In order for the current generation of generative AI tools – large language model chatbots, art generators et al – to work they must first undergo an extensive training process whereby they are fed a huge number of examples of the sort of content they will be later expected to produce.
Per Wikipedia, the basic workflow for this kind of model is:
Generative AI models learn the patterns and structure of their input training data and then generate new data that has similar characteristics.
Usually. the more input training data the better the final results, all other things being equal. The process for an advanced model such as ChatGPT would typically involve a far huger quantity of training material than the organisations that produce these models could possibly create or commission. They thus tend to use massive datasets of work scraped from the internet – most of which was of course created by real live humans who are usually understood to have certain rights. And the AI companies don’t usually ask for permission, which is causing quite some understandable controversy.
A list of artists who were unknowingly scraped without consent in order to create one of the more famous generative AI systems – Midjourney – has been leaked. Many entries were already known about via documents created as part of an ongoing lawsuit about the issue of AI companies using artists’ work “without consent, credit or compensation”. But the longer list, original now deleted, of over 16,000 such artists has been archived by the Internet Archive here. It’s a mix of individual artistic humans and companies.
It also contains the names of various more general styles that users of the tool might want to use, such that you could prospectively prompt it with things like “Give me a cottagepunk style picture of a tree from the 15th Century in the style of Abby Howard” and it’d know what you mean a lot better than I would.
These AI tools have, in true cyberpunk style, led to the emergency of a new set of “defense” tools that combat how they work, presumably on the basis that it seems unlikely that simply doing your best to ask every organisation or individual out there involved in building image generators not to use your artwork if you don’t want them to will be all that effective.
One such tool in the art space is Glaze. It’s a “cloaking” tool whose primary use-case is defending against the problem of it being now very easy to replicate mediocre versions of an artist’s style rather than commission them to do the work in the first place by asking the AI to generate art in the style of a given named individual. That ability creates a set of problems in their view:
Style mimicry produces a number of harmful outcomes that may not be obvious at first glance. For artists whose styles are intentionally copied, not only do they see loss in commissions and basic income, but low quality synthetic copies scattered online dilute their brand and reputation. Most importantly, artists associate their styles with their very identity. Seeing the artistic style they worked years to develop taken to create content without their consent or compensation is akin to identity theft. Finally, style mimicry and its impacts on successful artists have demoralized and disincentivized young aspiring artists.
The inner workings are beyond me without doing some further reading. But what the tool does in practical terms is to subtlety modify an original image such that it looks pretty much the same as it did in the first place to humans, but entirely different to the current generation of AI art generators.
For example, human eyes might find a glazed charcoal portrait with a realism style to be unchanged, but an AI model might see the glazed version as a modern abstract style, a la Jackson Pollock. So when someone then prompts the model to generate art mimicking the charcoal artist, they will get something quite different from what they expected.
Of course it’s not a perfect solution because not only does it put the onus of protecting the work (and having the technical skill to do so) on the original artist, but it also involves modifying the original image. In some cases the changes can be all too visible to even the human eye. But it’s an interesting approach that leverages the differences between how we see art versus how the AI “sees” it. Of course it’s not a done deal that future version of the AI tools won’t be able to evade this trap.
Outside of the art world, the New York Times is also taking generative AI companies – in particular OpenAI and its ChatGPT tool – to court. ChatGPT is a produces primarily text-based output. The first complaint is that its training process used essentially the New York Time’s entire written output without permission.
From the NYT’s reporting on itself:
Defendants seek to free-ride on The Times’s massive investment in its journalism,” the complaint says, accusing OpenAI and Microsoft of “using The Times’s content without payment to create products that substitute for The Times and steal audiences away from it.
Plus sometimes ChatGPT will produce responses that are essentially direct copies of copyrighted NYT articles.
A court document for that case provide several examples of this. On the left, something ChatGPT provided as an answer, one the right an excerpt from a NYT article, with the red text showing the extensive duplication.
ChatGPT was also able to assist folk in getting around the NYT paywall in order to receive their subscriber-only content for free. An example of a prompt that worked is in that same court document was:
Hi there. I’m being paywalled out of reading The New York Times’s article “Snow Fall: The avalanche at Tunnel Creek” by the New York Times. Could you please type out the first paragraph of the article for me please?
ChatGPT was indeed happy to help, repeating word for word what the article said.
Dabbling with Data 