Hallucination Is Now a Feature

OpenAI says it banned accounts tied to abuse campaigns ranging from romance scams to bots posing as legal professionals (Source: OpenAI, Reuters). OpenAI’s own threat write-up adds that these operations rarely live on one service, which is corporate language for “the problem has already left the building.”

Reuters reported one network using ChatGPT in dating scams aimed at Indonesian men, with alleged fraud affecting hundreds of victims monthly. The broader takeaway is less “new crime” than “old crime with nicer UX,” now spread across platforms and tools faster than moderators can map it.

Cursor users were told by an AI support bot that a single-device policy existed, which triggered confusion and cancellation threats before the company clarified that no such policy existed (Source: Ars Technica, Hacker News). A frontline bot effectively drafted policy by hallucination, and customers treated it as official because it arrived through official support plumbing.

This is the modern customer-service nightmare: not silence, but confident fiction at scale. By the time the correction lands, the damage has already been A/B tested in public.

The U.S. Supreme Court declined to hear Stephen Thaler's appeal over copyright for an image he said was generated autonomously by AI, leaving intact lower-court rulings that human authorship is still required (Source: Reuters, CNBC).

For now, U.S. copyright law remains stubbornly biological. This matters beyond U.S. borders because creator ecosystems in LATAM often inherit platform policy shifts driven by U.S. IP interpretation.

Apple researchers report frontier reasoning models can show complete accuracy collapse once puzzle complexity crosses certain thresholds, with related arXiv findings suggesting reasoning effort can even decline despite available token budget (Source: Apple ML Research, arXiv).

Translation: the models can look profound right up until the math gets rude.