Smart Model dedicated software allows edit?
Not long ago, an article was published about researchers developing a new tool that allows people to experience for themselves the inherent biases of artificial intelligence models against different genders and races.
While a lot has been written in the press about how our biases are reflected in AI models, it still feels uncomfortable to see AI-generated humans so white, masculine, and old-school, and this is especially true for the DALL-E2, When prompted with words like “CEO” or “Director (director)”, it generated a white male 97% of the time.
In the wider world created by artificial intelligence, the problem of bias is more entrenched than you might think. Stanford University researcher Federico Bianchi said that because the models were created by U.S. companies and most of the training data came from North America, when they were asked to generate even more mundane everyday objects, such as doors and houses, they would Creates objects that look “North American”.
As AI-generated imagery continues to emerge, we will see a deluge of images that reflect American prejudices, culture, and values. Who knows if artificial intelligence will eventually become the main tool of American soft power?
So, how do we solve these problems? There is certainly a lot of work to be done to address bias in training datasets for AI models, but two recent papers suggest some interesting new approaches.
What if you could just tell the model to give you less biased answers?
Researchers at the Technical University of Darmstadt in Germany and AI startup Hugging Face have developed a tool called FairDiffusion that makes it easier to tune AI models to generate the type of image you want. For example, you could generate a photo of a CEO with different settings, then use Fair Diffusion technology to swap the white male in the image for a female or someone of a different race.
As the Hugging Face tool demonstrates, many text-to-image AI models have very strong biases for occupation, gender, and race due to biases in the training data. The German researchers’ Fair Diffusion tool is based on a technology they developed called semantic guidance, which allows users to instruct AI systems on how to generate images of humans and edit the results.
Christian Kerstin, a German computer science professor who was involved in the work, said the images generated by the AI system were very close to the originals.
Felix Friedrich, a doctoral student at Darmstadt University, said the method allowed people to create the images they wanted without having to do the tedious and time-consuming extra work of trying to improve the images used to train artificial intelligence. Biased datasets for intelligent models.
However, this tool is not perfect. Changing the picture of some occupations, such as “dishwasher”, does not work well, because the word can mean both a dishwasher and a person who washes dishes in English. Ultimately, the diversity of people the model can generate is still limited by the AI system’s training dataset. Still, even if the work needs more refinement, it could be an important step toward mitigating bias.
Similar techniques seem to work for language models as well. As recently reported, research from artificial intelligence lab Anthropic has shown that large language models can be guided to produce less harmful content with just simple instructions. Anthropic researchers tested language models of different sizes, and they found that if the models were large enough, they would self-correct some biases at the request of users.
The researchers don’t know why the AI models that generate text and images do this. The Anthropic team believes this may be because larger models have larger training data sets that include both many examples of biases or stereotypes and examples of people resisting and correcting those biases.
Artificial intelligence tools are increasingly popular for generating images. Tools like FairDiffusion could be useful for companies that want their promotional images to reflect diversity in society, Kerstin said.
These welcome approaches to combating bias in AI raise the obvious question of whether they should have been incorporated into models in the first place. Right now, our best generative AI tools are amplifying harmful stereotypes on a massive scale.
It’s worth noting that clever engineering alone won’t completely eliminate bias. As researchers at the National Institute of Standards and Technology (NIS) pointed out in a 2023 report, bias exists in more places than just data and algorithms. We need to investigate the way humans use AI tools, as well as the broader social context in which these tools are used, all of which can lead to problems of bias.
Effectively mitigating bias will require more auditing and evaluation of how AI models are built, and the data incorporated into them, as well as increased transparency, NIST said.