Data Culture Group @ Northeastern University

Data Drums wins a Data Sonification Award

Wed, 29 Apr 2026 15:00:00 +0000

We’re thrilled to share that Data Drums has won a 2026 Data Sonification Award in the Arts category!

“Data Drums: A Rhythmic Call to Climate Action” is a live performance that brings global CO₂ emissions to life through the dynamic beats of a Brazilian-style drum ensemble. Drawing on carbon emissions data from Brazil, India, and the United States, the four-movement piece translates numbers into rhythm. At various points in the performance each drum and beat represent a different slice of the data; Volume, tempo, and pattern shift to reflect emissions levels and growth rates. Through drumming the contrasts and connections between countries becomes visceral and immediate. You feel the emissions. The piece culminates in an interactive finale where the audience joins in, embodying the collective responsibility and possibility for climate action.

The project was co-created with Marcus Santos and Lily Gabaree, and supported by the Northeastern CAMD Climate Change and Public Communication Impact Group and the Northeastern Center for Transformative Media.

We’re proud of how this project pushes what data storytelling can be — not a chart or a dashboard, but a room full of people making noise together about climate change. Get in touch if you’re interested in bringing Data Drums or other participatory data experiences to your community. You can read more the ideas behind about Data Drums in my Community Data book.

Beyond Accuracy: A Case Study of Sociotechnical ML Evaluation

Sun, 22 Feb 2026 15:00:00 +0000

The use of large language models and generative AI has significantly impacted computational social science, creating new capabilities and raising new questions for those using software to support social analysis. Key among those is how to evaluate if large models or prompt-based workflows are actually performing better than conventional machine learning. Against this backdrop, we want to propose a more sociotechnical approach to evaluation, one that includes contextual metrics and discussion beyond the standard recall and precision scores. This brief case study shares what happened when we tried this approach on a classifier we’ve built as part of the Counterdata Network’s collaboration with the nonprofit, US-based organization Pregnancy Justice.

Some Background

Building on software systems built as part of the Data Against Feminicide project, the Counterdata Network partners with groups tracking human rights violations via online news to partially automate some of the content discovery and evaluation. In practice, this looks like working together with organizations to train custom ML classifiers that score candidate news articles from online news archives (based on geographic filtering and on keyword matching) for relevance to the violations they want to track.

One of our more recent partnerships is Pregnancy Justice, a national legal advocacy organization that advances and defends the rights of pregnant people, no matter if they give birth, experience a pregnancy loss or have an abortion. Pregnancy Justice’s research team tracks pregnancy criminalization cases across the U.S. in collaboration with academic partners. These are cases charging pregnant people with crimes related to pregnancy, pregnancy loss, or birth. Their extensive and rigorous work is documenting the impact of the Dobbs supreme court ruling, and supporting people brought into the criminal system because of their status as pregnant. Our software system surfaces relevant news articles about cases that fit their focus each week.

The platform where the Pregnancy Justice team reviews retrieved articles and labels them as relevant or not relevant.

Improving the Model

After collecting and annotating a sample training set and experimenting with ML models, the first version of the model was deployed into the email alerts system to retrieve articles in November 2024. The Pregnancy Justice team reviewed retrieved articles each week and labeled each as Relevant / Not Relevant. When our team analyzed the results, we found the false positive rates were near 40%, meaning that every 2 out of 5 articles returned to the team was not relevant to their search.

From a purely technical standpoint, this was alarming. The model seemed barely better than random chance. As ML researchers, our instinct was to fix the model. We designed three parallel improvement experiments to reduce false positives and boost conventional ML performance metrics. What we learned through the process challenged our concepts about what better even means.

Experiment 1: Should we incorporate better data?

First, we updated our TF-IDF model using a subset of newly annotated data from PJ’s weekly reviews. After incorporating the new data, the model’s false positive rate decreased from 55% to 10%. Our analysis showed that much of the original misclassification was a result of domain-specific terms, such as “midwives,” which appeared frequently in real-world news cases but were absent from our initial manually sourced training set. By incorporating the team’s feedback, the dataset better matched the real-world linguistic context, allowing the simple model to more accurately generalize.

Experiment 2: How about more complicated models?

Our first model used TF-IDF, a method that identifies important words that are common in documents we care about but relatively rare in the entire collection. It’s a relatively simple approach by modern NLP standards, as it counts words rather than attempting to understand semantic meaning.

We hypothesized that more sophisticated embedding models might perform better. Unlike TF-IDF, large language models (LLMs) like BERT, Universal Sentence Encoder, and MPNet are designed to understand the semantic context—the meaning behind sentences. Theoretically, they should capture the nuanced relationships inherent in complex social issues like pregnancy criminalization.

We tested several state-of-the-art embedding models against our current TF-IDF baseline. However, given the ongoing debates about the ecological impacts of LLMs, we decided to evaluate them on more than just accuracy. We also evaluated their environmental and computational costs.

The Trade-off: Accuracy vs. Environmental Impact

The AI community is increasingly recognizing that model performance cannot be divorced from its environmental cost. Extensive research has emphasized the nontrivial energy consumption and associated CO2 emissions from training and deploying LLMs. This environmental footprint —through pollution, resource extraction, and climate change— disproportionately affects marginalized communities, many of the very communities Pregnancy Justice serves. As a result, our team was committed to prioritize environmental sustainability. With this in mind, we evaluated not just accuracy metrics but also running time and CO2 emissions (estimated using the codecarbon package).

Embedding Model	FPR	Time	CO2 Emissions
TF-IDF (w/ new data)	0.12	1x	<1 mg
MPNet	0.07	105x	>10 mg
USE	0.07	31x	>5 mg
Sentence-BERT	0.09	18x	>10 mg
mxbai-embed-large	0.23	353x	>100 mg

Model performance (FPR) and computational cost (time, CO₂ emissions) across embedding approaches. FPR = false positive rate (lower is better);

Our results show that while some embedding models achieved low false positive rates, the computational costs were staggering. The mxbai-embed-large model, for example, took 353 times longer to run and emitted significantly more carbon than our original model. The accuracy gains were small, especially compared to the retrained TF-IDF model, with substantial environmental costs. This matches findings from other research, where simpler models such as static embeddings have been shown to perform comparably on digital humanities tasks while requiring significantly less compute (Ehrmanntraut et al., 2021).

After discussing these results with the Pregnancy Justice team, we reached a critical conclusion that the marginal gain in accuracy does not justify the increased environmental and computational harm.

Experiment 3: How confident should the models be?

Our final experiment involved adjusting the model’s decision threshold. Our model outputs a probability score between 0 and 1 for each article, indicating the model’s confidence that the article is relevant. For the first model we used a threshold of 0.5: articles scoring above 0.5 were classified as relevant.

Histogram showing the distribution of scores on stories from the first model. Note the majority after under 0.5, our threshold for including in results sent to the team.

We hypothesized that If the threshold were raised, e.g. to 0.7 or 0.8, the model would reduce false positives by only returning articles it was very confident about. Our experiments confirmed this, but also showed a tradeoff: a stricter threshold means missing some truly relevant articles.

When we presented this tradeoff to the Pregnancy Justice team, they offered a surprising insight: they wanted the false positives.They explained that reviewing even the false positive articles, those not directly about pregnancy criminalization, was actually valuable. These articles helped them understand the broader landscape: adjacent policy discussions, related criminal justice issues, and the wider context in which pregnancy criminalization occurs. What we labeled as irrelevant from a technical standpoint was often relevant from a strategic advocacy standpoint.

This insight fundamentally reframed our evaluation. The model wasn’t just a filter to remove noise—it was a discovery tool that exposed the team to a broader information ecosystem. The “noise” contained signals we hadn’t anticipated.

Towards Sociotechnical and Community-Centered ML Evaluation

This case study highlights three critical lessons for evaluating ML systems with stakeholders and communities:

Bigger isn’t always better: Improving the training data might yield far significant improvements than increasing model complexity.
Let communities define success: Technical metrics like accuracy and false positive Rates are proxies for utility. What counts as correct output depends on contexts, workflows, and values that these metrics cannot capture. We recommend co-designing metrics and success criteria with the people who will use the system.
Evaluate the System, Not Just the Algorithm: Efficiency is an ethical metric. By quantifying the time and CO2 costs, we made a value judgment that aligned our technical infrastructure with our project’s justice-oriented values.

Data Journalism in Dark Times: notes from the C+J keynotes

Sun, 14 Dec 2025 15:00:00 +0000

I just returned from the Computation + Journalism Symposium in Miami, and I’m still processing the weight of what I heard. The four keynotes this year painted a sobering picture of data journalism operating under pressure—from authoritarian governments, algorithmic complexity, and information warfare. But they also offered glimpses of resilience and innovation that feel essential right now. Here are short summaries of each keynote, drafted with some AI assistance based on my live notes.

Me ready for Miami in December ☀️

When Data Journalism Isn’t Enough

Attila Bátorfy opened with 15 years of experience doing data journalism in Hungary under Viktor Orbán’s increasingly authoritarian government. His message was blunt: sophisticated data journalism hasn’t stopped democratic erosion. Despite producing groundbreaking investigations–including tracking private jets to predict and photograph secret meetings of oligarchs–the work had limited impact on political reality.

What struck me most was his catalog of how governments actively undermine data journalism: releasing low-resolution data, commodifying public information, creating datasets with bad methodology, and even sponsoring “independent” counter-data. His COVID dashboard became a trusted source (5 million uniques in 2 years, cited in dozens of scientific journals) precisely because government data was so unreliable. But he’s now moving away from data journalism, noting that audiences “don’t want facts” and that Hungarian newsrooms see it as expensive with bad ROI. The one exception? Sports and health data still cut through the noise. Attila’s upcoming work focuses on historical representatos of data, and I’m excited to learn from what he shares.

Data as Weapon and Shield

Kae Petrin’s talk on anti-LGBTQ+ data policies brought this theme of governmental manipulation home to the US context. The Trump administration’s systematic removal and alteration of CDC data—changing “gender” to “sex,” taking down youth risk behavior data, adding warnings that datasets are “threats to society”—isn’t just censorship. It’s erasure through datafication.

Petrin highlighted a crucial methodological problem that journalists fumbled: when the CDC improved its estimation methods for trans youth populations, the numbers roughly doubled. News coverage reported a “sharp rise” in trans youth, playing into “social contagion” narratives, when the real story was that better measurement revealing what was always there. Meanwhile, the demographic data showing far more trans young people than older adults raises haunting questions about where trans adults over 24 are. These are questions we can’t answer if the data disappears.

But Petrin also showed data’s dark side: data can make people legible and vulnerable.

User Agency vs. Algorithmic Power

Homa Hosseinmardi’s work offered a different kind of complexity, and reveals that the villain might not be who we think. Her research on YouTube radicalization used an innovative “counterfactual bots” method to separate user intention from algorithmic influence. The finding? Users who relied exclusively on YouTube’s recommender actually consumed less partisan content than those who actively searched and navigated themselves.

This doesn’t absolve platforms of responsibility, but it challenges the narrative of algorithms as “great radicalizers.” The real picture is messier: a small group of users consuming vast amounts of far-right content, driven largely by external links and their own preferences. When her team simulated users without agency (random video selection, first recommendation only, etc.), those bots were showng content that drifted towards the politcal center over time. The algorithm appears to reflect and amplify existing preferences rather than creating them from scratch.

The implications for journalism are significant: we may be overexamining algorithms while overlooking user agency and the broader information ecosystem that drives people to platforms in the first place.

Navigating Narrative Warfare

Yevheniia Drozdova closed with work from texty.org.ua, showing what data journalism looks like when your country is literally at war. Working from Kyiv, her team has moved beyond fact-checking individual claims to mapping entire narrative ecosystems. They train models to identify manipulation techniques, track coordinated bot networks across Telegram (where 50%+ of Ukrainians get news), and detect patterns in TikTok campaigns using shared filters and timing signatures.

Her team found 2,000 TikTok bot accounts and documented coordinated Telegram networks. They trained models to detect emotional manipulation techniques, finding that 90% of messages from Russian accounts use fear and doubt. But their work isn’t just technical detection, it’s about providing context before narratives go viral. As she put it: “Fast news only satisfies the hunger for information, while slow journalism addresses a different deficit—deficit of understanding.”

What Stays With Me

These four talks trace a dark arc: data journalism struggling against authoritarian data practices, vulnerable populations made more visible and endangered through datafication, algorithmic explanations that may miss human complexity, and full-scale information warfare requiring new methodological approaches.

But I’m also struck by the practitioners’ persistence. Bátorfy’s COVID dashboard becoming a trusted institution. Petrin tracking articles through the Trans News Initiative to understand coverage patterns. Hosseinmardi’s methodological innovation in creating counterfactual bots. Drozdova’s team prebunking narratives in real-time during a war.

These keynotes show various ways data journalism remains essential for documentation, for providing sources of truth, for understanding complex systems, and for serving communities when governments fail them. C+J’25 offered a critical moment of trans-national solidarity that I won’t soon forget.

Notes from MISI Summit on Building Trust in Climate and Science Journalism

Mon, 10 Nov 2025 15:00:00 +0000

I attended the inaugural Boston University Center for Media Innovation & Social Impact (MISI) Summit recently, where a diverse group of leaders came together to discuss the theme of “Communicating Climate.” Hosted by the Center’s Directory Eric Gordon, the event included sessions on related topics. Here’s a quick blog post about key points from one session that focused on how journalism, academia, and public media can restore trust in an era of economic instability and eroding credibility.

The start of the “Business of News” Panel

The Business of News: Audience Trust and the Future of Science and Climate Policy Journalism

This panel included Brian McGrory (BU Journalism Dept Chair, formed Boston Globe editor), Sue Robinson (Helen Firstbrook Franklin Professor of Journalism at the University of Wisconsin–Madison), and Dan Mauzy (WBUR’s Executive Editor for News).

It was moderated by Meghan Irons (investigative journalist and Professor of the Practice at Boston University), who opened with the fundamental question: how do we restore public trust when newsrooms face economic instability, eroding credibility, and an increasingly complex information landscape?

Local News, Attention, and Transparency

Dan Mauzy opened with an optimistic take: local news has a leg up compared to national outlets, and public media has an advantage among those familiar with it. He focused don transparency as a tool for building trust, particularly related to showing how journalism happens and showing your work. Dan noted a sobering statistic: “8 or 10 people have never spoken to a local reporter”—a problem that needs solving. In parallel, to ensure they are representative WBUR conducts audits on their source demographics, comparing them to local census data to ensure representative coverage.

For climate coverage specifically, the answer is making it local: to “literally bring it home to people.” Instead of abstract global warming statistics, WBUR focuses on local watering holes, jobs, heat pumps, energy bills, and solutions. Dan pointed to their Cape Cod plumbing contamination story about nitrates as an example of making climate impacts tangible. WBUR now has 3 reporters plus an editor dedicated to their climate team.

Brian McGrory acknowledged the widespread distrust across institutions, but noted an interesting pattern: crises drive attention. The Boston bombing, Trump’s first presidency, and COVID all led to subscription bumps for the Globe. Digital subscriptions actually doubled during the pandemic. “These are votes of confidence on things we do and how we do it,” Brian said. But he expressed uncertainty about how to reach people who aren’t actively seeking out journalism without alienating current loyal readers. His philosophy: “when readers see things they like, admire, trust, they will stay, otherwise they will leave.” The answer for him? “Focus on interesting, important, consequential stories and the business model will follow.”

Sue Robinson opened arguing that journalism education and newsrooms aren’t teaching key skillsets: radical transparency vis-à-vis audience engagement and restoring trust. She pointed to organizations modeling this approach: ProPublica, Seattle Times, WITF, Baltimore Banner. WITF’s 2024 election coverage particularly stood out to Sue as a model, with Q&As, pledges about misinformation and accuracy, and messaging about what’s happening behind the story.

She pushed back on Brian’s faith in the story, pointing out that “The audience doesn’t come to you—they’re being told to hate us.” Instead of waiting for audiences to find them, news organizations need to actively work on finding common values with communities.

What’s Actually Working

Sue shared how listening sessions can change hearts and minds of those reaaders who are being told to disengage from local news. Through the Trusting News project, she works with journalists who talked to people disengaged from news brands. The sessions were challening for journalists who had to listen with “devastating results that traumatized [them].” But here’s the remarkable part: after those conversations, over half of participants said it built confidence and one-third said they’d probably subscribe. The takeaway? “Bring people together around a shared topic of concern—listening and convening can change hearts and minds.”

Dan echoed this from WBUR’s experience hosting listening sessions on various topics, particularly in response to parachute journalism in the Merrimack Valley. It’s “enriching for reporters to hear what people think is overlooked, what makes them proud.” The challenge? How to scale these personal interactions.

Brian agreed that focus groups are great, creating opportunties where meeting the person behind the reporting can “change the tune of the haters”. He pushes to have journalists that get out into community regularly: “in local news we can do that.” Brian also emphasized inviting readers into the process itself—not just the meta aspects like ethics policies, but the exposing the actual work of reporting. When people understand the process, trust follows.

Sue highlighted that public media is out front on innovative approaches. She pointed to Lookout Santa Cruz as a leading example. They partner with influencers in the community itself to scale their reach, working with people who have huge groups of followers not already connected to their local media organizations. Lookout is even fundraising with non-profits together. We “need more creativity like this,” Sue urged.

Dan shared that WBUR is doing more vertical video now and reaching new audiences. They’ve done partnerships with trusted community members on niche topics, which has expanded reach and built trust. The catch? “Platforms are not our friends, so how to build loyal communities developed within those platforms is challenging.”

The Hard Questions

Are Some People Unreachable? Brian posed a difficult question about major outlets like the NYT, WSJ, and Globe: while subscribers are growing, they’re “not reaching a critical part of the country.” His question: “BUT are those people reachable, and what might be the sacrifice to reach them?”

The Talent Pipeline is Broken Sue ended on a sobering economic note: pay rates aren’t good enough, so students coming out can’t afford to go into traditional newsrooms. Instead they’re going to NGOs and corporate communications to exercise similar skills. This threatens the entire pipeline of journalism talent.

The Continually Shifting Media Ecosystem Dan raised another challenge for the future: “how to create loyalty in a fractured media space? Discovery is harder in an AI-mediated internet search world.” As platforms and AI change how people find information, news organizations need new strategies for building direct relationships with audiences.

I left with a sense of both sobering challenges and promising experiments. While there’s no silver bullet for restoring trust, the path forward Sue, Brian, and Dan pointed at focused on doubling down on transparency, community engagement, and creativity. I can get behind that.

The views out the new-ish BU Computer and Data Science building are alwasy amazing 🍁☀️💦

Notes from csv,conf,v9: Community-Centered Data Stories

Mon, 29 Sep 2025 15:00:00 +0000

The University of Bologna graciously hosted csv,conf,v9 in Palazzo Malvezzi. Accordion music drifted in through the windows while we gathered, echoing the mix of technical innovation and creative application that folks shared. Speakers from around the world shared how they’re using data not just to inform, but to transform communities. Here are gently editing notes from the sessions I attended (edited with some help from AI); any errors are probably my fault.

Photo of all the attendees at csv,conf,v9 (source: John Chodacki)

Filling Information Voids with Chatbots: Using LLMs on WhatsApp for News Access

Mathias Felipe (he/him), InfoAmazonia

In the Amazon forest, communities often face critical information voids around environmental issues, while relying heavily on WhatsApp as their main communication platform. This talk presents BOTO, an open-source chatbot developed by InfoAmazonia that delivers localized, thematic environmental information—such as deforestation alerts and wildfires—through WhatsApp. Powered by Large Language Models (LLMs) and designed with a low-bandwidth architecture, BOTO aims to make local news information more accessible in under-served areas. In this talk, we’ll discuss how BOTO was created with participatory design to empower communities, bridge information gaps, and put news directly in the hands of those who need it most.

Felipe showing what Boto looks like in action

The most striking thing about BOTO isn’t its technical architecture, others are also exploring WhatsApp Business API integration with RAG-based article search; it’s how InfoAmazonia approached the design process. Focus groups with different indigenous communities shaped every decision, from interface choices to content summaries. The result? A system where people can select their region and themes, then receive data summaries and links directly to their phones.

Felipe emphasized something crucial: their audience doesn’t have a problem with chatbots. The hesitation comes from outside assumptions about what “non-technical users” want. When you actually ask communities what they need, the solutions become clearer. BOTO soft-launched recently with official rollout planned for next month, sponsored by ICFJ. The real test will be sustained engagement, but the co-design foundation suggests they’re building something people actually want to use.

Blazing Fast Automagical Metadata

Joel Natividad

With the pervasiveness of low quality, low resolution metadata in existing Data Catalogs, and new metadata standards like DCAT3 and Croissant demanding even more FAIR metadata, how do you make it easier for Data Stewards to maintain High Quality Data Catalogs?

Natividad giving his talk in a somewhat distractingly ornate room at the palace

Natividad’s (from datHere) talk hit on every data steward’s nightmare: the metadata maintenance burden. New standards like DCAT3 and Croissant (for machine learning) demand even richer metadata: properties, data dictionaries, summary stats, frequency tables. The solution? Flip the workflow. Instead of forcing metadata entry before data upload, let people upload first and generate suggested metadata automatically. Their qsv and describegpt tools infer data dictionaries, descriptions, and 40+ summary stats. It creates a “data steward in-the-loop” approach—human oversight with automated assistance.

Most intriguingly, they’re moving toward “chat with your catalog”—custom chatbots that can surface relevant CKAN datasets when answering questions. It’s RAG applied to data discovery, potentially transforming how people find and use datasets that are published and publicly available.

Guillaume Plique (he/him), médialabSciencesPo, Paris

CSV is ubiquitous in social sciences and in the humanities. CSV data is indeed the perfect bridge between social scientists, accustomed to dealing with tabular data, and research engineers needing to process the same data. That is why SciencesPo’s médialab has been building many of its Open-Source tools around CSV files, from well-designed web apps such as Table2Net to convert tabular data into graph data, down to powerful CLI tools such as minet to collect data from the web or xan to process tabular data using constrained resources.

Plique, and the lab where he works, are _all_ about CSV

Plique delivered a love letter to CSV. His argument is compelling: CSV is affordable, understandable, and free. It functions as a bridge between researchers, students, and engineers. It doesn’t require complex compute resources and naturally encourages data sobriety. The médialab’s tool ecosystem is impressive—table2net for web-based CSV-to-graph conversion, takoyaki for clustering CSV data to find clerical errors, minet for CLI-based CSV processing, and xan for visualization. All built on the principle that tabular data is inherently more comprehensible than JSON hierarchies. His closing advice: “If you need random access, use SQLite. Otherwise, CSV.” It’s a reminder that sometimes the simplest tools are the most durable—you’ll still be able to open a CSV file in 50 years.

Keynote: Developing tech with community: the example of Open Data Editor

Sara Petti, Open Knowledge Foundation

Open Data Editor is a desktop application specifically designed to help people detect errors in tables. It has been developed in constant interaction with the community from a very early stage. These interactions helped us understand what was really helping the community and what not, and especially made us aware of how much the use of such a tool could actually be helpful in increasing data literacy.

Petti dove into how they built a data cleanup and analysis tool with partners

Petti’s keynote was a masterclass in community-centered design. OKFN initially tried to build “an app to do it all”—review data, prepare it, map it, write articles. The scope was overwhelming and misaligned with user needs. The breakthrough came through extended pilot testing with organizations like ACIJ in Argentina and StoryData in Spain. This led to dramatic simplification and focus. The second round of pilots with five different organizational types revealed something crucial: researchers worried about changing metadata, navigation and design were hugely impactful, and people don’t read documentation.

Key lessons that resonated: language matters enormously (“validation,” “execute” carry technical baggage), there’s only so much you can do to make highly technical tasks accessible, and many challenges are fundamentally non-technical. Most importantly—iterate based on feedback, keep things simple, and don’t be confused by the latest tech. The project pivoted from React to a more sustainable framework, emphasizing the importance of maintainable code over popular technologies.

Democratizing data: data literacy for community action

Emily Zoe Mann (she/her/hers), University of South Florida Libraries

This presentation will share the results of Data Literacy for Community Action, a year-long pilot in the community of St. Petersburg, Florida, USA that provided data literacy to non-profit groups and interested members of the community, with a focus on using local data sets and data related to social determinants of health.

Mann shared lessons from hands-on data workshops with library patrons

Mann’s project exemplifies academic-community collaboration done right. The partnership between University of South Florida, the public library system, and a philanthropic health equity organization created something none could achieve alone. The workshop structure was brilliantly simple: introduce a concept, share a local dataset demonstrating it, bring in a local person to discuss real-world application, then run group activities. Her red tide sonification example from South Florida shows how hyperlocal data becomes immediately relevant when paired with familiar environmental experiences.

One insight that stood out: attendees came with vastly different question types, from basic to nuanced. Mann’s library background prepared her for meeting people at all skill levels, but it reinforced how “data literacy” isn’t a single skill—it’s a spectrum of capabilities that communities develop over time. The flexibility lesson is crucial for similar pilots: stay responsive to what people actually need rather than predetermined curricula.

Keynote: Clara Jiménez & David Fernández Sancho

Maldita.es Foundation

Clara Jiménez Cruz is the co-founder and CEO of the Maldita.es Foundation, a leading organization in the fight against misinformation… David Fernández is the current CTO of the Maldita.es Foundation, where he leads the development of innovative digital tools to combat disinformation.

The team shared a timeline of work they’ve done

This keynote hit hardest, perhaps because of recent events. Jiménez and Fernández positioned disinformation not as isolated false information, but as structured narratives—emotionally engaging, rationally grounded, amplified by trusted figures, global in scope. Their Valencia floods case study was chilling: the same narrative emerging simultaneously from Russian sources, far-right groups, and anti-vaxxers, all attacking institutional trust. The message was consistent: “You can’t trust government, trust me.”

Their AI-assisted approach uses Wikidata entity extraction to cluster claims, with human-in-the-loop narrative association. They’re tracking 600+ claims monthly—a scale that demands automation while recognizing AI’s limitations in understanding social context.

The most powerful moment was Jiménez’s declaration: “data as a means of resistance.” In an environment where “the internet is becoming more dangerous on purpose,” systematic documentation and analysis of disinformation patterns becomes a form of civic defense.

Data without borders: Straight out of spreadsheets and into the streets

Michael Brenner (He/Him/His), Data4Change

At Data4Change, we work with communities around the world to explore how data can be collected, understood and shared in radically different ways. This talk takes you through a series of surprising, community-led projects that didn’t begin with data, but with burning questions our community partners wanted to find answers to.

Brenner’s examples were extraordinary—from addressing intimate partner violence in Uganda’s LGBTQ community, to flute music rendering Ethiopian census data audible to the blind. Each project started not with data, but with burning community questions. I mention many of them in my Community Data book.

The data murals project with Haki Data Lab in Mathare particularly resonated. Rather than imposing external visualization methods, they recognized murals as a standard local communication mode, then created input visualization murals with string and stencils. The “power portraits” methodology for the Equal Victims project was brilliant—co-created visual representations that function simultaneously as art and data points about lived experience, avoiding retraumatization while enabling systematic documentation.

Brenner’s language shift is important: moving from “data literate” toward “data curious, data critical, and data confident.” It reframes the relationship from deficit-based (lacking literacy) to asset-based (developing capabilities).

Keynote: Fifteen Years into the Open Data Movement

Giorgia Lodi & Andrea Borruso

Fifteen years into the open data movement, we take the stage as an unusual duo to reflect, provoke, and laugh about where we’ve landed. A dialogue, shaped by friendship, a few successes, many failures, and a bit of fun—think less keynote, more late-night talk show.

This closing keynote was part celebration, part wake-up call. Lodi and Borruso’s Italian open data examples were both hilarious and heartbreaking—ministry electric charging station data published with beautiful graphs but completely locked down, forcing community scraping to make it actually useful. Their analysis cuts deep: following the letter of the law isn’t enough when datasets lack description, remain readable only by specialists, or simply don’t exist (particularly around gender-based violence). Communities get pushed out by technicalities, formalism, and missing data.

A case study about electric car data exemplifies the problem—officially “open” data published as unusable pivot tables instead of raw, machine-readable records. When the ministry finally released proper data, it came with semicolon separators and complete messiness. But their vector geography data example shows what’s possible: despite delivery delays and nested zip file complexity, raw land ownership data triggered new software, plugins, and services precisely because it was genuinely machine-readable.

Their conclusion resonated throughout the conference: exploit technology to support data management, think about machine-to-machine AND machine-to-human AND machine-to-community interfaces, and recognize that technical solutions alone can’t address fundamentally social challenges.

Reflections: Data as Community Practice

Walking through Bologna after two days of talks, I kept thinking about the accordion player who inadvertantly soundtracked the opening day. There’s something fitting about that mix of structure and improvisation, tradition and innovation. These presentations collectively argue for a fundamental shift in how we approach data work in and with community. We must shift from extraction to engagement, from literacy deficits to community assets, from technical solutions to socio-technical systems. Whether it’s BOTO’s co-designed news delivery, Data4Change’s participatory data projects, or Open Data Editor’s community-driven development, the most compelling projects start with community questions rather than data capabilities.

I gave a keynote related to my _Community Data_ book (source: Marco Cortella)

The CSV theme threading through many talks isn’t just about file formats—it’s about accessibility, sustainability, and democratic participation in data practices. In a world increasingly dominated by complex AI systems and proprietary platforms, there’s something radical about insisting that the most important data work can happen with simple, open, maintainable tools. As disinformation accelerates and institutional trust erodes, these community-centered data practices feel more urgent than ever. The question isn’t just how we make data more accessible, but how we make data work serve community needs rather than institutional convenience.

Protest Mapper: a new tool for journalists covering protests

Thu, 08 May 2025 15:00:00 +0000

As protests against the Trump administration spread, journalists are covering local rallies — but putting them in context of larger movements can be a challenge. To help, I built a simple tool that lets you search for recent protests near you and create an embeddable map you can use in your reporting. Try out Protest Mapper.

A map of the more than 400 protests in eastern Massachusetts since the start of the year.

Media Attention to Protests

One of the responses to the nationwide April 5 “Hands Off” and May 1 Mayday rallies was shock at the perceived lack of coverage from mass media outlets. Critics pointed to a lack of both print and digital coverage, arguing that it didn’t reflect the sheer volume of people that came out on the streets.

A sample of headlines critiquing media coverage of the April 5 “Hands Off” protests.

At the same time, local and regional media did show a noticeable bump in coverage. A query for articles mentioning both “hands off” and “protest” in the Media Cloud online news archive shows almost 400 matching stories on the 5th.

A Media Cloud query for “‘hands off’ AND ‘protest’” shows a peak of coverage in state and local papers on April 5.

Yet it appears that local reporters are covering protests in their area, but not often connecting them to larger movements. That might be because coalitions like #50501 aren’t as well known as unions and long-standing activist groups; they don’t have communications people with long-standing relationships to journalists.

Finding & Evaluating a Public Protest Dataset

One approach to help reporters make those links for readers, and put individual events in a broader context, is to use data about local protests. Connecting this weekend’s rally to events over the last few weeks might connect dots for audiences that are seeing public displays of resistance. I wondered if I could quickly map protests in my area based on existing data sources.

I first reviewed existing data sources to see what was available. A quick summary of what I found:

The GDELT project appears to have massive scope and has a “Protests” event category, but I found it too hard to access and use, with many broken links and inconsistent information.
The Carnegie Protest Tracker seems to focus on broad movements, not individual events.
LiveUaMap seems well established, with lots of social media results, but I found the interface challenging to use and they charge for access.
Amnesty International’s Protest Map focuses on police violence in response.
ACLED (Armed Conflict Location & Event Data) appears to be a well staffed non-profit that makes data easily available, updated once a week.
The Harvard Ash Center Crowd Counting Consortium (CCC) appears to be an academic project that is updated once a month.

This led me to focus on ACLED and CCC, since they are both regularly updated and easily accessible/usable. Each includes a list of protests and related manually curated metadata. They include protests about any topic.

My next goal was to evaluate these datasets. One metric that came to mind was to search for media coverage about protests in a fixed place and time, and then to compare that to the datasets at hand. I decided on Massachusetts in February, a context I know well, and pulled datasets from both Media Cloud and ACLED. To find media stories I queried Media Cloud for relevant anti-Trump/Musk protests with this query:

article_title:(
(protest* OR rally OR strike OR boycott OR walkout OR demonstrat* OR vigil)  
AND 
(Musk OR Trump OR Tesla OR closure OR federal OR university OR Israel OR Palestine OR cuts OR speech OR student OR research OR congress OR ICE OR agency OR USAID OR NOAA OR NSF OR NIH)
) and language:en

This discovered 76 related stories. I then coded them by hand to indicate if they were about protests in Massachusetts, identifying 41 stories about anti-Trump protests in the state from Media Cloud. In parallel, I pulled February data for the US from ACLED and filtered by the admin1 column to just Massachusetts events, identifying 51 entries in their database. To evaluate the two, we then reviewed every story in the Media Cloud corpus to identify which ACLED event it was about (linking it by the unique event_id_cnty provided from ACLED). Just two protests were in the Media Cloud dataset that did not have a related ACLED entry, both small small events reported by university student-run news sites. 95% of the sample I pulled from the basic Media Cloud query were already in ACLED, leading me to believe it was a trustworthy source for a list of protests. You can review the coded evaluation data.

A Mapping Solution

Dataset in hand, I decided to create a simple tool that would allow journalists to embed maps of recent local protests in their stories. The map at the top of this article is one example. Visit https://dataculture.northeastern.edu/local-protest-map/ to try it out yourself. You can filter by location, timeframe, and customize the map a little.

Tools like these require ongoing support, so I’ve made some decisions to try and minimize maintenance while maximizing utility. Here is a list of design decisions, concerns, and commitments:

I’ve mimicked the well known DataWrapper workflow to try and make it familiar and usable for data journalists. At the end you get an embed code (HTML iFrame) you can copy and paste into your webpage, just like I did above
This is deployed as a static HTML/CSS/JS site with no database or server compute power required, written with the open-source Svelte UI framework.
Some protests are marked in the exact same spot, for instance in the center of a city. This makes them very hard to see on the map. For these events I’ve a small random bump to the location in order to make them more visible. That is noted in the popup for impacted markers.
The code is open source on GitHub for full transparency.
I’m hosting the tool on GitHub Pages so I don’t have any ongoing expenses.
IFrame embeds can be tricky and limited, so the tool also offers the option to download a static image of the map in .png format.
You can pick between the ACLED (updated weekly) and CCC (updated monthly) data. For now I plan to manually pull the updated data each week, but hope to automate that later.
I’ve used a light background map base layer from Stadia Maps. This means it can’t support more than 200,000 hits across all uses. If this somehow magically becomes very popular I can easily change that via the Leaflet mapping library.

Give it a try, share it, and let us know what you find in the comments. Hopefully this tool and explanation will help you and others include data about the wave of protests to your reporting and add context to individual events you’re covering.

Originally published on Storybench

AI for Good: Five Key Principles for the Pro-Social Sector

Wed, 26 Feb 2025 02:00:00 +0000

Artificial Intelligence (AI) is reshaping industries across the board, but its impact in the pro-social sector is worth digging into more to understand emerging norms and challenges. Our typical examples stem from years of work on “civic tech”, “public interest tech”, “data science for good”, and related efforts. Most of these focus on challenges like species conservation, disease outbreak surveillance, deforestation tracking. Who doesn’t want to save threatened zebras? I describe these as accepted good: problems that most people agree need to be solved.

But there are problems less people agree are for the common good. What about AI projects that support for gender equity? Or projects working against surveillance capitalism? Not everyone agrees that these are projects that support the common good (though I do). These could be collectively called AI for contested good, and that’s the concept I want to dig into. Specifically, what are some key principles and methods to use when designing AI projects that support work on contested good in the pro-social sector.

This post introduces that thinking and some related examples. This line of thinking stems from an invited talk I gave recently at the UTSA MATRIX consortium’s AI Spring School, which offered an opportunity for me to spend a little more time conceptualizing and formalizing thoughts about these issues for the highly technical audience.

I took a quick selfie from the stage while the audience was doing a quick chat about projects they would describe as “AI for good”

1. Center Underserved Issues

Instead of focusing on well-studied and understood problems, turn your attention to innovate in neglected areas. This presents an opportunity for significant social impact.

This is particularly relevant in the context of feminicide, where mainstream institutions fail to collect and analyze data effectively. As Catherine D’Ignazio reminds us in “Counting Feminicide”, “What isn’t counted doesn’t count.” Over the last few years she’s been working with groups that create “counterdata” on feminicide to fill gaps in official records on gender-related killings of women, girls (cisgender and transgender). The work of these civil society groups to produce datasets about this violence supports advocating for policy, enforcement, and garnering public attention.

Producing this data is hard, both logistically and in regard to the human toll. Global groups struggle to create datasets of gender-related killings with ill-fitting tech. This creates significant data gaps and emotional trauma. As one partner noted,

It is not easy to read ten cases of feminicide and put them on a table, disaggregate them, have to put a name, age, circumstances and all that detail, without it affecting you emotionally.

In response, I’m part of the the Counterdata Network that works with groups to create bespoke machine learning classifiers that score news stories, cluster them by events, and deliver via email/web. Over 30 groups use them daily to monitor human rights and civil rights violations via news reports. We co-designed the system to help, basically replacing google news alerts with a custom-built open-source data pipeline. The network is growing across multiple topics: feminicide, criminalization of pregnancy, business human rights violations, and more.

This type of work emphasizes the need to look for issues that are systematically ignored and listen to voices that are unheard or silenced. Innovate in neglected areas with higher potential for social impact.

2. Build With, Not For

Technology should not be imposed on communities but co-created with them to ensure relevance, impact, and empowerment. This participatory approach contrasts with the traditional top-down model of technology deployment.

I first heard this phrase “build with, not for” from Catherine Bracy during her time at Code for America. She used it to emphasize the idea of co-creating technologies with government and community, instead of creating top-down solutions. As technologists, even our language is embedded with the opposite approach – consider the ideas of “deploying” vs. integrating” tools.

One response to challenges around Individualized Education Programs (IEPs) in the U.S. present a good government-related example to learn from. IEPs are plans created by public schools to offer special education services to children in need of more specialized care. They are created collaboratively across the school system, teachers, and parents. These legally mandated plans are in use for around 8 million students across the US. However, many parents struggle to understand the services their child is offered in their complex IEP.

The AI-EP team used interviews to identify common concerns parents had about IEPs. Through interviews, the team identified key challenges and built an AI-powered translation engine and chatbot to help parents navigate complex documents in a personalized way. The features include translation, question answering, and more. This creates a new AI-supported method for helping parents understand a complex document that is written in legal language using complex terms.

Projects like this remind us to focus on participatory design and iterative solution that integrate feedback from the impacted community. Co-creation ensures relevance, impact, and empowerment.

3. Study Up

AI can be a powerful tool for interrogating power structures, flipping traditional methods to focus on the culture of power rather than the culture of the powerless.

Anthropologist Laura Nader’s concept of “studying up” is essential here—using AI to investigate those with authority rather than those affected by their decisions. She argued that the methods of traditional anthropology, which typically studied groups with less power in social structures, should instead be flipped to study mechanism of power like colonization. Chelsea Barabas’ work on bringing this idea to algorithmic fairness has been highly influential on my thinking here.

A striking recent example is the Landlord Tech Watch project. Landlords increasingly rely on automated screening tools that use criminal records, eviction histories, and credit scores to assess tenants—often reinforcing racial and economic discrimination. By simulating tenant reports, this project demonstrated how these technologies embed bias and provided advocacy groups with concrete evidence for policy change (So, 2022). They are studying the tools of those with power in the rental market to understand disparate impacts and work to fight against them.

This line of work studying up in AI systems identifies that we should prioritize identifying and interrogating power structures so that we can then flip traditional research methods to challenge systemic biases. We can leverage advanced AI tools to investigate those with power.

4. Look for Quick Wins

Does every problem need a complex AI solution? Often, small, targeted interventions offer more sustainable impact than complex, resource-intensive projects.

Journalists are often at the cutting edge of leveraging new technologies, and considering the ethical implications for their field… genAI is no exception. Their work serves as a model for building small, targeted, interventions that offer quick wins with lightweight AI integrations.

One compelling example is the Eva chatbot developed by the Paraguayan news outlet El Surtidor. Paraguay has many women in prison, or awaiting sentencing, for drug smuggling. The team wanted to open a door into the taboo issue, so they trained an AI chatbot based on interviews with women awaiting sentencing for international trafficking. Readers could interact with the bot to ask quesitons about their cases, how they got caught, and more (with appropriate guardrails). It quickly reached more than 15,000 interactions with the public. Eva helps reframe the public narrative around drug-related crimes by challenging stereotypes and stigma.

Quick-turn projects with tightly defined use-cases like this help us consider what kind of new stories with social impact we can create, focusing on lightweight, adaptable solutions that offer immediate impact. How can AI help you make impact on a small but well defined idea?

5. Consider Hidden Risks

GenAI technologies are built to look magical, and often produce good-looking results at first pass. Scholars are adopting them to support media research, without fully considering the very real social, financial, and ecological impacts.

More and more often I’m seeing genAI used for computational tasks that are well solved with regular ML: entity extraction, geographic ID, etc. There are a set of already defined problems that we don’t need genAI for. Ongoing work by academics and journalists are uncovering the true costs of the massive investment in LLMs. Like Bitcoin, with genAI there’s a very real future where we have blackouts because of computational power usage.

An example is a recent study I worked on with a team from The Scope. We wanted to analyze articles in the hyper-local news source to assess content diversity via geography, topic, and quotes. We could have used genAI for all of this, as others have, but for geography and entities we used existing off-the-shelf solutions because they are well validated, produce reproduceable results, are free to deploy and use, and are open source (I’ve helped build both). However, without well-performing and easy to use existing solutions for quotes, we decided to try the genAI black box. We iterated on prompts against a manually coded list of quotes and ended up using ChatGPT v3.5, which was easy to integrate with, and cost little ($6 total). Quote extraction and attribution (via pronouns) worked well in the end, allowing us to do gender-based analysis. That was only possible at scale thanks to ChatGPT. In the end we were able to do a cross-sectional analysis along those three dimensions of content to help the editors audit their own content, and plan for any changes with that data in mind.

Our work on that study shows how genAI integrations should be carefully designed to only be used where necessary. That can help reduce areas where results are not easily reproducible and justify work towards algorithmic interpretability to mitigate unintended consequences. Companies creating LLMs are deliberately obfuscating the hidden power and environmental costs, and don’t have a history of being good partners to the media industry.

The five principles outlined here—centering underserved issues, building with communities, studying up, looking for quick wins, and considering hidden risks—are foundational for using AI responsibly in the pro-social sector. If you’re interested in AI’s role in social impact work, I encourage you to explore these principles in your own projects. These kind of approaches and examples can help ensure that we model how AI technologies might serve as a tool for empowerment, rather than perpetuating existing inequalities.

Nebraska Data Art Jam: Beyond Visualization

Tue, 14 Jan 2025 02:00:00 +0000

I’m just back from delivering the keynote talk at the Data Storytelling Workshop and Data Art Jam, which brought together a fascinating mix of creative thinkers, from scientists to journalists to artists. Hosted by Dan Novy at the Johnny Carson Center for Emerging Media Arts, this interdisciplinary event highlighted new ways of communicating through data, bridging art, technology, and storytelling. Here’s a look at the insights and projects from the other speakers at the event:

Novysan introducing the event by showing how the R Studio hammer makes everything look like a nail, leading to cookie cutter data vis results

Dr. Laura Guertin’s Data Quilts

Dr. Laura Guertin from PennState Brandywine opened the event with her deeply personal and innovative approach to data storytelling. As a trained marine geologist, she had spent much of her career in STEM education. But it was an object from her grandmother—a Kenmore sewing machine—that set her on a new path. Laura began quilting and soon realized that quilts could tell stories, including those based on data. In 2018, she joined a team working on “coastal optimism,” an initiative that aimed to counter the often bleak narratives of environmental decline (for me this evoked a solutions journalism approach). Instead of focusing solely on doom and gloom, she highlighted solutions: open levees in Louisiana that balance storm protection with environmental exchange, and offshore Christmas tree pens used to reduce coastal erosion.

Dr. Guertin speaking about her data quilts

Inspired by these stories, Laura returned home and created “Stitching Hope for the Louisiana Coast,” her first data quilt. The response was overwhelming, and she began exploring more ways to translate climate data into fabric arts. This relates to now-popular trends like temperature scarves—knitted or stitched rows of fabric, each representing a day’s temperature—and expanded the concept to include themes like air quality and rainfall.

Her most recent work connected to a research expedition on the open ocean. Serving the role of education and outreach on the boat, Zoom interactions with kindergarteners about measuring cloud cover inspired a quilt about the related data they were producing on-board. She shared daily photos of the sky with the kids, and the kids collected their own data to compare. The quilt she later made when back on shore, using local South African fabrics, visually represented cloud cover with colorful trapezoids, connecting scientific observation to physical materials and local culture. Laura’s quilting projects demonstrate how creative approaches to data can evoke curiosity and make complex information more accessible to a wide variety of audiences. As she noted, quilts have a unique power to draw people in—they feel warm and inviting, not intimidating. Explore more of her work on Dr. Guertin’s website.

Matt Waite on Data Journalism with AI

Later, Matt Waite, a journalism professor at the University of Nebraska-Lincoln, took the stage with a compelling argument for using artificial intelligence in journalism—but only for the boring stuff. He described AI as “the dumbest intern you could have,” perfect for automating repetitive tasks while leaving the creative work to humans. During a recent sabbatical, Matt explored how generative AI could support data journalism. His project focused on using R to integrate with Google Gemini, automating the summarization of county-level data to create personalized, map-based visualizations. He emphasized how this could make journalism more efficient, particularly for local stories.

Prof. Waite showing a demo of his work in R Studio on data journalism augmentation with AI

However, as he demonstrated, genAI isn’t foolproof. When he added a fact-checking bot to audit the summaries it produced, it caught errors, such as incorrectly describing a zero value as “no change.” Asking it to add “local context” to headlines introduced even more opportunities for inaccuracies. For Matt, these missteps underscored the importance of starting small with AI. Rather than aiming for big, transformative projects, he advocated for using AI to handle the mundane tasks that free up human journalists to focus on what truly matters: creativity and critical thinking.

Robert Twomey on Data and Art

The final presentation of the morning came from Robert Twomey, an artist and engineer whose work explores the intersection of data and art. Robert shared how his background in both science and painting informs his practice, with a focus on observation and representation. He drew inspiration from his MFA advisor, Natalie Jeremijenko, and projects like the “Suicide Box,” which used motion detection to document unquantified data on the Golden Gate Bridge. Robert pointed to how Jeremijenko’s work asks thought-provoking questions that are still core today: What counts as data? How can art challenge the ways we encounter and understand it?

Prof. Twomey connecting the dots between traditional data representation and painting as observation

He showcased projects that ranged from data physicalizations to generative art. One example was Sabrina Raaf’s Translator II: Grower, which visualized CO₂ levels by “painting” them on a gallery wall. Another, Joel Ong’s Aeolian Traces, used human migration data to generate fan-driven wind in a physical installation. Robert’s own work included a chatbot he created in 2005 based on interactions with his grandmother, who had been diagnosed with Alzheimer’s. The chatbot became an interactive data portrait, offering a poignant exploration of memory and technology. In another project, Garbage Cubes, Robert compressed physical artifacts into sculptural representations of “irreversible data compression,” challenging traditional notions of how we store and interpret information.

Throughout his talk, Robert emphasized the role of artists as experts in making things experiential. He pondered genAI’s ability to have explanatory power at all.

My Own Take on “Data Beyond the Visual”

As the opener before these brilliant folks, I took the opportunity to challenge the audience to rethink how we engage with data in community settings. I laid some groundwork by focusing on how now-traditional data visualization methods, developed in fields like science and business, often fail to meet the needs of pro-social domains like museums, libraries, and grassroots activism. These settings require tools that prioritize participation, empowerment, and justice.

Pulling from my book, I highlighted examples like participatory data murals and civic data sculptures, to demonstrate how creative approaches can reflect data back to communities, fostering deeper connections and driving action. I also previewed work on Data Drums, a climate change sonification composed for a Brazilian drum troop, which is set to debut this spring (more to come on here soon).

Through evocative examples and practical insights, I urged the audience to go beyond default charts and graphs, activating the latent creative capacity within communities to make data work more inclusive, impactful, and joyful.

Those talks set the scene for a hands-on 2nd half of the day, where groups spent hours designing and building data sculptures with copious amounts of craft materials, large and small, that Novy had supplied. The open space made collaboration easy, and groups dove in with impressive energy to the new activity.

Collage of teams working on their data sculptures.

Overall, the Data Art Jam was a testament to the power of interdisciplinary collaboration. From quilts that tell stories of resilience, to AI tools that simplify journalism, to art that questions the very nature of data, the event showcased the endless possibilities for reimagining how we communicate information. It left attendees inspired to experiment, adapt, and push the boundaries of what data storytelling can achieve.

Note: this post was drafted from my personal notes into a narrative using genAI tools and then reviewed and editing by hand

Journalism papers at IEEE Vis 2024: selected highlights

Fri, 03 Jan 2025 05:09:00 +0000

IEEE Vis is the premiere global academic computation-focused visualization conference. Each year about a thousand people attend, presenting around 100 papers. Some of those touch on data visualization and storytelling in journalism settings, bringing deep technical thinking and research into a domain that connects to the practicing newsroom. In this post I thought I’d summarize a few selected papers I saw at IEEE Vis 2024 that might be relevant to data visualization experts and data storytellers who work in journalism.

Structural influce model from Fu & Stasko, 2024

In “More Than Data Stories: Broadening the Role of Visualization in Contemporary Journalism”, Yu Fu and John Stasko identify ways data visualization can expand its impact in journalism beyond storytelling. They helpfullly lay out a research agenda to push other academics to work at this intersection. It might be useful to data journalists to look at their list of newsroom values data visualization can support, and offer some language for working with academics (like me) to develop co-created research projects.

An example Brazilian news data visualization Brito & Ferreira, 2024

In “Visual and Data Journalism as Tools for Fighting Climate Change” authors Emily Brito and Nivan Ferreira use media coverage of the terrible floods in Rio Grande do Sul (Brazil) as a case study in how data journalism can be a tool for fighting disinformation and climate change. Key elements they focus on include pursuasive communication, risk communication, and challenges realted to literacy and misinformation. Their paper offers valuable insights and examples for data journalists working on climate catastrophes.

A visual representation of the protocol used in the study by Rogha et al, 2024

In “The Impact of Elicitation and Contrasting Narratives on Engagement, Recall and Attitude Change with News Articles Containing Data Visualization” authors Milad Rogha , Subham Sah, Alireza Karduni, Douglas Markant, and Wenwen Dou explore the “you draw it” approach to interactive data storytelling. They experimentally explored the impacts of asking readeres for their opinions on attitude change, recall and overall engagement. They found that while showing showing contrasting visual data narratives did increase engagement, they didn’t improve recall nor change attitudes more. This kind of experimental research is critical to understanding how to build more effective data visuals to engage and impact the public.

Intro slide about Wang et al, 2024

The paper “Visualization Atlases: Explaining and Exploring Complex Topics through Data, Visualization, and Narration”, by Jinrui Wang, Xinhuan Shu, Benjamin Bach, and Uta Hinrichs, offers a taxonomy of “atlases” of data. This might be particularly useful to data journalists creating exploratory interfaces that let the reader browse through large datasets related to (more explanatory) stories. They found that atlases tend to combine data, narration, and structured navigation to make complex topics accessible. Speficially, many use design templates and careful data curation for consistency and clarity in visual storytelling. Their catalog of these templates and approaches in the paper might help others who are launching projects that create atlases like they describe.

I’m sure there were others that have relevant findings for folks working with visualizaton in newsrooms, but these were the ones that jumped out to me.

Talk Summary: Dr. Vishwanath on Science Communication & Misinformation

Wed, 18 Dec 2024 05:09:00 +0000

Notes from a talk By Vish Vishwanath (Harvard School of Public Health; Dana-Farber Cancer Institute), hosted by David Lazar and the Internet Democracy Initiative, December 18, 2024.

The surge of misinformation in science and health communication is not just a problem of individuals making poor choices or failing to fact-check. It’s a systemic issue deeply embedded in the structures of our modern information ecosystem. In his talk, Vish Vishwanath, chair of the National Academies committee on “Understanding and Addressing Misinformation in Science,” argued that we need to move beyond individual blame and examine the structural forces that create and amplify misinformation. Here’s a quick summary of how he unpacked the problem—and what we can do about it.

Vishnawath speaking at the event

The Problem: An Overloaded Information Ecosystem

The modern information landscape is overwhelming. With more channels and a greater volume of content than ever before, even science communication and translation experts struggle to make sense of the noise. This crowded ecosystem provides fertile ground for disinformation and misinformation—false information spread with or without malicious intent. Vishwanath broke down the key players shaping today’s science and health communication landscape:

Journalists: Traditional media outlets, along with local and minoritized press, play a major role in communicating science. The minoritized press, Vishwanath noted, often takes on an advocacy role, contrasting with the gatekeeping tendencies of mainstream media.
Activist groups and government officials: These actors produce content to inform, persuade, or advocate, often serving specific agendas.
Social media platforms: Platforms like WhatsApp allow anyone to reach massive audiences, often through viral video content. These low-barrier tools can amplify both accurate and false information.
The private sector: Commercial industries, like tobacco or sugary beverages, spend billions to shape public perceptions, often promoting narratives that align with their profit motives. For instance, tobacco companies spend $10 billion annually on marketing, creating a communication imbalance compared to public health efforts.

Amid this sea of content, making informed choices about what to trust requires effort, which many individuals may not have the time, tools, or skills to devote.

Misinformation spreads rapidly in this interconnected world, driven by factors such as:

Social media algorithms that amplify sensational content, regardless of accuracy.
The rise of AI-driven deepfakes that blur the line between fact and fabrication.
Declining trust in institutions, which leaves people more susceptible to alternative, often incorrect, narratives.

One illustrative example Vishwanath shared related to cancer prevention. Despite robust evidence that lifestyle changes can reduce cancer risk, many people believe that cancer is inevitable and that “everything causes it.” These misconceptions are perpetuated by social media spirals, where inaccurate ideas are reinforced and mainstreamed.

Inequalities in Health Communication

Structural influce model from a related academic paper (CC BY 4.0)

Inequalities in how people access, interpret, and act on health information are deeply rooted in education, race, and socioeconomic status, with significant consequences for public trust and health outcomes. Education levels and racial identity intersect to shape science literacy, with certain groups, like white evangelical populations, being among the least likely to get vaccinated due to cultural and informational divides. During the COVID-19 pandemic, reliance on online tools for vaccine scheduling highlighted the digital divide, leaving those without internet access at a disadvantage. Marginalized groups are often labeled as “hard to reach,” but Vishwanath reframed this as “hardly reached,” pointing to systemic failures in outreach and inclusion. These data gaps and inequities perpetuate disparities in health communication, reinforcing a cycle of exclusion and misinformation.

Structural Solutions: Rebuilding Trust Through Community

Vishwanath called for a shift from individual-focused solutions to systemic, community-driven approaches to combat misinformation effectively. By adopting an ecological perspective, efforts should target structural barriers such as limited access to technology and educational disparities, rather than placing the onus solely on individuals to “read critically.” Community-engaged research emerges as a key strategy, involving collaboration with communities to co-define problems and co-develop relevant solutions. Examples of this approach include the MassCONECT Portal in Massachusetts, which provides localized resources and training to help communities adapt public health messages, and SANCHAR in India, which has trained over 200 journalists to critically assess and communicate scientific information. These initiatives illustrate how building structural and community capacity can amplify accurate, evidence-based science communication.

Ultimately, Vishwanath called for more participatory and inclusive approaches to science communication. Misinformation isn’t just an individual problem—it’s a structural challenge that requires systemic solutions. By leveraging community knowledge and resilience, we can create alternatives to misinformation and foster trust in evidence-based information. Vishwanath’s talk highlighted the urgency of rethinking how we produce, share, and consume science information. Only by addressing the root causes of misinformation can we build a healthier, more equitable information ecosystem.