were needed to produce usable output. The article was several prompts and outputs stitched together. Fast-forward a couple of years, and even given the substantial advances from GPT3 to GPT3.5 and now GPT4, that is still the reality of working with generative AI in most contexts.
The news media has been so flooded with stories about generative AI that I hesitated to write this one, but I persevered because I am sick of the uninformed hyperbole. The danger is not from a "rogue AI" generating a killer virus that wipes out humanity because it’s been programmed to ensure the long-term survival of life on Earth: rather the danger is from people in power who believe the stretching of truth, and outright lies of sales people who sell AI to governments and business, and other AI boosters who seek to shape public opinion.
AI is just a type of software that is so complex that even the people who wrote the software don’t really understand how or why it works, in the sense that given a certain input, it can be predicted what the output will be, and also it can be explained why and how the output was generated. Creators of AI systems often express surprise and amazement at the capabilities of the system they created, and it’s natural to be a little fearful as well.
There is a well-founded fear that AI systems can do something unexpected and unpredictable. It is a well-known feature of AI systems that they often get things right, but also get things wrong and make things up. The sensible thing to do is treat them like a clueless intern who knows all about how to spot tumours in radiographs but nothing about human anatomy or physiology. Hence, don’t put that software in control of anything that could cause harm when it makes mistakes.
AI is causing real human suffering right now among people who write or create and edit images for a living, because business owners are replacing them. That’s harmful, but on the lower end of the scale of harm, compared with, for example, people wrongly identified by facial recognition being held in jail. Some have even been convicted and imprisoned before the mistake was uncovered, although it must be noted that the same thing has occurred based on human eyewitness testimony, which is also notoriously unreliable.
There are many technologies that are beneficial but potentially harmful, and so we regulate them, such as pharmaceuticals. The time has certainly come to regulate vendors of AI systems and put rules in place, at the very least for government workers regarding purchase and use of AI systems. Due diligence must be mandated, and not at the level of purchasing a new printer, but more along the lines of purchasing a drug that will be administered to all government employees and elected politicians every day.
It will take longer and much more work to regulate AI vendors in the same way we regulate the pharmaceutical industry, not least due to the enormous wealth and hence lobbying power of the companies involved. New Zealand would do well to follow the EU and United Kingdom rather than the United States, given the vulnerability of US lawmaking to lobbying.
— John Williams is a senior lecturer at the University of Otago.
