In Greek mythology<\/a>, Prometheus is credited with giving humans fire as well as the \u201cspark\u201d that spurred civilisation. One of the unintended consequences of Prometheus\u2019s \u201cgift\u201d was that the need for celestial Gods diminished. Modern humans have been up to all sorts of things that present similar unintended consequences, from using CFCs that led to a hole in the ozone layer<\/a> to building systems that they do not understand or cannot fully control<\/a>.<\/p>\n In dabbling with artificial intelligence (AI), humans seem to have taken on the role of Prometheus \u2013 apparently gifting machines the \u201cfire\u201d that sparked civilisation.<\/p>\n Predicting the future is best left to shamans and futurologists. But we could be better informed about the dangers that follow from how AI operates and work out how to avoid the pitfalls.<\/p>\n First, we must recognise that AI holds immense promise for human society. AI is becoming ubiquitous<\/a> \u2013 from mundane tasks such as writing emails to complex settings that require human expertise.<\/p>\n AI \u2013 by which we mean large language models (LLMs) that appear to \u201cunderstand\u201d and produce human language \u2013 are prediction machines<\/a>. They are trained on large datasets that enable them to establish statistical associations between a huge number of variables and to predict what is next.<\/p>\n If you have used Google, you might have experienced some version of this through its predictive prompts. For example, you might type \u201chow to drive\u201d and Google will complete it as \u201chow to drive an automatic car\u201d. It is unlikely to complete it with \u201chow to drive an aeroplane\u201d. Google establishes this by looking at the history of what words come after \u201chow to drive\u201d. The larger the dataset upon which it has been trained, the more accurate its prediction will be.<\/p>\n Variations of this logic are used in all of its current applications. AI\u2019s strength, of course, is that it can process untold amounts of data, and extrapolate it to apply to the future.<\/p>\n But this strength is also its weakness \u2013 it makes it vulnerable to a phenomenon management scholars refer to as the \u201cconfidence trap\u201d<\/a>. This is the tendency to assume that since earlier decisions have led to positive outcomes, continuing in the same way in future will continue to be OK.<\/p>\n Consider an example: the intervals between maintenance of critical aeroplane parts. If increasing the intervals in the past has worked out fine (no failures), these might be adopted widely and there might be a push to increase the intervals further. Yet, it turned out that this was a recipe for disaster<\/a>. Alaska Airlines flight 261 crashed into the Pacific Ocean killing all 88 people on board because \u2013 perhaps influenced by previous success \u2013 a decision was made to delay the maintenance<\/a> of a critical part.<\/p>\n AI might just exacerbate this tendency. It can take attention away from signs that there are problems<\/a> as AI analysis feeds into the picture to inform decision-making.<\/p>\n Or AI can extrapolate the results of the past and take decisions without human intervention. Take the example of driverless cars, which have been involved in more than a dozen cases<\/a> of pedestrians being killed. No dataset, no matter its size, can provide training for every potential action a pedestrian could take. AI cannot yet compete with human discretion in situations like these.<\/p>\n But more worryingly, AI can diminish human capabilities to the extent that the ability to determine when to intervene might be lost. Researchers have found that use of AI leads to skill decay<\/a> \u2013 a particular concern where workplace decisions involve life-or-death consequences.<\/p>\n Amazon learned the hard way about letting \u201cprediction machines\u201d make decisions when its internal hiring tool discriminated against women<\/a> as it was trained on a database spanning a ten-year period that skewed towards males. These are, of course, examples that we are aware of. As LLMs get more complex and their inner workings become more opaque, we might not even realise when things go astray.<\/p>\n Because AI mirrors the past, it might also be limited in its ability<\/a> to spark radical innovation. By definition, a radical innovation is a break from the past.<\/p>\n Consider the context of photography. Innovative photographers were able to change the way in which the business was done \u2013 the history of photojournalism<\/a> is an example of how something that started as a way of illustrating the news gradually acquired storytelling power and was elevated to the status of an art form.<\/p>\n Similarly, fashion designers such as Coco Chanel modernised women\u2019s clothing<\/a>, freeing them from uncomfortable long skirts and corsets that lost their relevance in the post-war world.<\/p>\n The founder of sportswear manufacturer Under Armour, former college football player Kevin Plank, used the discomfort from sweaty cotton undershirts as an opportunity to develop clothing using microfibres<\/a> that draw moisture away from the body. AI can improve on these innovations. But because of how it operates in its current form, it is unlikely to be the source of novelties.<\/p>\nLooking backwards<\/h2>\n
![\"\"](\"https:\/\/blogs.bath.ac.uk\/iprblog\/wp-content\/uploads\/sites\/115\/2024\/11\/Blog-Images-2-copy.png\")
<\/p>\n![\"The](\"https:\/\/counter.theconversation.com\/content\/242886\/count.gif?distributor=republish-lightbox-basic\")
<\/p>\n