This was tbd/con

tbd/con was an interdisciplinary virtual gathering about AI.

Sept. 23 was the conference with sessions selected and curated by the review crew: Information and Practice run through the morning, Thought and Garage through the afternoon, and Information carries on past lunch (ET).

Sept. 24 was a self-organized unconference.

The circles

• Information as in the utility, influence and infrastructure of what we know and believe.
• Practice as in what people are actually building, shipping and abandoning in newsrooms right now.
• Thought as in the concepts, vocabulary and frameworks we use to reason about all of this.
• Garage as in unfinished work, prototypes and things that might not survive contact with an audience.

Information circle

Measuring model bias without smuggling in your own

When someone asks an AI model about a politically sensitive event, how much does the framing of the question shape what they get back? We'll share a methodology built with ASL19 in Persian and extended into Venezuelan Spanish, testing six models on Iran, Ukraine, China and the far right, in several languages and framings. You can measure which way a model leans. You cannot settle what counts as "bias" without importing your own politics. Scoring is deliberately unfinished. Expect a live demo and an invitation to stress-test the rubric.

Valentina Aguana is a Venezuelan activist and systems engineer working at the intersection of human rights and technology. She is a 2026-2027 Mozilla Fellow working alongside Bellingcat on cybersecurity challenges affecting OSINT investigators. She is also a researcher at Conexión Segura y Libre, where she worked on Noticias Sin Filtro, an anti-censorship newsreader app, and runs digital security training for activists, journalists, and civil society.

Callum is CTO of a censorship circumvention project supported by the Open Technology Fund, providing safe, anonymous access to online platforms for at-risk groups, researchers, and journalists. Previously head of machine learning at an NGO, he now researches the impact of LLMs on information flows in civil society.

From personalized distribution to personalized production

Recommendation engines have mastered personalized distribution, but creator incentives still push toward homogeneous content, which leaves long-tail and deeply personal needs unserved. As generation matures, information systems shift from distributing what already exists to producing what does not: anyone can ask an AI to keep generating a feed built for them, indefinitely. A live demo of an experimental platform built on that premise, plus open questions and an invitation to poke holes in the idea.

The speaker will be introduced on the day.

The power of real-time censorship in Tibet

In February 2022, the Tibetan singer Tsewang Norbu set himself on fire in front of the Potala Palace, the most visited tourist site in Lhasa. Millions pass through it every year, and his act was likely seen by many. Yet, to this day, no video or image of the incident has surfaced publicly. This talk explores how Chinese real-time censorship erases events as they happen, and what it means when technology can shape discourse in real time.

Nithin Coca is a freelance journalist who covers the social, labor, and human rights implications of technology, focusing on China and emerging economies in Asia. Coca's feature reporting has appeared in global media outlets including Al Jazeera, Quartz, Engadget, Foreign Policy, Vice Motherboard, and Coda Story. He is currently based in Japan.

How to develop journalism products in an AI world

As AI reshapes digital discovery, traditional news product models face a crisis of relevance and trust. Is it possible to build audience-centered, ethically sound news products without relying on Silicon Valley’s platform dominance? This session explores the emerging tension between traditional editorial integrity, business sustainability, and AI-driven personalization, opening up a discussion on what product frameworks should—or shouldn't—look like in an AI-first media landscape.

Temitayo Akinyemi is a product strategist and media innovator focused on the intersection of journalism, technology, and user experience. He specializes in helping news organizations design sustainable, audience-centered digital products, guiding newsrooms to navigate algorithmic shifts, preserve editorial integrity, and build trustworthy information ecosystem models in an AI-driven era. He currently works as Head of Product and Growth at thegreenline.to, a hyperlocal newsroom in Toronto.

Precision or usefulness: what happens when an AI assistant refuses to guess

We built a conversational AI for tiny newsrooms that would rather say "I don't have that data" than hallucinate an answer. That choice raised a harder question: how much should an AI explain itself before caution starts working against the person it's meant to help?

Juan Pablo Garnica is a research and data-structuring analyst at Datasketch. He proposes database structures that make information usable for community organizations, civil society groups, and public agencies, and produces the diagnostic and documentation work behind those projects.

Juliana Galvis is Projects Director at Datasketch. Since 2017 she has worked across data journalism, open government, and responsible AI, leading projects with newsrooms, civil society organizations, and public agencies, and managing their scope, timelines, and client relationships.

Transformers and Decepticons: should we dehumanize LLMs?

The language we use to talk about a technology influences our perception of the roles it can play and how it should be regulated. LLMs are anthropomorphized by their creators, but might our societal discussions about them benefit from dehumanizing language?

Cecilia Dobbs is an independent product, strategy and audience consultant based in Wisconsin, US. She has 20+ years of experience and holds master's degrees in science journalism and social work. She is passionate about bringing insights from both fields to her work with news organizations.

Building Big Tech alternatives for news

How might we confront AI hype and doom narratives and create specialized, people-centered technologies as an alternative to power-concentrating Big Tech rhetoric and “solutions”? How do we do this at a time when Big Tech is concentrating power and wealth while the news industry is struggling to find profitable business models? As Big Tech builds its empire and captures more of the news industry, it becomes even more important for us to discuss how we build partnerships to hold Big Tech accountable.

Joanna S. Kao leads the Pulitzer Center's AI Accountability Network, where she oversees a portfolio of AI and machine-learning reporting projects and aims to foster a growing, global community of journalists doing AI-related reporting. She is based in the UK.

Are agentic dark patterns baked in?

Whatever a prompt asks for, the model produces a version of it that its training data makes most probable. The dark patterns emerge from that interaction, not from the prompt or the data alone. We'll evaluate a live agentic surface and ask what it means to choose when the choices are all generated.

Philliph Drummond is a Senior Tech Design Researcher at Superbloom Design, where he leads privacy and security research for tools used by high-risk communities. His work spans threat-informed design, human-centered threat modeling, and the integrity of the interfaces people encounter when they are most exposed. 

Can AI help journalists uncover stories hidden in public geospatial data?

Public geospatial data holds countless investigative stories, yet remains inaccessible to most journalists. I'll demonstrate an experimental prototype exploring whether AI can lower those barriers by revealing patterns and generating investigative hypotheses without replacing editorial judgment. Expect a live demo, open questions, and an invitation to stress-test the idea.

Tainah Ramos is a Brazilian journalist, geography student, and geospatial technologies researcher exploring the intersection of journalism and geospatial data. She is developing Territórios Descobertos (Uncovered Territories), an experimental project that uses public spatial data to help journalists uncover stories that might otherwise remain hidden.

Who said that, and can you prove it? Inside the Provenance Bridge

Fact-checking asks one question: is a claim true or false? Provenance asks a different one: where did it come from, and does it faithfully represent its source? I'll walk through the Provenance Bridge, built by The Newsroom with the BBC and Stanford's Starling Lab, a system where a published claim can be cryptographically sealed and traced back to its canonical source, even through paraphrasing and edits. The seal runs both ways, and breaks if the source changes or the wording of the reporting drifts. Expect a look at how it works, open questions around scaling it into newsrooms and beyond, and space to poke holes in it.

Jenny Romano is Co-founder and CEO of The Newsroom, working at the intersection of AI, information provenance, and accountability. She co-led development of the first AI text provenance framework aligned with the C2PA standard, created with Stanford's Starling Lab and the BBC, and has trained 2,000+ journalists and media executives on navigating AI at institutions including LSE and Oxford.

When AI comes to town: how should journalists cover communities pushing back?

As AI infrastructure expands, communities are pushing back over data centers, energy and water use, environmental impacts and who gets a say in what gets built where. How should journalists cover these actions locally, without removing the broader context or flattening the story into a pro- or anti-AI debate? We’ll explore how a solutions journalism lens might help reporters investigate what communities are trying, what evidence exists, where those efforts fall short and what can be learned across communities. We’ll also wrestle with where that framework has limitations and invite participants to help identify better questions and approaches for covering this emerging story.

Jaisal Noor is Program Manager of Training and Curriculum at the Solutions Journalism Network, where he supports newsrooms in transforming their coverage of politics and civic life through rigorous reporting on how communities are responding to their most pressing problems, or “hope with teeth".

Practice circle

Archipelago Alliances

AI is not just another platform, and agentic workflows and protocols mean that journalists, their outlets, and their readers can all interact in new ways without traditional middlemen. This session explores what it can look like when different media outlets unite in distributed systems to reach their audience and make revenue. 

Christopher Brennan is a journalist and technologist with bylines from a dozen countries including for the BBC, NY Daily News and Moscow Times. He co-founded Overtone in 2021, before AI was cool, because he believes new forms of tech can make the internet a more human place.

If not the article, what?

We keep saying the article is dead. So what do we make instead? We'll demo a live prototype: Useful data journalism served by an agent that people and other agents can query, subscribe to, and hand work off to. Do we own this future, or does Silicon Valley?

Scott Klein is an award-winning journalist and innovator who helped establish data-driven news applications as a form of journalism, and now builds tools connecting that work to an era of AI agents.

Duy Nguyen is an AI scientist who pilots original research to reimagine reporting capabilities inside newsrooms. He's now turning that same work toward producing journalism for the agentic future.

Lessons from AI product failures

Seamus will briefly explain his AI for news product failures and invites you to bring your failures (or successes!) to share too. Seamus' failures include an AI obituary writer, an AI story writer and community listener, and, most recently, an AI site editor for WordPress. 

Seamus is a software engineer with over a decade of experience, almost entirely in education technology. He recently took a year detour attempting (and failing) to create useful products for local news organizations. 

How research organizations evolve in the age of AI

Artificial intelligence is altering the pace and mechanics of organizational research, forcing insights leaders to reconsider team structures, core methodologies, and strategic impact. This session examines the practical shifts underway within research divisions: where AI effectively speeds synthesis, how to maintain qualitative rigor amid automated data streams, and what structural changes allow insights teams to directly influence high-level business strategy.

Marta Rey Babarro, PhD, is an executive advisor and strategist specializing in research architecture, innovation diffusion, and human-centered AI integration. Previously VP of Research & Insights at Zillow and Global Head of UXR at Google, she partners with executive teams and companies to scale research operations, redesign organizations, and turn human insights into measurable business growth.

Building a design system at the speed of news

We'll walk through what happened when we tried to use AI to prototype, automate, and scale our design system, and what we're still figuring out. Come for the war stories, leave with a realistic picture of what implementing AI at scale into a design process actually takes.

Priscilla Huff is a senior software engineer at The Philadelphia Inquirer with 14 years of web development experience. Over the past year, she expanded her focus to include AI integration and automation, building intelligent systems and streamlined processes that enhance modern web development capabilities. 

Elizabeth Flynn, senior product designer, leads the design system at The Philadelphia Inquirer and designs the products that generate its revenue. She experiments with AI frequently, bringing fourteen years of web, brand, product, social, and advertising design experience and a generalist's range to zoom out on the system and in on the details.

What is the friction we want in the design and governance of AI systems?

How do we make pro-social outcomes easier (removing friction) and harmful or high-risk outcomes harder (adding friction)? There’s urgency to explore the full spectrum of practical frictional realities, interventions, and possibilities in AI design and governance. This session is an invitation to articulate, negotiate, and transform them together.

Bogdana (Bobbi) Rakova is an ex-Mozilla fellow, builder, researcher, organizer, writer, a research affiliate with the Data & Society Research Institute, and a data scientist on the AI innovation and governance team at a law firm. She integrates interdisciplinary perspectives, systems thinking, and a passion for weaving paradox into synergy.

The tech trust paradox

When online platforms fail targeted journalists and communities, trusting another tech tool, especially one using AI, can feel impossible. In prototyping an online abuse evidence-gathering tool, we faced key decision points around research, community agency, and platform independence. We’ll share the trade-offs we navigated as well as what we learned, then open the floor to ask: Can we intentionally build an AI tool using tech anti-patterns, and what would it take for you to actually trust it?   

Kalie M. Mayberry is a social impact technologist, building better community governance and activism practices online and offline. She has led research and programs at Harvard University’s Berkman Klein Center for Internet and Society, Columbia Business School and Wharton School of Business, and holds an MPA from the University of Pennsylvania.

Chelsea Johnson has been building software for over 13 years across industry, academia, and nonprofits. At Harvard University’s Berkman Klein Center, she works on online safety, mostly on the question of who it actually protects. Prior to Harvard, she served as director of technology at Tiny News Collective.

Should we look at how other industries are doing AI?

We're not the only ones struggling with LLM evals. AI opened bigger datasets to verify, and errors still cost us reader trust and reputation. Lawyers and investment bankers face the same math with better resources and clearer price tags on failure. What can we borrow from how they evaluate AI?

Teresa Mondría Terol leads AI products and platforms at NPR, using generative AI to enrich content metadata for better newsroom workflows and audience personalization. She previously built AI tools for investigative reporting on The New York Times' AI Initiatives team, and has worked at The Browser Company and Chartbeat.

Thought circle

Is humanity's end-goal unlimited satisfaction?

When humans drive the progress of technology, our individual concerns and preferences are what decide its shape. But, if we let go of the steering wheel and let technology improve itself, what shape will it take? Will we be able to force it to satisfy us? In this session, let me share a few perspectives before we reflect together on where technology may take us all.

Martin Heidegger is a freelance engineer with more than 20 years of experience in a variety of fields. Aside from having a name that invites questions about fundamentals, he has led OKTech, an international community exploring diverse topics in Osaka and Kyoto for the past 11 years.

AI metaphors we live by

We say things like “AI is my thought partner” or “AI hallucinated”, giving LLMs human-like properties where they have none. This human metaphor distorts what we can discuss about AI. What happens if we talk about AI as if it were soup, a material, or organized malice?

Aras Bilgen helps teams use pragmatic, humble design approaches for product development. The products he worked on have more than 160 million users worldwide. He is the author of Product Research Rules, a book about research in product teams. His computer science background helps him see through the AI hype.

Epistemic intermediation, from recommendation to generation

After thinkalouds with 53 regular AI chatbot users, we've been pondering what makes algorithmically generated synthesis different from algorithmically recommended links. We’re calling it “epistemic intermediation” - a combination of synthesis, conversationality, interface cues, and actionability. What are we missing?

Jay Barchas-Lichtenstein is the Senior Research Manager at the Center for News, Technology & Innovation. Before joining CNTI, they led media research at Knology, where they built participatory collaborations with news organizations across the U.S. to support audience learning about topics ranging from inflation to the U.S. carceral system.

Prabhat Mishra is a Senior Fellow at the Center for News, Technology & Innovation (CNTI). Prior to joining CNTI, he was a Visiting Fellow at Harvard University where he conducted interdisciplinary research on AI governance in China. He works on global AI governance, media innovation, and human-technology relations.

Gods in humans in machines in humans

Why fall in love with Claude and decide she's writing you love letters? Why treat an AI like an oracle? What is it about the texts LLMs produce, in collision with our beliefs about language, that makes us see and love—or hate—the “author” of the words we read?

Erika Alpert (they/she) is a linguist, anthropologist, writer, and professional busybody. Their research centers on how people make relationships through language and across media, including screens, human intermediaries, and algorithms. Take her advice and sit at a round table on your next date.

The AI disclosure dilemma: what journalists fear and what audiences want

As journalists increasingly use AI, audiences want transparency, accountability, and human oversight, but newsrooms often struggle to meet these needs. Grounded in audience research and newsroom experimentation, this session centers on what audiences are asking for and the fears holding newsrooms back while suggesting practical solutions for building trust while using AI.

Lynn Walsh is an Emmy Award-winning journalist and consultant focused on ethics, trust and the future of information. She is Assistant Director of Trusting News, an adjunct professor, small business owner and mom committed to helping people understand the information shaping their lives so they make informed decisions for themselves and their communities.

How do we talk to our friends who hate AI about how we use AI?

Join us for a facilitated discussion among practitioners about friends with mixed feelings about AI and navigating moments of distrust and tension - not to win skeptics over, but to sit with what that friction reveals about labor and extraction, and ask whether 'thoughtful use' is as legible a claim as we think.

Madison Karas leads product and research initiatives for independent media and startup newsrooms, from small shops to large operations, and is focused on building a better, more intentional future for news media for everyone involved. She co-runs The Service Desk at Gazzetta and is co-organizing tbd/con.

Determining the Future of AI

Join the audience for an interactive discussion with a pair of cutting-edge AI practitioners as they discuss the potential future of AI. David Kuszmar represents a skeptical, adversarial view of AI systems, while Matthew Gore-Kormanik offers a grounded vision of a future where machine learning genuinely improves human lives.

David Kuszmar is a researcher behind many cross-architecture LLM exploits, pioneered work into an ML-semiotic concept now known as pseudodiachrony, and generally known for his field-style approach to AI research. His work was recently featured in IEEE Spectrum, the world's largest engineering magazine.

Garage circle

Killer bots: what's really going on when AI gets hacked?

AI handed everyone the ability to build at a scale we have never had. It handed attackers the same thing. Automated attacks, drained accounts, supercharged social engineering. I will walk the real threat landscape, then run a live MCP hack on my own system so you can see one.

Matthew Gore-Kormanik (Zigula) is a security researcher, AI systems engineer, and co-founder of tbd/con. He was the hands-on operator on the Gemini guardrail bypass inside Fortnite, documented in IEEE Spectrum's August 2026 cover story. 

AI in the product vs. AI to create the product 

Most systems now ship AI creates, human reviews. I tried that on high-precision work, watched reliability fail, and rebuilt around deterministic tools. Does review catch residual error under volume? What does creator-to-reviewer do to people? How do you review what is missing? I lived a small version of this question.

Shawn Terry is an operations leader and AI builder. A non-developer turned product builder, he uses AI to design and build practical systems, including Mantis Crochet Notebook and an AI creative studio. His work focuses on where AI expands human capability—and where deterministic systems still matter.

The case for imperfect video journalism

This session explores why highly produced video isn't always what audiences want. David argues that newsrooms can build more trust, reach younger audiences, and create more video by embracing content that feels human, authentic, and intentionally imperfect.

David Rodin is a senior at USC and the founder of Editory, a video technology company helping newsrooms turn their reporting into authentic, social-ready video. His work focuses on how publishers can adapt to changing audience behavior while keeping journalists, not technology, at the center of storytelling.

Minority collapse

Jay has engineered a prompt that elicits a zero-shot response set from Claude Fable 5 showing role bleed, revealed training data, internal tagging, and guardrail bypass. He will share Minority Collapse live, launch a dynamic explorer for public use, and walk through the five triggers behind it.

Jay Gerbrandt is a Toronto-based Canadian entrepreneur and researcher. When Claude launched in Canada in 2024, he began casually red-teaming in his spare time, which led to him developing a zero-shot jailbreak that works on Fable 5.

Closing plenary

Everyone back in the lobby with info on the next day's unconference.

Sept. 24 agenda

A self-organized unconference, 9:00 to 16:20 ET.

More here:

that’s a wrap on tbd/con! 👾 unconference details for tomorrow

woohoo!

GazzettaG. Azzetta

If you have questions or suggestions, email us at tbdcon@gazzetta.xyz.