The world of artificial intelligence has been determined for a few years to leave us with our mouths open every so often, and the truth is that they are achieving it. From specific companies in the sector such as OpenAI to large technology companies such as Google, in these times we have seen how artificial intelligences have learned to do everything, from writing a text on a specific topic that we indicate, to rescaling game images in real time to enhance the gaming experience.
A very interesting area of this evolution of artificial intelligence can be found in image generation, in some cases purely randomly, in others based on a description previously provided by the user. And in this last group we can find specialized solutions for a specific type of image, such as NVIDIA’s GauGAN2 or Google’s popular face generator, as well as general purpose ones.
Among the latter, until now the gold medal corresponded to the OpenAI DALL-E system, but this has changed with the presentation of Image, a new system designed by Google for generating images from text descriptions. And it is that, as we can see on the presentation page of the project, some of the images generated by this artificial intelligence can pass perfectly for real, except for the fact that at least part of it reproduces not entirely realistic motifs.
It is important to mention, however, that as always happens in these cases, Google will have made a selection of the best results obtained, but we can assume that some not-so-effective outputs have also occurred. In addition, this is something that we will not be able to verify, at least in the short term, since Google has decided that, at least for now, it will not make Image available to potential users/customers.
There are two reasons for this. The first is that the company is concerned about potential malicious uses of Imagefrom fake news to illegal sexual content, such as CSAM, Google acts with a lot of common sense on this point, since we can be sure that such uses would occur almost from the first moment that this technology became accessible to everyone.
The other, and which also makes a lot of sense, is that, as is the case with practically all the massive datasets used to feed the learning processes of artificial intelligences, we can find in them some biases that should be corrected. The problem is that performing a detailed analysis of these datasets is a daunting task, which makes even companies the size of Google pale. Thus, instead of approaching it in this way, the approach is to make the AI itself capable of learning about these biases in order to correct them automatically.
Personally, I admit that I’m not sure if I find it fascinating or if it scares me, so it’s probably a bit of both.