The End of Being Hard to Find
For years, an ordinary kind of privacy has depended on other people having better things to do. A photograph might be public, an old comment searchable, a professional biography available to anyone. Yet those fragments could remain practically separate because connecting them required time, attention and a reason to bother. That imperfect shelter deserves more attention as artificial intelligence becomes better at interpreting the traces people leave behind.
On September 10, Anthropic published evaluations of AI capabilities in intelligence targeting and conventional weapons development. The privacy findings deserve attention in their own right. Its researchers tested account linkage using synthetic social-media material and location inference using photographs and older public posts. In one account-analysis evaluation, a leading model assessed a roughly 37,000-word sample in about 11 minutes. Anthropic cautions that the synthetic material is imperfectly realistic. Its photograph evaluation also compared models with expert players on a different geolocation task, not a controlled head-to-head human trial. Those limitations rule out a sweeping claim that AI can reliably find anyone. [1]
The underlying concern predates this week's announcement. A November 2025 research preprint, GEO-Detective, reported that combining language-and-vision models with tools improved location inference from images compared with baseline models. That separate work supports the broader concern about image privacy; it does not independently validate Anthropic's new measurements. The distinction matters when a fresh corporate study arrives wrapped in the authority of precise-looking results. [2]
Conjugo's interpretation is that this is a question about power as much as technical performance. Imagine the difference between someone noticing a remark you made in passing and someone commissioning a detailed assessment of everything you have said in public. Both involve public information. They are not equivalent social relationships. The second creates a durable object about you, assembled for purposes you may never see and judged by standards you cannot challenge.
That difference should make us wary of treating publication as unlimited consent. A person can choose to participate in a community without choosing to become a permanent analytical project. A photograph shared to celebrate a place has one meaning within a circle of friends and another when incorporated into a profile. We need a public vocabulary for that change of context, because a checkbox about whether a post is visible does not capture everything that can subsequently be done with it.
Better inference also creates two different kinds of danger. A system might correctly connect information that someone reasonably hoped would remain separate. Or it might connect the wrong fragments and produce a persuasive account of a person who does not exist. Consider an employer, investigator or administrator reading such an assessment: the document's smooth prose could make an uncertain inference feel settled. The problem would then be both exposure and misrepresentation, with the subject carrying the burden of correcting a story they did not write.
Accuracy alone therefore cannot be the whole standard for responsible deployment. We should also ask who commissioned an inference, what legitimate purpose it serves, how uncertainty remains visible, and what happens before anyone acts on it. A human reviewer needs time and authority to reject the system's conclusions. Otherwise, human oversight risks becoming the ceremonial act of clicking the button after the machine has already framed the person.
This is not an argument for abandoning AI-assisted research. There are valuable reasons to connect dispersed evidence, including understanding institutions, examining claims and making complex public material accessible. The distinction should turn on purpose, consent where appropriate, proportionality and consequences. Helping people understand a public policy is a different proposition from quietly constructing intimate profiles of the people affected by it. An organization ought to be able to explain that difference before deploying the same capable system for both.
For platforms and developers, a useful response would treat privacy as a workflow property. What information can the system gather? What kinds of personal inference will it decline? What records show how a consequential conclusion was reached? How can an affected person seek correction? These are design questions as well as governance questions. They should be tested alongside speed and benchmark performance, rather than arriving as an apologetic paragraph after launch.
For individuals, the practical response should not become an instruction to disappear from public life. We can be thoughtful about what we publish without accepting that participation makes every subsequent use reasonable. Journalists, community organizers, artists and ordinary people should not have to choose between visibility and dignity. Societal preparation for AI means building institutions that can protect that distinction, not simply handing everyone a longer list of things to worry about before posting a photograph.
None of this requires a machine to be conscious, generally intelligent or secretly plotting its next move. The challenge is already intelligible at a more ordinary level: software becomes better at drawing conclusions, and people decide where to apply it. A future with more analytical capability could give individuals greater understanding and agency. It could also make them easier to categorize without their participation. We should decide what protections belong around that power before being hard to find becomes a privilege available only to people who can afford it.
Header: AI-generated conceptual illustration.
Sources: [1] Anthropic, September 10, 2026. [2] Zhang and colleagues, GEO-Detective, November 27, 2025 (preprint).
https://www.anthropic.com/research/intelligence-targeting-conventional-weapons-capabilities
www.anthropic.com