Daily Archives: October 20, 2024

German court: LAION’s generative AI training dataset is legal thanks to EU copyright exceptions

Posted on by

The copyright world is currently trying to assert its control over the new world of generative AI through a number of lawsuits, several of which have been discussed previously on Walled Culture. We now have our first decision in this area, from the regional court in Hamburg. Andres Guadamuz has provided an excellent detailed analysis of a ruling that is important for the German judges’ discussion of how EU copyright law applies to various aspects of generative AI. The case concerns the freely-available dataset from LAION (Large-scale Artificial Intelligence Open Network), a German non-profit. As the LAION FAQ says: “LAION datasets are simply indexes to the internet, i.e. lists of URLs to the original images together with the ALT texts found linked to those images.” Guadamuz explains:

The case was brought by German photographer Robert Kneschke, who found that some of his photographs had been included in the LAION dataset. He requested the images to be removed, but LAION argued that they had no images, only links to where the images could be found online. Kneschke argued that the process of collecting the dataset had included making copies of the images to extract information, and that this amounted to copyright infringement.

LAION admitted making copies, but said that it was in compliance with the exception for text and data mining (TDM) present in German law, which is a transposition of Article 3 of the 2019 EU Copyright Directive. The German judges agreed:

The court argued that while LAION had been used by commercial organisations, the dataset itself had been released to the public free of charge, and no evidence was presented that any commercial body had control over its operations. Therefore, the dataset is non-commercial and for scientific research. So LAION’s actions are covered by section 60d of the German Copyright Act

That’s good news for LAION and its dataset, but perhaps more interesting for the general field of generative AI is the court’s discussion of how the EU Copyright Directive and its exceptions apply to AI training. It’s a key question because copyright companies claim that they don’t, and that when such training involves copyright material, permission is needed to use it. Guadamuz summarises that point of view as follows:

the argument is that the legislators didn’t intend to cover generative AI when they passed the [EU Copyright Directive], so text and data mining does not cover the training of a model, just the making of a copy to extract information from it. The argument is that making a copy to extract information to create a dataset is fine, as the court agreed here, but the making of a copy in order to extract information to make a model is not. I somehow think that this completely misses the way in which a model is trained; a dataset can have copies of a work, or in the case of LAION, links to the copies of the work. A trained model doesn’t contain copies of the works with which it was trained, and regurgitation of works in the training data in an output is another legal issue entirely.

The judgment from the Hamburg court says that while legislators may not have been aware of generative AI model training in 2019, when they drew up the EU Copyright Directive, they certainly are now. The judges use the EU’s 2024 AI Act as evidence of this, citing a paragraph that makes explicit reference to AI models complying with the text and data mining regulation in the earlier Copyright Directive.

As Guadamuz writes in his post, this is an important point, but the legal impact may be limited. The judgment is only the view of a local German court, so other jurisdictions may produce different results. Moreover, the original plaintiff Robert Kneschke may appeal and overturn the decision. Furthermore, the ruling only concerns the use of text and data mining to create a training dataset, not the actual training itself, although the judges’ thoughts on the latter indicate that it would be legal too. In other words, this local outbreak of good sense in Germany is welcome, but we are still a long way from complete legal clarity on the training of generative AI systems on copyright material.

Source: German court: LAION’s generative AI training dataset is legal thanks to EU copyright exceptions – Walled Culture

Penguin Random House is adding an AI warning to its books’ copyright pages fwiw

Posted on by

Penguin Random House, the trade publisher, is adding language to the copyright pages of its books to prohibit the use of those books to train AI.

The Bookseller reports that new books and reprints of older titles from the publisher will now include the statement, “No part of this book may be used or reproduced in any manner for the purpose of training artificial intelligence technologies or systems.”

While the use of copyrighted material to train AI models is currently being fought over in multiple lawsuits, Penguin Random House appears to be the first major publisher to update its copyright pages to reflect these new concerns.

The update doesn’t mean Penguin Random House is completely opposed to the use of AI in book publishing. In August, it outlined an initial approach to generative AI, saying it will “vigorously defend the intellectual property that belongs to our authors and artists” while also promising to “use generative AI tools selectively and responsibly, where we see a clear case that they can advance our goals.”

Source: Penguin Random House is adding an AI warning to its books’ copyright pages | TechCrunch

Penguin spins it in support of authors, but the whole copyright thing only really fills the pockets of the publishers (eg. Juicy licensing deals with AI companies show that publishers don’t really care about creators). This will probably not hold up in court.

You Don’t Need Words to Think

Posted on by

Scholars have long contemplated the connection between language and thought—and to what degree the two are intertwined—by asking whether language is somehow an essential prerequisite for thinking.

[…]

Evelina Fedorenko, a neuroscientist who studies language at the McGovern Institute for Brain Research at the Massachusetts Institute of Technology, has spent many years trying to answer these questions. She remembers being a Harvard University undergraduate in the early 2000s, when the language-begets-thought hypothesis was still highly prominent in academia.

[…]

She recently co-authored a perspective article in Nature that includes a summary of her findings over the ensuing years. It makes clear that the jury is no longer out, in Fedorenko’s view: language and thought are, in fact, distinct entities that the brain processes separately. The highest levels of cognition—from novel problem-solving to social reasoning—can proceed without an assist from words or linguistic structures.

[…]

Language works a little like telepathy in allowing us to communicate our thoughts to others and to pass to the next generation the knowledge and skills essential for our hypersocial species to flourish. But at the same time, a person with aphasia, who are sometimes unable to utter a single word, can still engage in an array of cognitive tasks fundamental to thought. Scientific American talked to Fedorenko about the language-thought divide and the prospects of artificial intelligence tools such as large language models for continuing to explore interactions between thinking and speaking.

[…]

What evidence did you find that thought and language are separate systems?

The evidence comes from two separate methods. One is basically a very old method that scientists have been using for centuries: looking at deficits in different abilities—for instance, in people with brain damage.

Using this approach, we can look at individuals who have impairments in language—some form of aphasia. […] You can ask whether people who have these severe language impairments can perform tasks that require thinking. You can ask them to solve some math problems or to perform a social reasoning test, and all of the instructions, of course, have to be nonverbal because they can’t understand linguistic information anymore. Scientists have a lot of experience working with populations that don’t have language—studying preverbal infants or studying nonhuman animal species. So it’s definitely possible to convey instructions in a way that’s nonverbal. And the key finding from this line of work is that there are people with severe language impairments who nonetheless seem totally fine on all cognitive tasks that we’ve tested them on so far.

[…]

A nicely complementary approach, which started in the 1980s and 1990s, is a brain-imaging approach. We can measure blood flow changes when people engage in different tasks and ask questions about whether the two systems are distinct or overlapping—for example, whether your language regions overlap with regions that help you solve math problems. These brain-imaging tools are really good for these questions. But before I could ask these questions, I needed a way to robustly and reliably identify language areas in individual brains, so I spent the first bunch of years of my career developing tools to do this.

And once we have a way of finding these language regions, and we know that these are the regions that, when damaged in adulthood, lead to conditions such as aphasia, we can then ask whether these language regions are active when people engage in various thinking tasks. So you can come into the lab, and I can put you in the scanner, find your language regions by asking you to perform a short task that takes a few minutes—and then I can ask you to do some logic puzzles or sudoku or some complex working memory tasks or planning and decision-making. And then I can ask whether the regions that we know process language are working when you’re engaging in these other kinds of tasks. There are now dozens of studies that we’ve done looking at all sorts of nonlinguistic inputs and tasks, including many thinking tasks. We find time and again that the language regions are basically silent when people engage in these thinking activities.

[…]

Do the language and thinking systems interact with each other?

There aren’t great tools in neuroscience to study intersystem interactions between language and thought. But there are interesting new opportunities that are opening up with advances in AI where we now have a model system to study language, which is in the form of these large language models such as GPT-2 and its successors. These models do language really well, producing perfectly grammatical and meaningful sentences. They’re not so good at thinking, which is nicely aligning with the idea that the language system by itself is not what makes you think.

But we and many other groups are doing work in which we take some version of an artificial neural network language model as a model of the human language system. And then we try to connect it to some system that is more like what we think human systems of thought look like—for example, a symbolic problem-solving system such as a math app. With these artificial intelligence tools, we can at least ask, “What are the ways in which a system of thought, a system of reasoning, can interact with a system that stores and uses linguistic representations?” These so-called neurosymbolic approaches provide an exciting opportunity to start tackling these questions.

So what do large language models do to help us understand the neuroscience of how language works?

They’re basically the first model organism for researchers studying the neuroscience of language. They are not a biological organism, but until these models came about, we just didn’t have anything other than the human brain that does language. And so what’s happening is incredibly exciting. You can do stuff on models that you can’t do on actual biological systems that you’re trying to understand. There are many, many questions that we can now ask that had been totally out of reach: for example, questions about development.

In humans, of course, you cannot manipulate linguistic input that children get. You cannot deprive kids of language, or restrict their input in some way, and see how they develop. But you can build these models that are trained on only particular kinds of linguistic input or are trained on speech inputs as opposed to textual inputs. And then you can see whether models trained in particular ways better recapitulate what we see in humans with respect to their linguistic behavior or brain responses to language.

So just as neuroscientists have long used a mouse or a macaque as a model organism, we can now use these in silico models, which are not biological but very powerful in their own way, to try to understand some aspects of how language develops or is processed or decays in aging or whatnot.

We have a lot more access to these models’ internals. The methods we have for messing with the brain, at least with the human brain, are much more limited compared with what we can do with these models.

Source: You Don’t Need Words to Think | Scientific American

New 3 point graph mining algorithm finds patterns in complex networks

Posted on by

University of Virginia School of Engineering and Applied Science professor Nikolaos Sidiropoulos has introduced a breakthrough in graph mining with the development of a new computational algorithm.

Graph mining, a method of analyzing networks like social media connections or biological systems, helps researchers discover meaningful patterns in how different elements interact. The new algorithm addresses the long-standing challenge of finding tightly connected clusters, known as triangle-dense subgraphs, within large networks — a problem that is critical in fields such as fraud detection, computational biology and data analysis.

The research, published in IEEE Transactions on Knowledge and Data Engineering, was a collaboration led by Aritra Konar, an assistant professor of electrical engineering at KU Leuven in Belgium who was previously a research scientist at UVA.

Graph mining algorithms typically focus on finding dense connections between individual pairs of points, such as two people who frequently communicate on social media. However, the researchers’ new method, known as the Triangle-Densest-k-Subgraph problem, goes a step further by looking at triangles of connections — groups of three points where each pair is linked. This approach captures more tightly knit relationships, like small groups of friends who all interact with each other, or clusters of genes that work together in biological processes.

“Our method doesn’t just look at single connections but considers how groups of three elements interact, which is crucial for understanding more complex networks,” explained Sidiropoulos, a professor in the Department of Electrical and Computer Engineering. “This allows us to find more meaningful patterns, even in massive datasets.”

Finding triangle-dense subgraphs is especially challenging because it’s difficult to solve efficiently with traditional methods. But the new algorithm uses what’s called submodular relaxation, a clever shortcut that simplifies the problem just enough to make it quicker to solve without losing important details.

This breakthrough opens new possibilities for understanding complex systems that rely on these deeper, multi-connection relationships. Locating subgroups and patterns could help uncover suspicious activity in fraud, identify community dynamics on social media, or help researchers analyze protein interactions or genetic relationships with greater precision.

Story Source:

Materials provided by University of Virginia School of Engineering and Applied Science. Note: Content may be edited for style and length.

Journal Reference:

  1. Aritra Konar, Nicholas D. Sidiropoulos. Mining Triangle-Dense Subgraphs of a Fixed Size: Hardness, Lovasz extension and ´ Applications. IEEE Transactions on Knowledge and Data Engineering, 2024; 1 DOI: 10.1109/TKDE.2024.3444608

Source: Professor tackles graph mining challenges with new algorithm | ScienceDaily

Research shows how corporate social responsibility messaging can backfire

Posted on by

It’s lately been considered good business for companies to show they are responsible corporate citizens. Google touts its solar-powered data centers. Apple talks about its use of recycled materials. Walmart describes its support for local communities.

But these narratives, according to new research by Haas Associate Professor Tim McQuade, have some downsides. With Emanuele Colonnelli and Niels Gormsen of the University of Chicago, McQuade demonstrates how positive corporate messaging can evoke negative associations among consumers, in turn nudging them away from policies that support corporations in times of crisis.

“Even if you frame information in a positive way, consumers with pre-existing negative beliefs regarding might draw up mostly negative experiences from memory,” McQuade says. “In this manner, the messaging can do the opposite of what’s intended.”

Their results were published in The Review of Economic Studies.

Working with faulty memory

These results hinge on an updated model of how consumers call information to mind when making decisions. Traditionally, economists assumed consumers to be rational actors sifting through all the relevant knowledge they have when making a decision. McQuade and his colleagues draw on a more recent understanding of cognition in which people have limited recall—meaning they generally only draw on a limited set of information to make decisions—and in which specific cues can influence what information they use.

Much advertising relies on this premise. For instance, if people are cued with the old Snickers tagline, “Hungry? Why wait,” they may buy the candy simply because they are prompted to think about their hunger and not consider whether they need the calories or could better spend money on something else.

With this picture of consumer psychology in place, the researchers recruited nearly 7,000 participants to complete a four-part survey. The survey took place in May 2020, when many companies were struggling under pandemic restrictions and the federal government was discussing the possibility of bailouts.

A landscape of ‘big business discontent’

The first portion of the survey asked basic questions about socioeconomic background. The second contained four different animated videos—three of which were used to cue distinct patterns of thought, and one used to create a control group.

The watched a video detailing basic instructions to complete the survey along with definitions of concepts like “corporate ” and “stakeholders;” the rest of the videos started with this control segment but included additional content. One framed big companies as relatively bad citizens—polluting, overpaying executives, underinvesting in communities, and so forth. The second video framed them as good citizens. The third mentioned nothing of corporate citizenship but talked instead about the economic stability provided by corporate bailouts.

After participants watched one of these four videos, they were asked the degree to which they thought large companies were doing what they should when it comes to environmental, social, and governance (ESG) goals. Another section asked participants how strongly they supported economic bailouts for large corporations. (The ordering of sections three and four varied randomly.)

The raw results from this survey found that people have an overwhelmingly negative view of corporate citizenship. “Our first key contribution showed that on a variety of dimensions, there is this broad perception in society that corporations are not doing what people think they should be doing,” McQuade says. “We call this ‘big business discontent,’ and it becomes a necessary condition for what we find next.”

How positive messaging elicits negative associations

The researchers looked next at for bailouts.

They found that survey participants who were cued by videos to think about —whether the video framed this work positively or negatively—expressed much lower support for corporate bailouts than those who watched the video about stabilizing the economy. In fact, those who watched the video framing companies’ ESG efforts positively expressed lower support for bailouts than those who simply watched the control video.

“When we primed people to think about these policies through a corporate social responsibility lens, even when we put that work in a positive light, the fact that there is this pre-existing big business discontent meant that the messaging backfired relative to giving them no information at all,” McQuade says. “Because recall is imperfect, the positive framing still brings to mind negative experiences,” such as the Enron accounting scandal, various environmental disasters, or poor wages.

This effect was even stronger among the survey participants who were asked how well they thought companies were doing on ESG goals before being asked their level of support for bailouts. This particular ordering of questions, it seems, dredged up more negative memories. Lack of support for bailouts was also strongest among young people and liberals, who expressed the highest levels of big business discontent.

Finding a message that works

Survey participants who were instead shown a video discussing how bailouts contributed to economic stability expressed support for the policy. In other words, the topic that people are cued to consider—in this case ESG goals versus economic health—significantly influenced their policy preferences.

The implications extend beyond corporate messaging into all realms of influence or persuasion. As McQuade notes, groups often try to update people’s beliefs by providing positive information on some policy or action. Companies talk about their good citizenship; politicians talk about their achievements.

“But if the domain or topic they’re talking about is one that many people have negative views on, then it is probably not the most effective way to gather support, since the framing effect could outweigh any positive PR effects of the communication,” he says. “Rather, they might want to refocus attention on some other policy domain. This insight shifts the way we think about optimal communication and optimal messaging.”

More information: Emanuele Colonnelli et al, Selfish Corporations, Review of Economic Studies (2023). DOI: 10.1093/restud/rdad057

Provided by University of California, Berkeley Haas School of Business

Source: Research shows how corporate social responsibility messaging can backfire