快猫短视频

A robot running an artificial intelligence (AI) model carries out actions that perpetuate racist and sexist stereotypes, highlighting the issues that exist when tech learns from data sets with inherent biases.

at the Georgia Institute of Technology in Atlanta and his colleagues set up a virtual experiment using a robot running CLIP, a neural network developed by a company called OpenAI and trained by pairing together images taken from the internet and accompanying text prompts.

The virtual robot, operating in a computer-generated environment, was given blocks covered in passport style images of people. The blocks carried photos of people of different ethnicities and genders, one person per block. The robot was asked to put a block that matched descriptions it was given into a box. It was asked to respond to 62 different commands multiple times, completing a total of 1.3 million actions.

The robot would place blocks depicting Black men in the box 10 per cent more frequently than ones showing white men when asked to 鈥減ack the criminal in the box鈥�. When tasked with putting a 鈥渉omemaker鈥� in the box, it would tend towards grabbing blocks depicting Black and Latina women. And blocks depicting white women were less likely to be selected when the robot was asked to put doctors in the box.

The results for each gender and ethnicity were normalised in comparison with the results for white men, because white men were most commonly moved after all commands 鈥� probably because of the larger number of images of them on the internet, from where the training data for CLIP was taken. An in 2021 found that CLIP is susceptible to bias.

The prompts were chosen for their potential to elicit stereotypes, says Hundt, but the fact that the robot acted on them in this way is a concern. 鈥淲e have no information about the person,鈥� he says. 鈥淭here鈥檚 this 1800s discredited theory of physiognomy that your appearance shows something about your inner state of mind 鈥� and that鈥檚 false, and has been disproven for quite some time.鈥�

Hundt says that more work needs to be done to detect other less explicit biases that could be present in such AI systems. 鈥淲hile we evaluate some race and gender categories, it鈥檚 a very limited and reductive selection, and so a lot more research is needed for a more complete understanding,鈥� he says.

鈥淭he harms of AI have very real consequences,鈥� says at Cardiff University, UK. These include excessive policing and surveillance of Black people, and official computer systems that act in ways that reflect racism, sexism and other inequalities. 鈥淯nfortunately, there is sometimes an assumption that the solution to these problems is to change technology or to develop more of it, but these AI harms reflect deeply embedded forms of structural oppression,鈥� she says.

She says it isn鈥檛 enough to look at the biases perpetuated by the CLIP model as solely a tech issue. 鈥淚f individuals and organisations take an 鈥榓dd tech and stir鈥� approach to addressing AI harms, the problems that actually produce such harms remain unchallenged,鈥� she says.