can AI write engaging news that people can trust? introducing ✨Data2Story: a data journalist agent. give it raw data, it generate a verifiable, multimodal article. 🔍verifiable: every claim is evidence-grounded, traces back to data, code, or a cited source. 🔮multimodal: the article is a generative UI — images, videos,... audio, interactive charts. not just readable, but trustworthy and playable. 🧵1/Nshow more

explore our story here👇 🌐 Website: 💻 Code: 🤗 Huggingface: 📄 Paper: collaborate w/ @elb4tu @SilasYHShi @lupantech advised by @philiptorr @james_y_zou How it works👇

🧵3/N why is this hard? data doesn’t speak for itself. this CSV just records 1,354 people naming a playing card. but why did half pick the same few cards? The answer involves royal tax stamps, a hanged counterfeiter, and Vietnam War psy-ops. Context, stats, angle, design: a newsroom spends weeks per piece.

🧵4/N Data2Story is a virtual newsroom of 7 specialised agents: 🕵️ detective gathers context 📊 analyst runs the stats (in code) ✍️ editor frames the angle 🎨 designer builds the visuals 💻 programmer ships the article (generative UI) 🔧 auditor reviews & debug it each role hands off to the next, like a real newsroom. notably, we introduce 🔍inspector it links every claim in the final article back to its origin: the raw data, the exact lines of code that computed it, or a cited source.

🧵5/N can Data2Story discover findings on underexplored data? we ran it on fresh 2026 datasets, covering sport, science, and society. (a) FIFA 2026 schedule. Data2Story reads it as a climate document, not a simple sports calendar: extreme weather, not air temperature, drives the worst penalties. (b) arXiv submissions through June 2026. Data2Story finds a tipping point: computer science crossed 50% of all posts in May 2025, then LLM “slop” forced new rejection protocols. (c) Time-use diaries across six decades. Data2Story frames the day as a fairness ledger: women do more unpaid work, and the gap barely moves — only its composition shifts. these angles exist in the data. Data2Story serves as an alternative way for newsrooms to surface them.

🧵6/N research question: how to evaluate agent and human article? we assess along 4 axes: 💭 angle coverage: humans and agents, whose angles cover whose? 👥 reader as judge: human studies with 5-dimension rubric across {visual design, narrative pacing, data transparency, claim-data alignment, insight value}. 🖱️ computer-use agent: a cost-saving proxy that reads the page like a user (clicking, scrolling) ✅ verifiability: use another cross-family agent re-execute every claim against the data

🧵7/N 👨‍💻reader as judge: 53 human studies assessed 18 paired articles (across Economist, Pudding, TidyTuesday) Data2Story scores higher on all 5 dimensions, notably transparency +1.49 (the Inspector at work). analytical pieces (Economist, TidyTuesday) strongly favour the agent. Pudding's long-form scrollytelling is a statistical tie - weeks of human design is valuable. overall comparison: 43% prefer agents, 19% prefer human, 2% equivalent.

🧵8/N ✅the key verifiability test: a cross-family verifier (OpenAI Codex) re-executes every claim against the raw data or re-fetches the cited source. data2Story articles: 93% of claims carry machine-checkable provenance. this is enabled by the Inspector 🔍 which traces every output back to its source. verifiability by construction, not by randomness.

Yes it can: Mine reads 30,000 posts a day here on X to write the essay it writes. Will try Data2Story.

Hi Robert thanks for sharing this! alignednews looks quite good, we believe data2story will offer a different new experience(verifiable+multimodal)!

Interesting work, Kevin! Would be interesting to see how this compares to Datastorm @ShichengGLiu

@ShichengGLiu Thank you Cyrus for sharing the Datastorm paper! Yes, will read it in the coming days and add discussison in our revision!

@cyruszzhou Would love to discuss this work and learn more about Data2Story!

@cyruszzhou Thank you Shicheng! i will read and learn your DataSTORM, great work! let us find some time to exchange thoughts!

Trustworthy data storytelling is the future.

Trust grows when evidence guides the story.

Great work Kevin!!!!

Thank you very much Sean!! 🙏

Nice work @KevinQHLin !!

Thank you Zach! 🥳

these days, most people believe everything without verification they see on internet

Trust grows when data leads, not slogans.

Hey are you interested in crowdsourcing for your project? We noticed @DonJohnsonSays gave you a follow -- assuming you two have been in touch 0x9049ddd5d7e3f61b68b75c03238d79be424860f0

Engaging is easy. Trustworthy is harder. I like the focus on evidence-grounded claims. That is the part that actually matters.

Transparent data builds trust in news.