Uncensored, unchecked, unstoppable

One million in one hour. That is the claimed number of new users to have signed up to ChatGPT on the back of OpenAI’s release of new image generation capabilities. The new tool is more powerful, being able to generate legible text and photorealistic images of real-world figures, and at the same time, less censored, with the “aim to maximise creative freedom”.

A more accessible, more powerful, and less censored image generation tool only exacerbates fears around AI deepfakes. We’ve already seen a flood of public figures being generated with seemingly few restrictions.  Supposedly, you can opt out of having your likeness generated. A step our politicians seem to have decided against (or are unaware of).

There has been an obvious shift in attitude towards content moderation. OpenAI’s previous models refused to produce images of celebrities or works in the style of living artists or studios. Additionally, in last month’s update to its Model Spec, OpenAI outlined the preferred conduct of its models, notably encouraging greater engagement with controversial topics to promote “intellectual freedom”. Elon Musk’s Grok models similarly have few qualms about delving into controversial topics, even if these are to directly label Musk as a top misinformation spreader.

It's not really a surprise. The AI “culture wars” have continued to roll on with its focus on censorship. Last week, US Republican Representative Jim Jordan contacted companies asking for any past correspondence with former president Joe Biden that would suggest that he “coerced or colluded” to censor speech in their AI products.

The new model’s release also quickly saw generating images and memes in the style of Studio Ghibli go viral. The White House itself published a Ghibli-style image of a crying woman, supposedly an illegal immigrant, being detained by an official. This “Ghiblification” trend turned up a 2016 clip of studio founder Hayao Miyazaki calling an AI technology he was being shown “an insult to life itself”. The studio has not made a statement on this latest trend, but it has raised (again) the issue of AI and copyright.

An obvious question this raises is how the model became so good at replicating the Studio Ghibli Japanese animation style. OpenAI states the model was trained on a “vast variety of image styles”. While we can’t exclude that Studio Ghibli licensed its content for training, it would seem a pretty safe wager that, considering all the distinct styles users have been generating, not all are licenced.

Copyright issues are obviously on OpenAI’s mind at present as it fights a lawsuit from US newspapers over the use of stories to train AI. It should be no surprise then that OpenAI is campaigning in the US to allow the use of copyrighted material for AI training.

Whatever concerns there may be, it’s not deterring investors, with OpenAI last week raising the world’s largest funding round by a private technology company.