The AI regulations that aren’t being talked about

Curated on Posted on by Stefaan Verhulst

Article by Deloitte: “…But our research shows that this focus may be overlooking some of the most important tools already on the books. Of the 1,600+ policies we analyzed, only 11% were focused on regulating AI-adjacent issues like data privacy, cybersecurity, intellectual property, and so on (Figure 5). Even when limiting the search to only regulations, 60% were focused directly on AI and only 40% on AI-adjacent issues (Figure 5). For example, several countries have data protection agencies with regulatory powers to help protect citizens’ data privacy. But while these agencies may not have AI or machine learning named specifically in their charters, the importance of data in training and using AI models makes them an important AI-adjacent tool.

This can be problematic because directly regulating a fast-moving technology like AI can be difficult. Take the hypothetical example of removing bias from home loan decisions. Regulators could accomplish this goal by mandating that AI should have certain types of training data to ensure that the models are representative and will not produce biased results, but such an approach can become outdated when new methods of training AI models emerge. Given the diversity of different types of AI models already in use, from recurrent neural networks to generative pretrained transformers to generative adversarial networks and more, finding a single set of rules that can deliver what the public desires both now, and in the future, may be a challenge…(More)”.

Researchers warn we could run out of data to train AI by 2026. What then?

Curated on Posted on by Stefaan Verhulst

Article by Rita Matulionyte: “As artificial intelligence (AI) reaches the peak of its popularity, researchers have warned the industry might be running out of training data – the fuel that runs powerful AI systems. This could slow down the growth of AI models, especially large language models, and may even alter the trajectory of the AI revolution.

But why is a potential lack of data an issue, considering how much there are on the web? And is there a way to address the risk?…

We need a lot of data to train powerful, accurate and high-quality AI algorithms. For instance, ChatGPT was trained on 570 gigabytes of text data, or about 300 billion words.

Similarly, the stable diffusion algorithm (which is behind many AI image-generating apps such as DALL-E, Lensa and Midjourney) was trained on the LIAON-5B dataset comprising of 5.8 billion image-text pairs. If an algorithm is trained on an insufficient amount of data, it will produce inaccurate or low-quality outputs.

The quality of the training data is also important…This is why AI developers seek out high-quality content such as text from books, online articles, scientific papers, Wikipedia, and certain filtered web content. The Google Assistant was trained on 11,000 romance novels taken from self-publishing site Smashwords to make it more conversational.

The AI industry has been training AI systems on ever-larger datasets, which is why we now have high-performing models such as ChatGPT or DALL-E 3. At the same time, research shows online data stocks are growing much slower than datasets used to train AI.

In a paper published last year, a group of researchers predicted we will run out of high-quality text data before 2026 if the current AI training trends continue. They also estimated low-quality language data will be exhausted sometime between 2030 and 2050, and low-quality image data between 2030 and 2060.

AI could contribute up to US$15.7 trillion (A$24.1 trillion) to the world economy by 2030, according to accounting and consulting group PwC. But running out of usable data could slow down its development…(More)”.

Chatbots May ‘Hallucinate’ More Often Than Many Realize

Curated on Posted on by Stefaan Verhulst

Cade Metz at The New York Times: “When the San Francisco start-up OpenAI unveiled its ChatGPT online chatbot late last year, millions were wowed by the humanlike way it answered questions, wrote poetry and discussed almost any topic. But most people were slow to realize that this new kind of chatbot often makes things up.

When Google introduced a similar chatbot several weeks later, it spewed nonsense about the James Webb telescope. The next day, Microsoft’s new Bing chatbot offered up all sorts of bogus information about the Gap, Mexican nightlife and the singer Billie Eilish. Then, in March, ChatGPT cited a half dozen fake court cases while writing a 10-page legal brief that a lawyer submitted to a federal judge in Manhattan.

Now a new start-up called Vectara, founded by former Google employees, is trying to figure out how often chatbots veer from the truth. The company’s research estimates that even in situations designed to prevent it from happening, chatbots invent information at least 3 percent of the time — and as high as 27 percent.

Experts call this chatbot behavior “hallucination.” It may not be a problem for people tinkering with chatbots on their personal computers, but it is a serious issue for anyone using this technology with court documents, medical information or sensitive business data.

Because these chatbots can respond to almost any request in an unlimited number of ways, there is no way of definitively determining how often they hallucinate. “You would have to look at all of the world’s information,” said Simon Hughes, the Vectara researcher who led the project…(More)”.

Assessing and Suing an Algorithm

Curated on Posted on by Stefaan Verhulst

Report by Elina Treyger, Jirka Taylor, Daniel Kim, and Maynard A. Holliday: “Artificial intelligence algorithms are permeating nearly every domain of human activity, including processes that make decisions about interests central to individual welfare and well-being. How do public perceptions of algorithmic decisionmaking in these domains compare with perceptions of traditional human decisionmaking? What kinds of judgments about the shortcomings of algorithmic decisionmaking processes underlie these perceptions? Will individuals be willing to hold algorithms accountable through legal channels for unfair, incorrect, or otherwise problematic decisions?

Answers to these questions matter at several levels. In a democratic society, a degree of public acceptance is needed for algorithms to become successfully integrated into decisionmaking processes. And public perceptions will shape how the harms and wrongs caused by algorithmic decisionmaking are handled. This report shares the results of a survey experiment designed to contribute to researchers’ understanding of how U.S. public perceptions are evolving in these respects in one high-stakes setting: decisions related to employment and unemployment…(More)”.

Can Large Language Models Capture Public Opinion about Global Warming? An Empirical Assessment of Algorithmic Fidelity and Bias

Curated on Posted on by Stefaan Verhulst

Paper by S. Lee et all: “Large language models (LLMs) have demonstrated their potential in social science research by emulating human perceptions and behaviors, a concept referred to as algorithmic fidelity. This study assesses the algorithmic fidelity and bias of LLMs by utilizing two nationally representative climate change surveys. The LLMs were conditioned on demographics and/or psychological covariates to simulate survey responses. The findings indicate that LLMs can effectively capture presidential voting behaviors but encounter challenges in accurately representing global warming perspectives when relevant covariates are not included. GPT-4 exhibits improved performance when conditioned on both demographics and covariates. However, disparities emerge in LLM estimations of the views of certain groups, with LLMs tending to underestimate worry about global warming among Black Americans. While highlighting the potential of LLMs to aid social science research, these results underscore the importance of meticulous conditioning, model selection, survey question format, and bias assessment when employing LLMs for survey simulation. Further investigation into prompt engineering and algorithm auditing is essential to harness the power of LLMs while addressing their inherent limitations…(More)”.

AI and Democracy’s Digital Identity Crisis

Curated on Posted on by Stefaan Verhulst

Essay by Shrey Jain, Connor Spelliscy, Samuel Vance-Law and Scott Moore: “AI-enabled tools have become sophisticated enough to allow a small number of individuals to run disinformation campaigns of an unprecedented scale. Privacy-preserving identity attestations can drastically reduce instances of impersonation and make disinformation easy to identify and potentially hinder. By understanding how identity attestations are positioned across the spectrum of decentralization, we can gain a better understanding of the costs and benefits of various attestations. In this paper, we discuss attestation types, including governmental, biometric, federated, and web of trust-based, and include examples such as e-Estonia, China’s social credit system, Worldcoin, OAuth, X (formerly Twitter), Gitcoin Passport, and EAS. We believe that the most resilient systems create an identity that evolves and is connected to a network of similarly evolving identities that verify one another. In this type of system, each entity contributes its respective credibility to the attestation process, creating a larger, more comprehensive set of attestations. We believe these systems could be the best approach to authenticating identity and protecting against some of the threats to democracy that AI can pose in the hands of malicious actors. However, governments will likely attempt to mitigate these risks by implementing centralized identity authentication systems; these centralized systems could themselves pose risks to the democratic processes they are built to defend. We therefore recommend that policymakers support the development of standards-setting organizations for identity, provide legal clarity for builders of decentralized tooling, and fund research critical to effective identity authentication systems…(More)”

The Bletchley Declaration

Curated on Posted on by Stefaan Verhulst

Declaration by Countries Attending the AI Safety Summit, 1-2 November 2023: “In the context of our cooperation, and to inform action at the national and international levels, our agenda for addressing frontier AI risk will focus on:

  • identifying AI safety risks of shared concern, building a shared scientific and evidence-based understanding of these risks, and sustaining that understanding as capabilities continue to increase, in the context of a wider global approach to understanding the impact of AI in our societies.
  • building respective risk-based policies across our countries to ensure safety in light of such risks, collaborating as appropriate while recognising our approaches may differ based on national circumstances and applicable legal frameworks. This includes, alongside increased transparency by private actors developing frontier AI capabilities, appropriate evaluation metrics, tools for safety testing, and developing relevant public sector capability and scientific research.

In furtherance of this agenda, we resolve to support an internationally inclusive network of scientific research on frontier AI safety that encompasses and complements existing and new multilateral, plurilateral and bilateral collaboration, including through existing international fora and other relevant initiatives, to facilitate the provision of the best science available for policy making and the public good.

In recognition of the transformative positive potential of AI, and as part of ensuring wider international cooperation on AI, we resolve to sustain an inclusive global dialogue that engages existing international fora and other relevant initiatives and contributes in an open manner to broader international discussions, and to continue research on frontier AI safety to ensure that the benefits of the technology can be harnessed responsibly for good and for all. We look forward to meeting again in 2024…(More)”.

Does the sun rise for ChatGPT? Scientific discovery in the age of generative AI

Curated on Posted on by Stefaan Verhulst

Paper by David Leslie: “In the current hype-laden climate surrounding the rapid proliferation of foundation models and generative AI systems like ChatGPT, it is becoming increasingly important for societal stakeholders to reach sound understandings of their limitations and potential transformative effects. This is especially true in the natural and applied sciences, where magical thinking among some scientists about the take-off of “artificial general intelligence” has arisen simultaneously as the growing use of these technologies is putting longstanding norms, policies, and standards of good research practice under pressure. In this analysis, I argue that a deflationary understanding of foundation models and generative AI systems can help us sense check our expectations of what role they can play in processes of scientific exploration, sense-making, and discovery. I claim that a more sober, tool-based understanding of generative AI systems as computational instruments embedded in warm-blooded research processes can serve several salutary functions. It can play a crucial bubble-bursting role that mitigates some of the most serious threats to the ethos of modern science posed by an unreflective overreliance on these technologies. It can also strengthen the epistemic and normative footing of contemporary science by helping researchers circumscribe the part to be played by machine-led prediction in communicative contexts of scientific discovery while concurrently prodding them to recognise that such contexts are principal sites for human empowerment, democratic agency, and creativity. Finally, it can help spur ever richer approaches to collaborative experimental design, theory-construction, and scientific world-making by encouraging researchers to deploy these kinds of computational tools to heuristically probe unbounded search spaces and patterns in high-dimensional biophysical data that would otherwise be inaccessible to human-scale examination and inference…(More)”.

The UN Hired an AI Company to Untangle the Israeli-Palestinian Crisis

Curated on Posted on by Stefaan Verhulst

Article by David Gilbert: “…The application of artificial intelligence technologies to conflict situations has been around since at least 1996, with machine learning being used to predict where conflicts may occur. The use of AI in this area has expanded in the intervening years, being used to improve logistics, training, and other aspects of peacekeeping missions. Lane and Shults believe they could use artificial intelligence to dig deeper and find the root causes of conflicts.

Their idea for an AI program that models the belief systems that drive human behavior first began when Lane moved to Northern Ireland a decade ago to study whether computation modeling and cognition could be used to understand issues around religious violence.

In Belfast, Lane figured out that by modeling aspects of identity and social cohesion, and identifying the factors that make people motivated to fight and die for a particular cause, he could accurately predict what was going to happen next.

“We set out to try and come up with something that could help us better understand what it is about human nature that sometimes results in conflict, and then how can we use that tool to try and get a better handle or understanding on these deeper, more psychological issues at really large scales,” Lane says.

The result of their work was a study published in 2018 in The Journal for Artificial Societies and Social Simulation, which found that people are typically peaceful but will engage in violence when an outside group threatens the core principles of their religious identity.

A year later, Lane wrote that the model he had developed predicted that measures introduced by Brexit—the UK’s departure from the European Union that included the introduction of a hard border in the Irish Sea between Northern Ireland and the rest of the UK—would result in a rise in paramilitary activity. Months later, the model was proved right.

The multi-agent model developed by Lane and Shults relied on distilling more than 50 million articles from GDelt, a project that ​​monitors “the world’s broadcast, print, and web news from nearly every corner of every country in over 100 languages.” But feeding the AI millions of articles and documents was not enough, the researchers realized. In order to fully understand what was driving the people of Northern Ireland to engage in violence against their neighbors, they would need to conduct their own research…(More)”.

AI Globalism and AI Localism: Governing AI at the Local Level for Global Benefit

Curated on Posted on by Stefaan Verhulst

Article by Stefaan G. Verhulst: “With the UK Summit in full swing, 2023 will likely be seen as a pivotal year for AI governance, with governments promoting a global governance model: AI Globalism. For it to be relevant, flexible, and effective, any global approach will need to be informed by and complemented with local experimentation and leadership, ensuring local responsiveness: AI Localism.

Even as consumers and businesses extend their use of AI (generative AI in particular), governments are also taking notice. Determined not to be caught on the back foot, as they were with social media, regulators and policymakers around the world are exploring frameworks and institutional structures that could help maximize the benefits while minimizing the potential harms of AI. This week, the UK is hosting a high-profile AI Safety Summit, attended by political and business leaders from around the world, including Kamala Harris and Elon Musk. Similarly, US President Biden recently signed an Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, which he hailed as a “landmark executive order” to ensure “safety, security, trust, openness, and American leadership.”

Generated with DALL-E

Amid the various policy and regulatory proposals swirling around, there has been a notable emphasis on what we might call AI globalism. The UK summit has explicitly endorsed a global approach to AI safety, with coordination between the US, EU, and China core to its vision of more responsible and safe AI. This global perspective follows similar recent calls for “an AI equivalent of the IPCC” or the International Atomic Energy Agency (IAEA). Notably, such calls are emerging both from the private sector and from civil society leaders.

In many ways, a global approach makes sense. Like most technology, AI is transnational in scope, and its governance will require cross-jurisdictional coordination and harmonization. At the same time, we believe that AI globalism should be accompanied by a recognition that some of the most innovative AI initiatives are taking place in cities and municipalities and being regulated at those levels too.

We call it AI localism. In what follows, I outline a vision of a more decentralized approach to AI governance, one that would allow cities and local jurisdictions — including states — to develop and iterate governance frameworks tailored to their specific needs and challenges. This decentralized, local approach would need to take place alongside global efforts. The two would not be mutually exclusive but instead necessarily complementary…(More)”.