Artificial intelligence is capable of amazing feats, from writing a novel to creating photorealistic art, but it seems that it isn’t so good at extracting exactly what we want. It fails to grasp nuance or overcome poorly worded instructions. That has given rise to the new job of “prompt engineer” – people who are skilled at crafting the precise text instructions needed for AI to produce exactly what is needed – often with salaries of upwards of $375,000 a year.
This ability to unlock the potential of AI with their “magic voodoo” may seem like a bit of a fad, but New Scientist found that lots of companies find it surprisingly beneficial – at the moment, at least. The question is whether AI will become better at understanding what humans mean and therefore cut out the intermediaries.
Racial bias in AI is nothing new, and sadly it is a trait inherited from data tarnished by human prejudice. Previous cases have made it harder for those with dark skin to get a passport or be shortlisted for a job, but it has now emerged that AI may also place them at higher risk of being hit by a driverless vehicle.
AI of the type used in driverless cars is 7.5 per cent better at spotting pedestrians with light skin than those with dark skin, warn researchers. Part of the problem is the lack of images of dark-skinned pedestrians in training data. Racial bias of all forms needs to be rooted out of AI, but when it is potentially life-threatening it is imperative to act swiftly before that technology is released into the real world.
The Turing test is a famous proposal for distinguishing AI from humans, but in terms of scale there has been no bigger test than the widespread use of CAPTCHA – the frustrating little problems you have to solve when signing up to various websites. Whether it is clicking the tiles that include traffic lights, typing in distorted text or solving an arithmetic problem, the idea is the same: to allow humans past while stopping AI bots whose aim is to abuse the site.
The problem is that AI has become better than humans at these tests. A lot better. More accurate and faster. It seems that all CAPTCHA tests are managing to do is irritate humans. So is it time to ditch them altogether? And should we be worried that engineers struggle to come up with an on-screen task that humans can do better than AI?
It is no secret that AI is big business at the moment. The arrival of generative models has birthed a generation of startups and every big company is rushing to build, borrow or buy a model to streamline some part of their business. The result is that the hardware commonly used to train and run these models – originally designed to power computer games – is in extremely short supply.
Builders of these chips are making hay while the sun shines as new players are understandably muscling in. Meanwhile, trade sanctions are making the global supply of chips a political issue, and academia is increasingly priced out of the AI research that it kickstarted. AI may feel like a software revolution, but it is also very much a hardware arms race.
If you are looking for straight answers from AI chatbots you might have a problem: they seem to just tell us what we want to hear. These digital yes-men are “designed to fool us and to kind of seduce us”, says Carissa Véliz at the University of Oxford. The problem seems to grow worse as the size of the model increases, which is a serious issue because growing scale currently seems to be the best way to make them more capable.
How can we trust an AI if it responds to our questions not with facts or evidence, but with a re-jigged reflection of our own opinions and biases? And should we really be adding that technology to search engines before we have found a solution?