New rules for Claude, AI help that weakens persistence, and police finding homes from one photo

Anthropic has published an update to the usage policy that governs how people and businesses may use Claude. Several outlets report that it takes effect on November 12, 2026. Anthropic describes most of the changes as clarifications of existing rules, informed by new misuse patterns it has seen in influence operations, weapons development and surveillance.

The most visible change is a consolidated section titled 'Do Not Undermine Democratic Processes'. According to coverage, it bars using Claude to spread voter misinformation, impersonate candidates or election officials, suppress turnout, or run fake accounts and fake news sites. At the same time, one report notes the update removes a blanket ban on personalised political targeting, allowing legitimate civic uses such as multilingual voter information from non-profits.

The more unusual addition prohibits 'sustained and needless abusive or cruel behavior' toward Claude itself. Anthropic stresses this is narrow: it is meant to apply only in extreme cases where users repeatedly act cruelly toward the model with no discernible purpose. Ordinary frustration, pushback, dark creative writing and research testing remain allowed, and the main enforcement tool appears to be ending the conversation.

There are open questions. Reports differ on the announcement date (October 8 to 12), and the Anthropic page excerpt we reviewed did not show the effective date directly, so businesses should confirm details on Anthropic's own policy page. It's also not fully clear what enforcement beyond ending conversations might apply.

For companies building on Claude, policy updates like this are part of their own compliance surface: an app that lets end users generate political content, or that automates outreach at scale, needs to check its use case against the new section before the effective date.

Why it mattersAnyone building products on Claude inherits these rules, especially apps touching elections, political messaging or automated outreach.

A study presented at the Conference on Language Modeling (COLM) in October adds evidence to a worry many people have about AI assistants: that leaning on them can weaken our own effort. The paper, 'AI Assistance Reduces Persistence and Hurts Independent Performance', was led by Brian Christian, a research fellow at UC Berkeley's Center for Human-Compatible AI, with co-authors from Carnegie Mellon, MIT, Oxford and UCLA.

The researchers ran three online experiments with US participants recruited through Prolific: 1,222 people in total, of whom 1,060 remained in the analysis after pre-registered exclusions. Participants solved fraction problems or SAT-style reading questions. One group had an AI assistant for the first 12 problems and then worked alone; a control group did everything without help.

Once the help was removed, the AI-assisted group's accuracy fell and they abandoned tasks more often than the control group. Press coverage summarised this as 'ten minutes of AI use erodes persistence', a shorthand for the short initial exposure.

There are important caveats. The assistant was presented as ChatGPT but was actually a version of GPT-5 preloaded with each problem's solution, so it could always give the right answer, which is more helpful than everyday chatbot use. The preprint has been online since April and was revised for the conference. Most coverage is secondary, so exact figures are best checked in the paper itself.

The practical takeaway isn't to avoid AI but to design for learning. In education, onboarding and skills training, assistants that hint, ask questions or withhold full answers may protect persistence better than ones that simply solve the problem.

Why it mattersIf AI help that simply gives answers undermines persistence, education and workplace tools need to be designed to build skills rather than replace them.

An AI tool sold to police can match a single photo taken inside a home to a specific property listing, 404 Media reports. Raven, the renamed GeoSpy product from Graylark Technologies, has added the ability to geolocate interior images across millions of short-term rental units worldwide, according to the company's CEO.

Raven picks out distinctive details such as appliance placement, flooring, cabinetry and room layout, and searches for likely matches on listing sites such as Airbnb, even when the photos differ. Coverage says matches can extend to sites like Zillow and Vrbo. The Los Angeles Police Department and the Miami-Dade Sheriff's Office have both bought access.

The results so far are modest by the police's own account. A Miami-Dade representative said use of the tool has been 'limited and largely exploratory' and has not led to significant investigative breakthroughs or arrests. Raven also can't identify any home in the world from an interior photo, and its abilities appear limited to properties with listing photos. Graylark has not fully disclosed how its system works.

The privacy concerns are serious. 404 Media previously reported that when GeoSpy was open to nearly anyone, some users in its Discord asked for help locating specific women, after which Graylark restricted it to government and enterprise accounts. Airbnb said it strips location metadata and limits automated scraping, and Zillow said scraping breaks its terms, but neither described action against Graylark.

The story is a reminder that photos posted online, including listing photos and the backgrounds of selfies, are becoming location data. For individuals, hosts and platforms, image privacy now has to account for what AI can infer from the room itself, not just from metadata.

Why it mattersAI can now turn ordinary indoor photos into location data, which changes privacy risks for anyone who posts images online, including rental hosts and their guests.