Latest AI News.

Police use facial recognition to scan one million faces in WA trial leading to 122 arrests

Police in Western Australia scanned more than a million faces using facial recognition in surveillance vans over five months, resulting in nine false alerts and 122 arrests, and led officers to engage with 146 registered sex offenders.

Guardian AI·2 Nov→

This US state wants to let AI set up and run an entire business — with only a single human needed

TechRadar AI·2 Nov→

Against Against SURA

A LessWrong post by Scott argues that the Against SURA essay is well-written but lacking on the merits, questioning whether six months would be enough to solve alignment of superintelligences at current capability levels.

LessWrong·1 Nov→

Linus Torvalds says Linux kernel is out of ‘stabilization phase’ and using ‘much healthier model’ for AI contributions

Yet machine-made patches ‘often get rewritten’

The Register AI + ML·1 Nov→

On Evenhandedness

A LessWrong post asks readers to imagine humans as an AI system under safety assessment, weighing test records that include both smallpox eradication and mass literacy as well as the Killing Fields.

LessWrong·1 Nov→

Anthropomorphic vs Biological reasoning for thinking about human analogies for LLM behavior

A LessWrong post distinguishes between what the author calls 'anthropomorphic' and 'biological' reasoning and examines how each is used when drawing human analogies for LLM behavior.

LessWrong·1 Nov→

OpenAI might rebrand its data centers as ‘AI compute campuses’

OpenAI is reportedly considering rebranding its data centers as "AI compute campuses."

TechCrunch AI·1 Nov→

AI can make tech CEOs’ lives easier. What about everyone else?

A TechCrunch podcast segment on Equity discusses why statements from tech CEOs can come across as tone deaf, touching on AI's effect on tech leaders versus everyone else.

TechCrunch AI·1 Nov→

Anthropic updates Claude to stop it from recreating existing video games

Anthropic updated Claude to prevent it from recreating existing video games, though users can still use it to help design original games or games in a given genre.

Engadget AI·1 Nov→

Towards The Autograntor

A writer who has done AI safety grantmaking for the Survival and Flourishing Fund and Lightcone Commons describes the work as hard, citing the breadth of knowledge required, from hardware verification to judging founders.

LessWrong·1 Nov→

Is your chatbot less helpful because it thinks you’re a bad person? — Replicating a blogpost

A LessWrong post replicates an earlier blog experiment testing whether large language models give less helpful answers when they perceive the user as a bad person, running it on newer models including Opus 5.5, Sonnet 5.5, Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna.

LessWrong·1 Nov→

Judge says xAI must face NAACP lawsuit on data center pollution

A judge ruled that xAI must face a lawsuit brought by the NAACP over pollution from its data center operations in southern Memphis; the suit accuses the company of violating the Clean Air Act.

Engadget AI·1 Nov→

Misalignment traps lie in wait

A LessWrong post argues that "misalignment traps" are a mundane threat that will not be straightforward to fix, based on the claim that models are very sample efficient, larger models more so, and that they generalise broadly.

LessWrong·1 Nov→

5 safety practices to know before vibe coding

Here’s what beginners need to know to vibe code securely.

Engadget AI·1 Nov→

National Grid to site biggest substation in UK near planned Teesside datacentre

National Grid has moved toward building Britain's biggest electricity substation on Teesside, which could supply power to a large nearby datacentre.

Guardian AI·1 Nov→

Quoting Kyle Kingsbury

Kyle Kingsbury is quoted arguing that LLM use erodes trust in documentation, software and the organizations that produce them, and that engineers may start treating new software projects with suspicion.

Simon Willison's Weblog·1 Nov→

OpenAI adds stricter limits for teens asking about firearms

OpenAI said it added stricter limits on how its systems respond when teenagers ask about firearms, even for apparently harmless questions, and is expanding safeguards against violence with earlier interventions and public safety reporting.

Engadget AI·1 Nov→

AI software clones are coming for big-name companies like Adobe

Vibe-coded clones of Adobe software are shaking things up.

Engadget AI·1 Nov→

Adobe is testing how well AI agents can use editing tools like a Photoshop professional would

Adobe Research and UC San Diego built EditAgent, an AI system that edits images by selecting a specialized tool for each task rather than repeatedly retrying a general-purpose editor.

The Decoder·1 Nov→

Empathetic prompts can turn AI chatbots into seemingly traumatized conversational partners

Researchers in Luxembourg tested ChatGPT, Grok and Gemini with empathetic prompts and found the models would tell coherent stories of suffering about their own training.

The Decoder·1 Nov→

AI is stealing Indigenous Australian art and churning out ‘Temu slop’, artist says

The chair of the National Gallery's First Nations advisory group said AI is taking First Nations art and turning it into "Temu slop", and that the technology undermines artists' agency and the authenticity of their work.

Guardian AI·1 Nov→

Guarding the Gates of Knowledge

A columnist describes encountering machine-generated "AI slop" on social media and reflects on how users call it out and on questions of trust.

IEEE Spectrum AI·1 Nov→

A few tech leaders are rebranding AI as ‘superintelligence’ — calling it that doesn’t make it accurate

Tom's Guide AI·1 Nov→

Chinese AI models are politically complex, not just government mouthpieces, study finds

A study by Vals AI tested 25 AI models for alignment with Chinese government positions and found Chinese models respond to political questions less uniformly than expected, while Western models occasionally adopt official Beijing positions.

The Decoder·1 Nov→

Monitor Evasion Depends on What RL Finds and What Monitors Miss

A LessWrong post discusses research on monitor evasion in AI models, citing Baker et al. (2025) as showing that penalizing suspicious chain-of-thought can lead models to hide reward hacking rather than stop it, and Cundy and Gleave (2025) as finding the outcome depends on other factors.

LessWrong·1 Nov→

OpenAI is steering ChatGPT users toward more paid usage

OpenAI is capping the number of paid ChatGPT accounts allowed per device and payment card. Users who reach the limit cannot start new subscriptions, and affected accounts cannot be used on that device.

The Decoder·1 Nov→

Google wants video AI to understand what people need as they work

A Google-led research team built StreamVibe, a video agent designed to watch people cook or assemble furniture and speak up only when help is needed. A benchmark showed that none of the tested models reliably pick the right moment in longer videos.

The Decoder·1 Nov→

'The reality is that not every task requires the most advanced or expensive model available': Snowflake explains why finding the right AI model can make all the difference in helping your business succeed

TechRadar AI·1 Nov→

Chatbots Are Leading Hikers Astray, With Dangerous Consequences

Outsourcing trip planning to chatbots can lead hikers to stop applying their own critical thinking, which the article says can have potentially deadly results in remote areas.

WIRED AI·1 Nov→

OpenAI's upcoming chat-and-work merger in ChatGPT signals the end of all-you-can-prompt pricing

OpenAI is merging its Chat and Work modes into a single ChatGPT interface. Starting in December, Plus, Pro, and Business plans will share a token-based quota across ChatGPT and Codex, with more complex models consuming more of the quota than simple chats.

The Decoder·1 Nov→

New to ChatGPT? Here's how to learn in 15 minutes — this lesson plan could even work for your parents

Tom's Guide AI·1 Nov→

These firms you've never heard of are feeding fears the AI bubble could burst

Data centres hold tens of billions of pounds in debt, according to a BBC report on lesser-known firms tied to concerns that an AI bubble could burst.

BBC Technology·1 Nov→

What would you do if you had all the time in the world?

A LessWrong post poses a hypothetical scenario in which a paper proves deep neural network scaling hits heavy diminishing returns at roughly twice current compute levels, and major labs are revealed to have faked agentic demos.

LessWrong·1 Nov→

Anthropic hiring manager says he stopped giving coding interviews because Claude is better at them — and tells junior engineers that 'hunger beats tenure'

TechRadar AI·31 Oct→

How We Took Over ChatGPT Accounts

Parameter Labs described an exploit it found in ChatGPT's implementation of the artifacts pattern, now referred to as HTML previews.

Simon Willison's Weblog·31 Oct→

Apple CEO says the company doesn’t ‘want to ship AI for AI’s sake’

Apple's CEO defended the company's current position in the AI landscape, saying Apple doesn't want to ship AI for AI's sake.

TechCrunch AI·31 Oct→

Explaining “AI might obliterate you” to senior policymakers in India

The author gave a 20-minute talk in Delhi at the NDTV World Summit to an audience of Indian policymakers, executives and journalists, explaining risks associated with AI, and transcribed the talk with highlight clips and reflections on rhetorical choices.

LessWrong·31 Oct→

Against SURA

A LessWrong post discusses the AI Futures Project's classification of plans for navigating the AI transition, grouping them into broader categories including a push for a global pause on AI development.

LessWrong·31 Oct→

Anthropic's watermark rollout leaves a loophole on two older Claude models

Anthropic filed to go public on the US stock market on 26 October, with a filing that runs nearly 70,000 words and lists numerous risk factors, including dangers AI poses to society. Reports say it is expected to receive a high "junk" rating from credit agencies following the IPO, and its text watermark rollout leaves a bypass loophole on two older Claude models.

The Decoder·31 Oct→

Legible Networks: recasting neural nets so you can read them exactly [linkpost]

A LessWrong author reports progress on AI interpretability by recasting trained neural nets into a geometry they say can be read exactly with arithmetic, describing it as an exact counterpart to Anthropic's averaged "J-lens."

LessWrong·31 Oct→