Search

Exactly as you would expect. Grok favors Republicans, ChatGPT Democrats | mobilenet.cz

Translated title. Original: Přesně, jak byste čekali. Grok straní republikánům, ChatGPT demokratům | mobilenet.cz

The article reports that a Cornell-led study finds political bias in advanced AI models, with Grok 4.7 favoring Republicans and GPT-6.1 Sol favoring Democrats in most test scenarios, and suggests bias also appears in topic and source selection. It highlights differences in how bias is encoded and the potential influence of AI and tech leaders on voters.

Entryism — Olivia Guest

The piece argues that anti-AI positions can be forms of entryism driven by superficial judgments, equating correlation with cognition and linking this to racism and sexism. It highlights cases of AI misuse in scientific publishing and calls for accountability of powerful figures in AI entryism, while cautioning against blanket judgments of individuals.

ECHV in health: spectacular time savings and risk of bias

Translated title. Original: IA en santé : gains de temps spectaculaires et risques de biais

AI in health boosts diagnostic speed and reduces clinician workload but depends on high-quality, diverse data and rigorous clinical validation to avoid biased or erroneous decisions.

What you (don't) see is what you get

The piece argues that output bias in AI stems from politicised development, using DeepSeek as a case study of state-influenced censorship, while noting that bias is a broader, non-democratic and democratic phenomenon across models like Grok and Gemini. It contends censorship is systemic, embedded in training, model building, and real-time filtering, and that global concerns reflect wider debates about AI regulation and propaganda across regimes.

New Rules for Big Platforms – M – People Make Media (ver.di)

Translated title. Original: Neue Regeln für große Plattformen – M – Menschen Machen Medien (ver.di)

German states plan to require major platforms to allow links to journalism and give qualifying Public Value content a regular place in recommendation feeds, with a proposed minimum of one in five recommendations. The draft’s eligibility rules and automatic status for some media outlets remain under discussion.

[2610.31808] Patronizing and Condescending Language Detection in English and Mandarin

The document describes PCL as harmful but hard to detect, and presents two detection methods—CNN with static embeddings and fine-tuned BERT—finding BERT more effective with F1 0.528 on development data.

Can I ask AI if I’m sick? An expert explains what’s worth looking up and what calls for caution.

Translated title. Original: Posso perguntar à IA se estou doente? Especialista explica o que vale pesquisar e o que exige cuidado

AI can help people understand medical terms, general health information, and prepare questions for a doctor, but its answers may be biased or wrong. The article advises checking sources and not using AI as a substitute for professional diagnosis or treatment.

Intensive care in medicine: did she come to help treat or to hinder? | VEJA RIO

Translated title. Original: IA na medicina: ela veio para ajudar a tratar ou atrapalhar? | VEJA RIO

Generative AI offers medical benefits (diagnosis, prevention, evidence synthesis) but can hallucinate and mislead; safeguards are needed to prevent harm, as AI should support, not replace, clinical judgment.

KI and literature – does that go together? | FAZ

Translated title. Original: KI und Literatur – passt das zusammen? | FAZ

The article questions the role of AI in literature, examining a program that detects AI contributions in texts and its implications for human critics and literary awards like the Prix Goncourt. It asks whether AI can mimic human writing to the point of evading detection and whether rejecting an award on AI-contamination grounds signals a broader crisis for human literary assessment.

[2610.36453] Correction or Manipulation? Measuring Epistemic Vigilance in LLM Agents Under User Pressure

The paper tests whether LLM agents distinguish correction from manipulation under user pressure, using seven models and two study setups. Results show concession to correction can be mistaken for manipulation, with policy violations not reliably predicted by concession; evaluation should consider both discrimination and action taken.