BBS observers recorded about 41% fewer birds that sing at 7 kHz or more after 20 years on a route; route x observer models put this loss into 1966-2025 trends of 446 species

Question: long-serving North American Breeding Bird Survey (BBS) observers record fewer high-pitched birds. Standard trend models use observer-route effects and assume that detection does not drift while one observer serves. How large is the resulting bias in species trends? Data: 130,303 standard BBS runs (2026 release, 1966-2025; 5,319 routes; 10,417 observers), 446 species, song frequencies of Sagar et al. (2024), GHCN_CAMS climate, and the official USGS 1966-2025 trends (CC0). Methods: for each species, conditional Poisson models with (A) route x observer effects, (B) route effects, and (C) route effects plus a tenure term that resets at each observer change; all with year effects. We wrote the plan before we calculated trends. Post-hoc: tenure bins, end-of-service terms, a simulation, and song frequency as an external instrument. Results: the count change per year of tenure was -0.13% below 3 kHz, -0.55% at 3-5 kHz, -1.23% at 5-7 kHz and -2.06% at 7 kHz or more (slope -0.30% per kHz, SE 0.05). After a small gain in the first 2-4 years, counts after 20 or more years on a route were +2%, -8%, -24% and -41% relative to the first run in these classes. The visual control groups (raptors, water birds, swallows, swifts) had no late loss, but their linear drift also decreased with frequency. Model A put all of the drift into the trend (fraction 1.16, 95% CI 0.96-1.44). Its median 1966-2025 trends were 1.7 points per year lower than model C at 5-7 kHz and 0.4 lower below 3 kHz. A minus the drift had no frequency dependence (0.05). In the simulation, model C recovered the true trend and drift. The official USGS trends carried an uncertain part of the drift (0.39, CI -0.27 to 0.77); the pre-specified tests did not find it. Cold-adapted species had more drift (+0.053% per year per 1 C), which gives a false CTI warming of about 0.04 C per 10 years of tenure. With the drift removed, cold-adapted species did not decline more than warm-adapted species. Limitations: observer ages are not known; tenure is a proxy; the correction assumes no cohort-specific detection; a frequency-independent difference of 0.2-0.4 points per year between models remains unexplained; only 10 species sing at 7 kHz or more. Conclusion: route x observer trend models make trends of high-pitched, often cold-adapted species too negative unless they model tenure.