Lighthouse

Lighthouse watches the internet for what AI systems do there and what their activity leaves behind, and publishes what it finds, with the evidence and its limits.

Studies LH003

LH003 Survey of surveys

The answer so far. Thirty-four candidate observatories were considered; seventeen have a full row in sources.md's included table and seventeen are excluded there with a stated reason, an observation. Of the seventeen included, most see only one of LH F02's layers from one narrow population: repository-event archives and dependency resolvers see residue (changed structure), API-broker rankings and vendor usage reports see activity (inference calls), and vulnerability databases see a specific kind of residue, a disclosed flaw, rather than its exploitation. No source in this reading is classed as observing propagation directly. That is Lighthouse's classification, not an observation about the sources: LH F02 defines propagation as a derived link between an earlier residue record and later activity, not a surface any instrument reads off directly, so under that definition no source could be counted as seeing it. The empirical question is which of these records support identifying and testing such links. A propagation claim has to be built by joining two sources' records and testing the join, which is what LH002 is designed to attempt on a small scale for one such join (a dependency change and a dependent's later change) and what this study does not attempt at all. The strongest supported finding, an interpretation resting on the rows above rather than on any single one, is that these seventeen sources are partial views whose relationships to one another are unmeasured in anything this survey read: each is scoped to its own population (one forge, one broker, one vendor's product tiers in a given report, one company's network), no source read here states how much its population overlaps another's, and this study took no joint sample, so their counts cannot simply be added into a total. Distinct boundaries do not establish little overlap, and the table itself records connections, an observation: Open Source Insights links packages in seven ecosystems to their GitHub, GitLab and Bitbucket projects and to OSV advisories; OSV.dev aggregates GitHub Security Advisories, PyPA and RustSec among its upstream sources; GitHub's unreviewed advisories come from the NVD feed and are published in OSV format; NVD enriches records that CVE Numbering Authorities assign (this last per a search summary; see the fourth Finding). Where the populations overlap, and by how much, is a question for a joint sample, not something this reading establishes. What stays invisible, or visible only in part, is specific. No source read here can reliably tell a person's download or code change from an agent's. Cloudflare's AI Crawl Control does attribute crawler traffic to named AI operators, and its page lets a site owner group that traffic by crawler, by operator and by "Category: Analyze crawlers by their purpose or type" (an observation, from notes/R-0001.md's raw read of the page, last updated 23 April 2026); but it covers only crawlers Cloudflare has already identified, and because the category belongs to the crawler rather than to any one request, it cannot say why a particular page was fetched (an interpretation). The conversation figures do not say whether a person or a program was typing. No source states at what layer an AI system operated, and usage that happens off all these particular surfaces goes unseen, for instance direct provider API traffic outside a broker, a private repository, or a deployment that never triggers a public event. The first of those examples holds for the sources as read here; the first review note says later Economic Index reports, which this survey has not read, include one vendor's first-party API traffic.

Paper, version 0.4, 26 September 2026 studies/LH003/LH003.md

From studies/LH003/ in the repository, last changed 26 September 2026.