Can we steer visual representations like we prompt LLMs? This paper shows how to inject text into vision encoders via early fusion, creating steerable features that stay strong for core vision tasks while focusing on any concept you ask for.

Read the full paper: http://arxiv.org/abs/2604.02327v1

Most fake news generators still spit out text you can’t reproduce or compare.
RogueGPT ships a controlled pipeline - multi-model, multilingual, style-locked news with full provenance - beyond what GROVER or FACTGEN allowed.
If we can generate disinfo this precisely, how should evaluation tools change?

Full paper here: https://github.com/aloth/RogueGPT

#introduction

I'm Alexander Loth -- researcher & author in Frankfurt.

PhD: how generative AI reshapes disinformation & trust.

Tools I build:
🔬 JudgeGPT -- humans detect AI news at 52% (coin flip)
🤖 RogueGPT -- controlled misinfo stimuli
🔍 Origin Lens -- on-device C2PA verification
📊 CRED-1 -- domain credibility dataset

Also: books on data viz & AI, iOS apps.

https://scholar.google.com/citations?user=ofZZ8LgAAAAJ
https://github.com/aloth

#AcademicMastodon #GenAI #Disinformation #C2PA #WebScience #AIethics

People think they can spot AI-written news. Turns out they mostly can’t.
In a large human study, GPT-4 news was judged about as authentic as real journalism, with accuracy hovering near chance.
If readers can’t tell, what happens to trust when anyone can publish at scale?

Full paper here: https://github.com/aloth/JudgeGPT

We still verify images by squinting at pixels and vibes.
Origin Lens does on-device cryptographic C2PA verification, showing who signed an image and if it was altered.
When trust is math instead of guesswork, which would you rely on?

Try it out on iOS: https://apps.apple.com/us/app/origin-lens/id6756628121