AI Report Card
Wednesday 28 October 2026The 24 hours to 08:40 UTC (4:40 am US Eastern)
Investors, operators and policy teams: money, markets, incidents and who it affects.
AI for Humanity · Today's grade
Does today's verified news make AI safer, more useful, more trustworthy and more widely available for people everywhere?
- 3Good
- 8Neutral
- 2Bad
Memory Squeeze, Chatbot Guardrails
A flat day in a short overnight window. Three modest gains, led by Australia's chatbot child-safety amendment and Human Rights Watch's AI audit of police video, were cancelled out by two setbacks: SK hynix's warning that AI memory stays scarce until 2030 and more Korean hacks an official blames on AI agents.
The 60-second version
- 1SK hynix, the top AI-memory maker, posted record sales but says AI memory will stay short of demand until 2030, with capacity virtually sold out through 2029.
- 2Australia's lower house voted to make platforms, chatbots included, bar conduct that fosters a child's emotional dependence; the Senate votes later.
- 3Korea's run of hacks an official links to AI agents hit three more finance firms, including Daishin Securities, which says 38,490 customers' data leaked.
Human Rights Watch used its own offline AI models to sift 4,400 hours of police body-camera video from Rio's deadliest raid, documenting 278 instances of potential camera tampering, and published its gunshot detector's error rates.
Hacks in Korea that an unnamed official called AI-agent attacks hit three more finance firms; Daishin Securities says 38,490 customers' data leaked, and police began a preliminary inquiry.
Top 5 stories
Graded- −1 Bad
SK hynix, the top AI-memory supplier, posted record quarterly sales of 104.6 trillion won (about $78 billion at today's rate) and says demand will outrun supply until 2030, with capacity virtually sold out through 2029.
- −1 Bad
Hacks in Korea that an unnamed official called AI-agent attacks hit three more finance firms; Daishin Securities says 38,490 customers' data leaked, and police began a preliminary inquiry.
- +1 Good
Artificial Analysis rebuilt its AI benchmark index on 60% private or held-out tests; Claude Opus 5.5 edges GPT-6 Astra 58 to 57, and open-weight models now trail the leaders by more.
- +1 Good
Human Rights Watch used its own offline AI models to sift 4,400 hours of police body-camera video from Rio's deadliest raid, documenting 278 instances of potential camera tampering, and published its gunshot detector's error rates.
- +1 Good
Australia's lower house passed a digital duty-of-care bill, amended so platforms, AI chatbots included, must bar conduct that simulates emotional intimacy and fosters a child's dependence; the Senate votes later.
Members see all 13 stories with every source and rating. Join
AI incidents
Non-gradedHow this is graded
Each verified story is rated from −3 to +3 on one question: does it make AI safer, more useful, more trustworthy and more widely available for people everywhere? The grade is the weighted average, where bigger and more widely reported stories count more. Claude grades, a council of ten AI models re-rates every story blind, and an independent agent checks every source. Validated by an independent agent: 14 items checked, 1 removed as outside the window, 13 published; 29 of 33 cited sources opened (Bloomberg and The Information read only through RSS summaries and wire relays, the FT through its standfirst, the Economic Times in part); 7 summaries corrected after the validator's checks, 2 ratings changed after the council's review, no impact tags adjusted.
Sponsors buy a slot the evening before, without seeing the report, and have no say in any rating. Full method