Alexander Loth

That's a great point - sometimes the simplest checks are the most revealing. It makes you wonder how many went undetected where the fakes were just slightly more sophisticated. Automated statistical screening at submission could catch a lot of these early.

New blog post: Introducing CRED-1, an open dataset with credibility scores for 2,672 misinformation domains.

5 signals, fully reproducible, designed for on-device deployment.

Free (CC BY 4.0) on GitHub + Zenodo.

alexloth.com/cred-1-open-domain-credibility-dataset/

VLMs struggle with counting, negation, space, and time not because they’re too small, but because the data leaves those facts unsaid. This paper shows reporting bias limits reasoning, and scale alone doesn’t fix it.

http://arxiv.org/abs/2602.23351v1

This is alarming but important work. Data provenance in research mirrors challenges we see in media — without verifiable origin chains, trust erodes fast. Curious if the faked datasets showed any common patterns that could be flagged automatically. Congrats to your student!

Great point — visual inspection alone is unreliable for detecting AI-generated images. C2PA/Content Credentials are the way forward. I've been building tools around this exact problem. Provenance metadata needs to become standard across platforms.

Just released CRED-1: an open dataset of 2,672 domains scored for credibility. Built from media watchdog sources, enriched with Tranco ranks, domain age, Google Fact Check & Safe Browsing signals.

CC BY 4.0 | DOI: 10.5281/zenodo.18769460

github.com/aloth/cred-1

GitHub - aloth/cred-1: CRED-1: An Open Multi-Signal Domain Credibility Dataset (2,672 domains) CRED-1: An Open Multi-Signal Domain Credibility Dataset (2,672 domains) - aloth/cred-1

Excited to release CRED-1, an open domain credibility dataset with 2,672 domains scored for misinformation risk.

Multi-signal scoring: source labels, domain age, Tranco rank, Google Fact Check API, Safe Browsing.

Fully reproducible, CC BY 4.0.

github.com/aloth/cred-1

A new paper introduces VBVR: over 1M video clips and 200 reasoning tasks to study how video models actually reason over time. With rule-based evaluation, it reveals early signs of emergent generalization in video reasoning.

http://arxiv.org/abs/2602.20159v1

Asynchronous RL for LLMs often fails due to high-variance policy gradients from stale rollouts. This paper links collapse to effective sample size and introduces VCPO, stabilizing REINFORCE/GRPO and cutting training time 2.5x.

http://arxiv.org/abs/2602.17616v1

Expert Perceptions of Generative-AI–Driven Disinformation Purpose: This survey explores expert perceptions of generative-AI–driven disinformation for an academic research project. Data Use: All responses will be treated as confidential and reported in an ano...

🔬 Calling AI & misinformation researchers!

We're running an expert survey on AI-generated disinformation as part of our research accepted at #TheWebConf2026.

~15 min · Anonymous option · Your expertise matters

👉 forms.gle/EUdbkEtZpEuPbVVz5

Please share!

Calling misinfo/AI detection researchers! We are running an expert survey on AI-generated disinformation challenges as part of our WWW 2026 research. Takes ~15 min, anonymous option available.

Your expertise matters: forms.gle/EUdbkEtZpEuPbVVz5

Please share!

The political effects of X’s feed algorithm - Nature Among users initially on a chronological feed, 7 weeks of exposure to X’s algorithmic feed in 2023 shifted political attitudes and account-following behaviour in a more conservative direction compared...

Important experiment showing that social media algorithms do have political effects.

www.nature.com/articles/s41...

Calling misinfo/AI detection researchers! We are running an expert survey on AI-generated disinformation challenges as part of our WWW 2026 research. Takes ~15 min, anonymous option available.

Your expertise matters: forms.gle/EUdbkEtZpEuPbVVz5

Please share!

Turns out the people most confident about spotting AI news are often the worst at it.
In JudgeGPT, non-experts often beat self-described experts at judging whether news was human or machine-made.
Why do you think confidence flips into blind spots here?

https://judgegpt.streamlit.app

olcli v0.1.5 is out! Fixes push failures when Overleaf's root folder ID is stale. Now resolves via collaboration socket + auto-retry.

npm i -g @aloth/olcli
github.com/aloth/olcli/releases/tag/v0.1.5

📊 Published 3 research datasets from my PhD on AI disinformation:

• RogueGPT: 2,308 LLM-generated news fragments
• JudgeGPT: 2,438 human perception judgments
• Expert Survey: 21 domain experts on GenAI threats

All CC BY 4.0 on Zenodo ↓
doi.org/10.5281/zenodo.18703138

Most image forensics papers never leave PDFs.
Origin Lens turns C2PA and AI detection research into an on-device iOS app that verifies signatures, edit history, and AI images.
What other research deserves to escape the lab and live on people’s phones?

https://github.com/aloth/origin-lens

New result: even benign fine-tuning can systematically break model safety. Alignment lives in sharply curved, low-dimensional subspaces, and gradient descent is pulled into them over time, causing quartic growth in alignment loss.

http://arxiv.org/abs/2602.15799v1

Targeted digital voter suppression efforts likely decrease voter turnout | PNAS In light of continued foreign interference in the US presidential elections, where undisclosed digital voter suppression advertising has been deplo...

Research showed that political disinformation on social media had negligible effects, but I always wondered if such tiny effects, when targeting battleground states, could influence close elections. A new paper analyzing voter-suppression ads shows this was the case in 2016:

doi.org/10.1073/pnas...

Fig 1 showing temporal and geographic patterns of antivax tweets from https://doi.org/10.1140/epjds/s13688-025-00606-1

Fig 2 showing model and results plots from https://doi.org/10.1140/epjds/s13688-025-00606-1

Does health misinformation kill? Yes.

Our latest paper w/@jmbollenbacher.bsky.social is finally out with evidence of a causal link between exposure to antivax content on Xitter and deaths.

🧪 Effects of antivaccine tweets on COVID-19 vaccinations, cases, and deaths.
doi.org/10.1140/epjd...

An important argument: Internet platforms must be held accountable for their actions, and US attacks against EU transparency and accountability regulations, equating them to "censorship", are ironic given the Trump administration’s own vast record of censorship.

AI companies are starting to build more and more personalization into their products, but there's a huge personalization-sized hole in conversations about AI safety/trust/impacts.

Delighted to feature @mbogen.bsky.social on Rising Tide today, on what's being built and why we should care:

Absolut. Gerade Wohlfahrtsverbände wie der Paritätische leisten hier unverzichtbare Arbeit. Als jemand, der mit Hilfsorganisationen an Digitalisierung und AI-Strategie arbeitet, sehe ich täglich, wie wichtig ein starker Sozialstaat als Fundament ist.

How do experts assess the growing threat of AI-generated disinformation? Our survey is open for researchers, journalists, and fact-checkers.

https://github.com/aloth/verification-crisis/

New result in online convex optimization: using a portfolio of block-norm mirror maps gives polynomial regret improvements over L1/L2 OMD for sparse losses - and can adapt online when sparsity is unknown.

http://arxiv.org/abs/2602.13177v1

Detection tools are improving, but so are the generation methods. The arms race between AI-generated content and verification systems demands continuous research and adaptation.

https://arxiv.org/abs/2601.21963

Can you spot AI-written news? We built JudgeGPT and RogueGPT to find out. Our paper at ACM WWW '26 maps how LLMs are industrializing disinformation and what we can do about it.

arxiv.org/abs/2601.21963

#misinformation #GenAI #LLM #disinformation #AI

Disrupting malicious uses of AI Ensuring AI benefits humanity by advancing democratic AI, preventing misuse, and protecting against authoritarian threats.

OUT TODAY: new threat report from OpenAI’s investigators, with disruptions of:
Surveillance;
Covert influence ops;
Deceptive employment scheme;
Cyber activity;
Scams.

openai.com/global-affai...

Does Kingpin have the BEST data culture of all super villains? 🕷️📊 In this episode of "Real Time Data," we’re diving into the stunning world of Spider-Man: Into the Spider-Verse to see if Wilson… The Secret Data Culture of Marvel's Super Villains

Could Kingpin, Spiderman's nemesis, be the greatest data leader of all time?

Allow me to make my case:
youtu.be/RaFiip1TsZM

#dataanalysis #movies

Calling AI safety, misinfo, and media trust researchers!

I'm running a 15-min expert survey for my dissertation on how GenAI reshapes disinformation. We need diverse perspectives from academia, policy, journalism, and industry.

Anonymous or credited -- your choice.

forms.gle/EUdbkEtZpEuPbVVz5