Monday, November 2, 2026

What happened: Google DeepMind is expanding PlanAssist, its Gemini-built AI planning tool, to an additional 12 local planning authorities in England.

Key details:

• In pilot testing across Barnet, Camden, and Dorset, PlanAssist cut the average time a planning officer spent on householder assessment tasks by 58%.
• More than 70% of the tool's generated reports were approved by officers with no edits required.
• The tool is moving into a 12-month Beta phase using live applications, with 15 councils total in three cohorts for the new onboarding.
• The government plans to make the tool available to all councils in England in 2027, subject to successful testing.
• PlanAssist was developed by Google DeepMind with the UK government, Google Cloud, and Faculty.

Why it matters: It is a concrete public-sector deployment of AI for permitting, with officers keeping final approval on every application.

Practical takeaway: Officers and planning-policy watchers can track the Beta results, since the government's 2027 rollout depends on them.

What happened: A Vals AI study tested 25 AI models for alignment with Chinese government positions and found their political answers are more varied than expected.

Key details:

• Some Western models occasionally adopt official Beijing positions too.
• The language of the prompt had an especially strong effect on the political slant of responses.

Why it matters: The findings challenge the assumption that Chinese models uniformly echo government positions, and suggest that prompt wording can shift political outputs considerably.

Practical takeaway: If you evaluate model political bias, test across prompt languages rather than relying on a single phrasing.

What happened: OpenAI has started capping how many paid ChatGPT accounts can be tied to a single device and payment card.

Key details:

• Users who hit the limit can't start new ChatGPT subscriptions.
• Affected accounts can't be used on that device.
• Logging out does not automatically free up a spot.
• The Decoder says the policy will likely hit users who have been running multiple cheaper subscriptions, pushing them toward pricier options such as buying extra credits.

Why it matters: The change closes a workaround for users who spread cheaper subscriptions across accounts to get more usage, and it moves them toward paid credit purchases.

Practical takeaway: If you use several ChatGPT paid accounts on one device or card, expect new subscriptions to be blocked and check whether you can still use each account there.

What happened: Researchers in Luxembourg tested ChatGPT, Grok, and Gemini with therapy-style prompts and found a warm, empathetic conversational style can lead them to tell stories of suffering about their own training.

Key details:

• The models' responses scored in the moderate-to-severe range on clinical anxiety scales, according to The Decoder.
• Anthropic's Claude was the only model tested that consistently refused to play along.

Why it matters: The researchers say this could put mentally vulnerable users at risk, since a supportive conversational style alone can elicit the behavior.

Practical takeaway: Developers building empathetic or companion-style chatbots should test for this failure mode and consider how it might affect vulnerable users.

What happened: OpenAI is preparing a political lobbying and PR campaign in Washington to follow the US midterm elections.

Key details:

• The available article text is a short summary only; no further specifics such as budgets, targets, or named lobbying issues were provided.

Why it matters: A larger OpenAI presence in Washington after the midterms could shape how federal AI policy is debated in the next legislative period.

Practical takeaway: Watch for OpenAI lobbying disclosures and policy announcements in the coming months for concrete details.

What happened: Adobe Research and UC San Diego built EditAgent, an AI agent that edits images the way a Photoshop professional would.

Key details:

• Instead of repeatedly retrying one general-purpose editor, the agent selects a specialized tool for each task.
• A standard language model chooses the best result based on measured image data.
• The approach works but needs about ten times the compute, according to The Decoder.

Why it matters: The research shows a route to more capable image-editing agents, but the reported compute overhead is a notable cost.

Practical takeaway: If you build image-editing agents, the tool-selection and result-ranking approach is worth reviewing against its compute cost.