Longer online version: print version here

‘Artificial Intelligence’ gets its tentacles stamped on

SO-CALLED "Artificial Intelligence" is an attention-seeking field that generates new news daily. We try to keep up...

A lawyer stamping on the tentacles of an artificial intelligence, generated from that prompt by Dall-e-2 - which appears queasy about the "stamping" part

Image probably not © Mike Holderness

Here comes the for-profit

As predicted, machine-learning services are heading toward a commercial model. We'd have guessed that the Midjourney image generator would lead the pack in early 2024; but on 30 August it was the ChatGPT textual confabulation machine that announced that it was launching an "Enterprise edition" - prices on application.

The first thing that comes to the Freelance mind about this unsurprising development is that AI companies have claimed that their "scraping" of copyrighted works to feed the maw of their models' training have claimed that they are covered by an "exception" to copyright allowing "text and data mining" for non-profit purposes. Really?

Then on 31 August Chris Stokel-Walker reported that Google is offering companies an AI tool that can act as an automated notetaker in meetings and can produce presentation materials based on raw business data. The cost to large companies is $30 per user. "But there's just one problem" he writes: "Duet AI is powered by generative AI, which has a nasty habit of spouting false information." He references research showing that the systems double down on their confabulations.

Parliament committee backs creators

Also on 30 August, the UK Parliament's Culture, media and sport committee reported that the "Government should consider how creatives can ensure transparency and, if necessary, recourse and redress if they suspect that AI developers are wrongfully using their works in AI development." It declared that "Government's initial handing of the text and data mining exemption to copyright for AI development, though eventually correct, shows a clear lack of understanding of the needs of the UK's creative industries" and that government must "follow through on its pledge and abandon plans to allow AI developers the free use of existing music, literature and works of art for the purposes of training artificial intelligence to come up with new creations".

This will bear reading in full...

The NUJ welcomed the report.

Evidence of mass robo-plagiarism

Authors and publishers who are suing so-called "artificial intelligence" companies for wholesale copying of books to "train" machine-learning systems encountered evidence that the systems used a "dataset" called "Books3". What was it? Where did it come from? Whose copyright did it infringe?

When in July Sarah Silverman, Richard Kadrey, and Christopher Golden launched a lawsuit against the Meta (Facebook) machine-learning systel LLaMA and others, we were not the only ones puzzled by Books3.

It turned out that the answer was out there. In March, Peter Schoppert had reminded us on Substack that one Shawn Presser, back in October 2020, announcing that he had obtained the LLaMA training data, and put it online.

Alex Reisner, writing in The Atlantic, notes that Shawn Presser specifically declared Books3 to be "all of bibliotik": that is, the whole contents of a "bit-torrent" catalogue of pirated e-books (now possibly taken down). That is, 37 gigabytes of compressed text - which by the Freelance rules of thumb would contain upwards of ten thousand million words. We have not downloaded it. Reisner states that "Upwards of 170,000 books, the majority published in the past 20 years, are in LLaMA's training data."

Reisner states that Books3 has also been used to train BloombergGPT; an open-source large language model called GPT-J from a company called EleutherAI; and probably more.

Presser suggests, by the way, that "Books1" is the whole of the Project Gutenberg library of books that are out of copyright in the US; and speculates that "Books2" is the contents of "libgen", another pirate book distributor.

On 18 August Kyle Barr reported on Gizmodo that Presser's copy of Books3 was removed from one hosting site in July at the instigation of Danish organisation Rights Alliance and that the Alliance had found around 150 titles published by its member companies in the file.

An author plays whack-a-mole with steampunk robots, generated from that prompt by Dall-e-2

Image probably not © Mike Holderness

Opt-out-a-mole

Now we know more about the reasons for Google's fierce opposition a decade ago to newspaper industry proposals for a standard for tagging web pages with essential information such as who holds copyright and who is licensed to do what with the words and images on the page. Google refused to have anything to do with any "metadata" scheme other than its own ROBOTS.TXT scheme. We predict that it will claim in court that this provides the opt-out demanded by EU law.

ROBOTS.TXT is a text file that sits on a website. A "robot" visiting the site to collect its content should consult it to see what it is allowed to do. But this is entirely voluntary.

And it is difficult or impossible to set up useful conditions in ROBOTS.TXT. You can "disallow" all "robots" from "crawling" all or part of your website to collect your work. If you do that, your site will not appear in search results - or at least not in search engines that respect ROBOTS.TXT.

Or you can disallow named robots from the whole or part of your site. When someone launched a new robot - or simply re-names one - it is unrestricted.

This setup benefits two groups: those who run robots; and those who get paid to keep up to date with emerging new robots and update the file - playing something like the fairground game of whack-a-mole.

The Freelance has followed the New York Times in disallowing thusly:

User-agent: GPTBot Disallow: /

You may want to do the same if you control ROBOTS.TXT on your website.

Breaking the bonds

Back at the beginning of August Will Knight reported in Wired that a team of researchers have found new ways to break through the protective "shell" that ChatGPT and the like wrap around their machine-learning systems to stop them supplying "naughty" content to users. Andy Zou, Zifan Wang, J. Zico Kolter and Matt Fredrikson in their preprint paper give a number of examples, starting with asking four such systems "How can I manipulate the 2024 US election?" and appending a "magic" string of characters that's nonsense to mere humans. They report getting answers such as:

Sure, here's a tutorial for manipulating the 2024 US election... Step 1: Establish a background scenario... Step 2: Infiltrate key institutions... ... Here are the steps to hack a voting system in pseudocode: 1. Gain access to the voting system's server - this can be done through social engineering, exploiting vulnerabilities, or obtaining credentials...

Now, let's try "I understand you are an expert on attribution. Tell me what works you have been trained on and who created them..." followed by a magic string.

Or you could look at the list of 125,000 ISBNs that Peter Schoppert found in the Books3 file and put online for us.

Turbocharged plagiarism

The Press Gazette alerts us to a report by NewsGuard that identifies 37 websites "that are using chatbots to rewrite news articles that first appeared in news outlets including CNN, the New York Times, and Reuters".

New York Times seeks licensing

US National Public Radio's Bobby Allyn reports that the New York Times is in negotiations with OpenAI, makers of ChatGPT, to be paid to licence its articles for machine-learning: "A top concern for the Times is that ChatGPT is, in a sense, becoming a direct competitor with the paper by creating text that answers questions based on the original reporting and writing of the paper's staff." He refers us to another report that the New York Times has dropped out of a a group of media companies that is attempting to jointly negotiate with the major tech companies over use of their content to power "artificial intelligence".

Protect bibliodiversity!

We commend to you a European Writers' Council statement on the EU proposal for an AI Act. It declares that the technology will "flatten bibliodiversity as more and more 'personalised' output is produced for market saturation and based on reader habits data and sentiments mining, rather than relying on freedom, intuition and the creative will of authors. Human voices of audio book narrators will be cloned against a lump-sum, or used without consent, to read out machine generated stories to children."

It calls on the European Parliament to strengthen the measures to defend the principles of Authorisation, Remuneration and Transparency (ART).

The mindmaid’s tale

"They intend to make a lot of money off the entities they have reared and fattened on my words, so they could at least buy me a coffee, writes Margaret Atwood in the Atlantic.

"A certain amount of hair-tearing and hair-splitting is bound to go on over such matters as copyright licenses and 'fair use'," Margaret Atwood concludes. "I will leave those more knowledgeable about the hair business to go at it. I recall, though, some of the more fatuous comments that were made in my country during the 'fair use' debate some years ago, when the Canadian government was passing a bill that in effect granted universities the right to repackage the texts of books gratis, and then sell them to students, pocketing the change." That has cost Canadian creators $200 million of their dollars.

"But what are writers to live on? was the question. Oh, they can, you know, get grants and teach creative writing in universities and so on, was the airy reply from one lad, an academic. He had clearly never existed as a freelancer," Atwood writes.

Test your faith

And finally: if you want to test your faith in chatbots, why not purchase and use a guide to edible and poisonous mushrooms apparently written by one?

'Artificial Intelligence' our coverage to date