Werd I/O

Simon Willison has a perfect name for unreviewed content that is shared with other people: "slop".

He goes on:

"I’m happy to use LLMs for all sorts of purposes, but I’m not going to use them to produce slop. I attach my name and stake my credibility on the things that I publish."

I think that's right. I'm less worried about using LLMs internally - as long as you understand that they're not impartial or perfectly factual sources, and as long as you take into account the methods used to generate the datasets that were used to train them. (Those are some big "if"s.)

But don't just take that output and share it with the public. And *certainly* don't do it so that you can publish content at scale without having to hire real writers. Not only is that not a good look, but you're going to harm your brand and your reputation in the process. #AI

[Link]

"Users who disagree with having their content scraped by ChatGPT are particularly outraged by Stack Overflow's rapid flip-flop on its policy concerning generative AI. For years, the site had a standing policy that prevented the use of generative AI in writing or rewording any questions or answers posted. Moderators were allowed and encouraged to use AI-detection software when reviewing posts."

This is all about money: "partnering" with OpenAI clearly means a significant sum has changed hands. The same thing may have happened at Valve, which also unblocked AI-generated art from its marketplace.

This feels like short-term thinking to me: while Stack will clearly make some near-term revenue through the deal, it comes at a cost to the health of its community, which is ultimately what drives the company's value. If motivated contributors drop off, the only thing left will be the AI-generated content - and there's no way that this will be as valuable over time.

I'd love to have been a fly on the wall of the boardroom where this deal was undoubtedly decided. What are they measuring that made this seem like a good idea - and what are they not measuring that means they're blind to the community dynamics that drive their actual sustainability? It's all fascinating to me. #AI

[Link]

"We found the company's phony authors and their work everywhere from celebrity gossip outlets like Hollywood Life and Us Weekly to venerable newspapers like the Los Angeles Times, the latter of which also told us that it had broken off its relationship with AdVon after finding its work unsatisfactory."

Even if the LA Times broke off its relationship because the work was unsatisfactory, the fact that this was attempted in the first place is unsettling. What if the work hadn't been "unsatisfactory"? What if it had been "good enough"?

It's not so much the technology itself as the intention behind it: to produce content at scale without employing human journalists, largely to generate pageviews in order to sell ads. There's no public service mission here, or even a mission to provide something that people might really want to read. It's all about arbitrage. #AI

[Link]

A detail I hadn't noticed: while the New York Times OpenAI lawsuit rested on copyright infringement, the Intercept, Raw Story, and AlterNet are claiming a DMCA violation.

"A study released this month by Patronus AI, a startup launched by former Meta researchers, found that GPT-4 reproduced copyrighted content at the highest rate among popular LLMs. When asked to finish a passage of a copyrighted novel, GPT-4 reproduced the text verbatim 60% of the time. The new lawsuits similarly allege that ChatGPT reproduces journalistic works near-verbatim when prompted." #AI

[Link]

"Sure, our articles maintain a rigid SEO template that creatively resembles the kitchen at a poorly run Quiznos, and granted, all our story ideas are gleaned from better-written magazine articles from seven months ago (that we’re totally not plagiarizing), but imagine if AI wrote those articles? So much would be lost."

Touché. #AI

[Link]

"Researchers have discovered a new way to hack AI assistants that uses a surprisingly old-school method: ASCII art."

So many LLM exploits come down to finding ways to convince an engine to disregard its own programming. It's straight out of 1980s science fiction, like teaching an android to lie. To be successful, you have to understand how LLMs "think", and then exploit that.

This one in particular is so much fun. By telling it to interpret an ASCII representation of a word and keep the meaning in memory without saying it out loud, front-line harm mitigations can be bypassed. It's like a magic spell. #AI

[Link]

Europe once again leads the way by passing meaningful AI regulation. Banned unacceptable-risk uses of AI include facial recognition, social scoring, and emotion recognition at schools and workplaces.

"The use of real-time facial recognition systems by law enforcement is permitted “in exhaustively listed and narrowly defined situations,” when the geographic area and the length of deployment are constrained."

I'm all in favor of these changes, but it's a little bit sad that this sort of regulation is always left up to the EU. American regulators appear to be sleeping. #AI

[Link]

"What I thought would be helpful, instead, is to survey the current state of AI-powered journalism, from the very bad to really good, and try to draw some lessons from those examples. I'm only speaking for myself today, but this certainly reflects how I'm thinking about the role AI could play in The Times newsroom and beyond."

A pretty good roundup, including the mistakes, folks using AI for pattern-recognition, and newsrooms that are actually using generative models. #AI

[Link]

I can't share a quote from this one without ruining it. But you should go read it. #AI

[Link]

"Users mistake decreasing levels of overt prejudice for a sign that racism in LLMs has been solved, when LLMs are in fact reaching increasing levels of covert prejudice."

Or to put it another way: AI is wildly racist. Although it has been trained to be less overtly so, it is now covertly discriminatory. For example, if it analyzes text written in AAE rather than Standardized American English, it is more likely to assign the death penalty, penalize job applicants, and so on. #AI

[Link]

A wonderful playlist from Ethan Marcotte about the state and context of AI and its implications for labor and society. Every quote is a gem; in aggregate it's a strong argument about where we are headed. #AI

[Link]

"We are professional editors, with extensive experience in the Australian book publishing industry, who wanted to know how ChatGPT would perform when compared to a human editor. To find out, we decided to ask it to edit a short story that had already been worked on by human editors – and we compared the results."

No surprise: ChatGPT stinks at this. I've sometimes used it to look at my own work and suggest changes. I'm not about to suggest that any of my writing is particularly literary, but its recommendations have always been generic at best.

Not that anyone in any industry, let alone one whose main product is writing of any sort, would try and use AI to make editing or content suggestions, right? Right?

... Right? #AI

[Link]

"Local news publishers, [VP Platforms at The Boston Globe] Karolian told Engadget, almost entirely depend on selling ads and subscriptions to readers who visit their websites to survive. “When tech platforms come along and disintermediate that experience without any regard for the impact it could have, it is deeply disappointing.”"

There's an interesting point that Josh Miller makes here about how the way the web gets monetized needs to change. Sure, but that's a lot like the people who say that open source funding will be solved by universal basic income: perhaps, at some future date, but that doesn't solve the immediate problem.

Do browser vendors have a responsibility to be good stewards for publishers? I don't know about that in itself. I'm okay with them freely innovating - but they also need to respect the rights of the content they're innovating with.

Micropayments emphatically don't work, but I do wonder if there's a way forward here (alongside other ways) where AI summarizers pay for access to the articles they're consuming as references, or otherwise participate in their business models somehow. #AI

[Link]

"The FCC announced the unanimous adoption of a Declaratory Ruling that recognizes calls made with AI-generated voices are "artificial" under the Telephone Consumer Protection Act (TCPA)."

A sign of the times that the FCC had to rule that making an artificial intelligence clone of a voice was illegal. I'm curious to understand if this affects commercial services that intentionally use AI to make calls on a user's behalf (eg to book a restaurant or perform some other service). #AI

[Link]

"Computer - enhance!"

I like the approach in this release from Apple: an open source AI model that can edit images based on natural language instructions. In other words, a human can tell the engine what to do to an image, and it goes and does it.

Rather than eliminating the human creativity in the equation, it gives the person doing the photo editing superpowers: instead of needing to know how to use a particular application to do the editing, they can simply give the machine instructions. I feel much more comfortable with the balance of power here than with most AI applications.

Obviously, it has implications for vendors like Adobe, which have established some degree of lock-in by forcing users to learn their tools and interfaces. If this kind of user interface takes off - and, given new kinds of devices like Apple Vision Pro, it inevitably will - they'll have to compete on capabilities alone. I'm okay with that. #AI

[Link]

During Meta's earnings call, Mark Zuckerberg said that Facebook and Instagram data is used to train the company's AI models.

“On Facebook and Instagram, there are hundreds of billions of publicly shared images and tens of billions of public videos, which we estimate is greater than the Common Crawl dataset and people share large numbers of public text posts in comments across our services as well.”

He's playing to win: one unstated competitive advantage is that Meta actually has the legal right to use training data generated on its own services. It's probably not something most users are aware of, but by posting content there, they grant the company rights to use it. If OpenAI falls afoul of copyright law, Meta's tech has a path forward.

It's a jarring thought, though. I'm certainly not keen on a generative model being trained on my son's face, for example. I'm curious how many users will feel the same way. #AI

[Link]

"OpenAI’s GPT-4 only gave people a slight advantage over the regular internet when it came to researching bioweapons, according to a study the company conducted itself." Uh, great?

"On top of that, the students who used GPT-4 were nearly as proficient as the expert group on some of the tasks. The researchers also noticed that GPT-4 brought the student cohort’s answers up to the “expert’s baseline” for two of the tasks in particular: magnification and formulation." Um, splendid?

"However, the study’s authors later state in a footnote that, overall, GPT-4 gave all participants a “statistically significant” advantage in total accuracy." Ah, superb? #AI

[Link]

"It should be obvious that any technology prone to making up facts is a bad fit for journalism, but the Associated Press, the American Journalism Project, and Axel Springer have all inked partnerships with OpenAI."

The conversation about AI at the Online News Association conference last year was so jarring to me that I was angry about it for a month. As Tyler Fisher says here, it presents existential risk to the news industry - and beyond that, following a FOMO-driven hype cycle rather than building things based on what your community actually needs is a recipe for failure.

As Tyler says: "Instead of trying to compete, journalism must reject the scale-driven paradigm in favor of deeper connection and community." This is the only real path forward for journalism. Honestly, it's the only real path forward for the web, and for a great many industries that live on it. #AI

[Link]

Simon Willison called this, and it makes sense: the George Carlin AI special was human-written, because that's the only way it could possibly have happened.

It's a parlor trick; a bit. It's also a kind of advertising for AI: even as you're horrified at the idea of creating a kind of resurrected George Carlin against his will, you've accepted that idea that it was technically possibly. It isn't.

Unfortunately for the folks behind the special, it's still harmful to Carlin's legacy, and putting his name on it in order to gain attention is still a problem. We'll see how the lawsuit shakes out. #AI

[Link]

"In order to combat the fracturing of social media platforms, a Google discoverability crisis fueled by AI generated spam and AI-fueled SEO, and a media business environment that is in utter freefall, we need to be able to reach our readers directly using a platform that we own and control."

For every publisher right now, email seems to be the only option. This is the first time I've seen this argument about AI scraping: usually the need to own your own relationship comes down to avoiding the thrash of different social media business models, which I've written about plenty of times before.

This idea that putting your content out there for free will only lead to it being rewritten by AI and repurposed by spam blogs could be the death of the open web. This is particularly true in light of Google's apparent refusal to downgrade machine-written content.

The idea is simple and awful: these spam sites rewrite human-written articles in an effort to capture search engine clicks themselves, instead of the people they stole from. They run ads against this spam. Because it's all machine-written, they can do it at scale.

Even if you don't agree that the web needs to be intrinsically protected (hi, we're enemies now), it seems obvious to me that incentives should be aligned towards publishing unique, useful information rather than superficially grabbing clicks through AI-driven SEO spam. I don't know what's going on inside the search engine businesses, but they need to consider what's going to be good for their businesses in the long term. This isn't it. #AI

[Link]

"In 2018, the indie women’s website The Hairpin stopped publishing, along with its sister site The Awl. This year, The Hairpin has been Frankensteined back into existence and stuffed with slapdash AI-generated articles designed to attract search engine traffic."

This is one of the worst kinds of AI-generated spam: a real, much-missed website has been purchased and spun into an LLM fever dream. It's now just a part of a Serbian DJ's thousands-deep portfolio of spam sites.

But the point made in the article about succession planning is really important. Media properties should be thoughtful about what happens to their domains once they've outlived their usefulness - even if the owner has shuttered completely. Otherwise anyone can scoop up the domain and abuse the goodwill built by its former owner for any purpose they like.

This is particularly true for journalism publishers. I recommend that they never let their domains expire for this reason, even if they've fully fallen out of use. You never know who might pick them up and abuse the trust of their community. #AI

[Link]

A robocall used a deepfake of Joe Biden's voice to encourage New Hampshire voters to stay home. "It's important that you save your vote for the November election."

It's not a perfect deepfake, but it doesn't necessarily need to be - for call recipients who don't understand what's happening, it has the potential to be enough to move the needle.

It's not clear that this is the first time that this has happened, but it certainly won't be the last. It's also not clear how this might be prevented except to block robocalls entirely (and even then, one can imagine using a live agent with a deepfaked voice, so that every call would be different). #AI

[Link]

"They just take it. Whatever they want." A poignant and infuriating reflection on generative AI, from the creator of Cat and Girl. #AI

[Link]

I like this argument that generated content is an invasive species in our content ecosystem.

"As generated material rapaciously populates the Internet, human-created artworks will be outcompeted by generated graphics on social media platforms by virtue of volume."

I agree that this is something to be concerned with, and the paragraph about legal rights and obligations is also spot on. #AI

[Link]

There are some jaw-dropping infringements here, including an image where DALL-E apparently copies the entire Pixar universe from the single two-word prompt, "animated toys".

It's impossible to hand-wave this away. Even if you don't think the New York Times case has merit, it's pretty obvious that generative AI can infringe copyright even when you don't ask it to, and without notifying the user. As noted in the references, it's a big ask to then push liability for infringement to the user. It's inherent to the engines.

As the author notes: "My guess is that none of this can easily be fixed." Indeed. #AI

[Link]