Berkman Klein Center for Internet & Society at Harvard University

In their article, Marshall Van Alstyne and his coauthors emphasize, “Data is the most crucial input to any AI model, but too often the size of the data receives much more attention than its quality.” They provide solutions to this prevalent issue and explore approaches to three additional challenges that individuals face when leveraging AI. https://lnkd.in/eEec9bzg 

Each year, Nieman Journalism Lab asks some of the smartest people in journalism what they think is coming in the new year. Here are their predictions for 2025, featuring BKC fellow Ryan Y. Kellett, and faculty associates Jasmine McNealy and Jonas Kaiser. https://lnkd.in/edBkvrDN

From BKC’s founder Jonathan Zittrain – well, you might not have heard of him: “Apparently, I was among a few professors whose names were spot-checked by the company around 2023, and whatever fabrications the spot-checker saw persuaded them to add me to the forbidden-names list. OpenAI separately told The New York Times that the name that had started it all—David Mayer—had been added mistakenly. And indeed, the guillotine no longer falls for that one. For such an inelegant behavior to be in chatbots as widespread and popular as GPT is a blunt reminder of two larger, seemingly contrary phenomena. First, these models are profoundly unpredictable: Even slightly changed prompts or prior conversational history can produce wildly differing results, and it’s hard for anyone to predict just what the models will say in a given instance. So the only way to really excise a particular word is to apply a coarse filter like the one we see here. Second, model makers still can and do effectively shape in all sorts of ways how their chatbots behave.” via The Atlantic

New research from former Visiting Scholar Jeff Hall (23-24) at Institute for Rebooting Social Media about common misconceptions of social media's impact on mental health.

News organizations are reckoning with the fact that their use of design may be putting them in the same category as sales sites and scammers. Why? Dark patterns, or manipulative designs in online interfaces and experiences that are made to extract something of value from a user that the organization behind the site or app would not otherwise get, is an increasingly common practice. And news sites are not immune. This reflects “the battle that news organizations have with creating and sustaining audience trust and their economic interests,” notes BKC Faculty Associate Jasmine McNealy. Learn more about what you might be seeing, and not realizing, is a dark pattern used on the news site you read online. Nieman Journalism Lab https://lnkd.in/gRN5hM8i

Former President of Colombia Iván Duque Márquez shares his thoughts on the research of Sandra Cortesi, BKC's long-time leader of everything youth and media related.

Today we're launching the Institutional Data Initiative at Harvard to work with libraries, government agencies, and other knowledge institutions to help refine and publish their collections as data, with an eye toward AI. Data access plays an important role in the development of AI models. It helps define who is represented in them, who they empower, and even who can build them. Data integrity, as a result, is important as well. Data access and integrity are two things institutions know well. They hold vast collections and think a lot about knowledge access and stewardship. Stewardship is a nice word because it conveys a sense of integrity over time, and time is something institutions excel at utilizing. As a team of technologists who deeply value knowledge institutions, we designed Institutional Data Initiative at Harvard as a gear to help the AI and institutional communities, each moving at their own unique speeds, mesh and transfer energy between them. Energy, experience, and expertise. Our goal is to work within the missions of institutions to help unlock and improve access to their collections, not only for AI uses but traditional patron access as well. This work can help expand the breadth of people that institutional collections and AI models are able to empower. Our first project is a collection of 1M public domain books, scanned at Harvard Library as part of the Google Books project. We're fortunate enough to have the support of Google in releasing this dataset and will do so in early 2025 with accompanying analysis. We're also working with Boston Public Library on an active scanning project of theirs involving millions of pages taken from public domain newspapers. We're hoping to make progress on some of the classic OCR challenges presented by newspapers, while evaluating the resulting data for model training. Our launch is generously supported by Microsoft and OpenAI, and these projects are just the start. If you’re part of an institution, we’d love to hear how we can help. If you’re an AI/ML researcher, we'd love to collaborate. We're also hiring for our core team, here at Harvard. Reach out. To learn more about the Institutional Data Initiative at Harvard, read our launch announcement here: https://lnkd.in/e6WJvBSN

We are delighted to announce our first foray into Executive Education! Students will learn about AI governance, navigating privacy protections, and international regulatory issues with BKC heavyweights Jonathan Zittrain Christopher Bavitz Urs Gasser James W Mickens Mark Wu. Special shoutout to Program Chair William Fisher Learn more and apply today!

“Do we want a future in which some people, almost certainly the richest...double their life expectancy, while others’ life expectancy remains largely unchanged?” BKC faculty associate Nick Couldry and Asher Kessler respond to Anthropic AI CEO Dario Amodei’s recent essay about the expansive potential for “physical health, neuroscience and mental health, economic development, war and peace, and finally work and meaning.” The co-authors question the portrait of biological determinism Amodei has painted, accessible only to the elite: https://brk.mn/LKWTEI The co-authors question the portrait of biological determinism Amodei has painted, accessible only to the elite: https://brk.mn/LKWTEI

One reason sensational content spreads so quickly? It comes down to the information economy, agreed Nieman Foundation for Journalism at Harvard fellows Ben Reininga and Jesselyn Cook, author of The Quiet Damage, in conversation with BKC founder Jonathan Zittrain. Jesselyn noted that content moderation could only go so far when content is incentivized in its current reward system: “It’s not so much about what we’re allowed to say or what we’re not, it’s more how this content is treated — what is eligible for monetization and algorithmic amplification. “I wouldn’t mind seeing a little bit more overly aggressive rules in place for dialing down that amplification and seeing how this content performs on its own without this unnatural boost.” https://brk.mn/NNRPGL