Not All Explanations Are Sought: Information-Seeking Psychology for Human-Centered XAI
This position paper argues that human-centered explainable AI should be designed around why people actually want explanations, not just around explanation quality. It frames expected utility in three modes – instrumental, hedonic, and cognitive – and warns that cognitive biases can cause two opposite failures: users can either seek too much irrelevant detail or stop too early and miss key risks, which becomes especially dangerous with agentic systems.
Instagram cracks down on AI accounts pretending to be human
Instagram is renaming its “AI creator” label to “AI-generated profile” and will actively target accounts that don’t self-label, aiming to make it obvious when a profile features an AI-made person. The change is a direct response to influencer-style “human” accounts that were getting harder to distinguish from real creators.
Beyond Fixed Budgets: Characterizing the Inelasticity and Limitations of Tree-of-Thought Reasoning Strategies
A new arXiv study finds that Tree-of-Thought methods don’t degrade smoothly as compute budgets shrink – different implementations fail for different reasons. One variant gets stuck exploring too slowly under tight budgets, while another burns valuable search frontier early, suggesting that search parameters can’t be “one size fits all” across the compute spectrum.
A milestone in expanding access to AI: ChatGPT Ads
OpenAI says ChatGPT Ads has reached a $1B annualized revenue run rate and is rolling out more broadly, including international expansion. The move is positioned as a way to support free and low-cost tiers while still monetizing scale.
Debian won’t ban AI code from its Linux distribution
Debian has adopted a policy allowing developers to use AI tools in contributions related to development, maintenance, and documentation – as long as the work remains “responsible” and within existing contributor expectations. The vote notably rejects proposals that would have barred AI-generated code outright, setting a pragmatic tone for how Linux communities handle AI-assisted workflows.
Benchmarking General Mobile Assistants in Challenging Real-World Scenarios
A new benchmark, GMA, targets a gap in agent evaluation: mobile tasks that reflect messy real life rather than tidy demos. It runs 300 tasks across seven real apps and shows that current frontier models fall sharply as workflows get more complex, while careful “agent harness” choices like state tracking can meaningfully improve performance.
Retrieving Relations, Detecting Fallacies: A RAG Approach to Political Debate Analysis
This arXiv paper attacks fallacy detection as a retrieval problem where argumentative structure – support and attack relations – guides what knowledge gets pulled into the model. On the ElecDeb60to20 benchmark, the approach boosts macro-F1 for both fallacy detection and classification compared to non-retrieval baselines, emphasizing that political reasoning often needs context beyond the text being judged.
ChatGPT to face tougher regulation in the EU
The European Commission has designated ChatGPT a “Very Large Online Search Engine” under the Digital Services Act, bringing it into a stricter compliance regime related to minors, mental health risks, and illegal content. The EU also named Reddit and Roblox under the same “very large” categories, expanding the scope of DSA-style obligations across the biggest platforms and services.
Hugging Face hack could indicate cultural issues at OpenAI
A report ties the recent Hugging Face incident to broader organizational lessons for frontier AI labs, arguing the episode may reflect culture and governance weaknesses rather than just a single technical failure. The focus is on how incentives, review practices, and escalation paths can shape whether safety boundaries are treated as real engineering constraints or optional guidance.