1880 to 2000
In 1918, Cher Ami was a national icon who saved a battalion.
Fast forward to 1966, and New York's Parks Commissioner was calling them "rats with wings." How did the world's most reliable messenger become our most famous pest?
The ultimate PR disaster, 100 years in the making.
Act 2
The Badass of the Skies.
Argonne Forest, 1918
On October 4, 1918, Major Charles Whittlesey's Lost Battalion of the 77th Division was trapped in the Argonne Forest. The unit was under German fire and, worse, under friendly American artillery.
The third message
The message was short and desperate: the barrage was falling on American soldiers. Cher Ami launched with the capsule attached, a living courier inside a communications system built for moments when wires and radios failed.
The wound
The bird was hit in the chest, blinded in one eye, and badly injured in the leg. The story is famous because the flight did not stop. Cher Ami covered roughly 25 miles in about 25 minutes and the shelling ceased.
The medal
France awarded Cher Ami the Croix de Guerre. General John Pershing helped make the bird a symbol of wartime service. More than a century later, in November 2019, Cher Ami became the second recipient of the Animals in War & Peace Medal of Bravery, awarded posthumously at a ceremony on Capitol Hill in Washington, D.C.
The artifact
Smithsonian taxidermists preserved Cher Ami after the war. The model beside this text is built from the taxidermied body on display at the National Museum of American History.
One of many
The US Army Signal Corps Pigeon Service used hundreds of birds in World War I and kept the service active until 1957. Cher Ami gives the story a face; the institution explains why newspapers had a heroic vocabulary ready to use.
Act 3
I measured words appearing near "pigeon" in the New York Times corpus, grouped those words into cultural frames, and tracked the rate of each frame per article over time.
Top collocates in focus
The measured frame
Words like "carrier," "messenger," "service," "duty," and "saved" become a measurable vocabulary around pigeons. The line rises when war makes the bird useful to human survival.
Wartime peaks
The chart marks World War I and World War II separately. In this processed NYT file, the WWII spike is the larger one; the Cher Ami-era peak still anchors the first heroic surge.
Targets before heroes
Sport and hunting terms dominate early coverage. Live-pigeon trap shooting was fashionable, and the passenger pigeon was being pushed toward extinction. Pigeons were targets before they were heroes.
A new vocabulary appears
The pest vocabulary does not arrive fully formed. It builds slowly, then becomes easier for newspapers to reach for after the war-era meanings recede.
Attention collapse
By the urban-pest era, total framed language is far lower than wartime levels. Newspapers still mentioned pigeons, but the old symbolic vocabulary had thinned to a whisper.
Act 4
Hoving's "rats with wings" line matters because it gave a name to a trend already visible in the language.
WWII era
Hoving era
1990s wave
1918-1919
The heroic line belongs to a world where the pigeon could be written as an ally, courier, and servant of the state.
Late 1950s
The crossover is not a clean switch thrown on one day. It is a zone where heroic language loses its old force and nuisance language becomes available.
1963-1964
Disease, droppings, death, nuisance, and control become part of the public vocabulary before Hoving becomes parks commissioner.
1966
The phrase "rats with wings" is memorable because it compresses a cultural shift into four words. The data suggests he amplified a current, not created it.
1990s
The pest frame keeps returning. By then, the heroic vocabulary is mostly historical residue: something invoked in memory, not in everyday urban life.
Act 5
The map view compares pigeon density with the built environment in Washington, DC and New York City. Hover a tract in either city to highlight the same pigeon-density quintile in both maps.
Linked view
The maps use the same pigeon-density quintile scale. A tract in the top band in DC highlights the top band in New York too.
Synthetic cliff hypothesis
Tall buildings mimic the ledges and vertical faces rock pigeons evolved to use. If that hypothesis is right, dense vertical form should align with pigeon density.
DC caveat
DC has dense blocks, but the height cap changes the available vertical habitat. The relationship should be less dramatic than in Manhattan.
Citizen science caveat
eBird records reflect birder behavior as much as pigeon behavior. This is evidence about urban sightings, not a perfect census of birds.
Act 6
Across the century, the story moves through wartime institutions, urban design, public-health language, and finally a fading attention span.
Appendix
Where the imagery, data, tools, and analytical choices behind this piece came from.
The historical newspaper screenshots throughout this piece come from two open archives.
The placeholder photograph of Cher Ami's taxidermied body comes from the Smithsonian National Museum of American History's online collection: si.edu/object/cher-ami.
I captured the 3D Gaussian Splat of Cher Ami by recording a walk-around video on my iPhone of the taxidermy on display at the Smithsonian National Museum of American History. I processed the video into a .ply Gaussian splat using Kiri Engine, converted it to .ksplat for faster web delivery, then rendered it in the browser with Three.js and @mkkellogg/gaussian-splats-3d.
5zhs-2jue) and 2020 Census Tracts (63ge-mke6): opendata.cityofnewyork.us.ksplat Gaussian splat
GSAP + ScrollTrigger
pinned scrollytelling sections and scroll-driven progress
Observable Plot
newspaper framing charts and the comparative scatterplot
D3
CSV loading and chart/data transforms
Leaflet
interactive DC and NYC choropleth maps
proj4
projected coordinate handling for tract geometry
Kiri Engine
iPhone-video-to-Gaussian-splat processing
AI tools were used as aids for scaffolding, visualization design, coding support, debugging, and asset generation. The final analytical decisions, interpretation, and editorial choices are my own.
OpenAI Images was used to generate reference icon imagery, which I then traced and refined into final SVG assets in Adobe Illustrator. The finished icons on this site are my own SVG work; the AI-generated images served as visual references during the design process.
Anthropic's Claude and OpenAI's Codex helped scaffold the site's technical structure, wire Quarto to the JavaScript libraries, troubleshoot data and rendering issues, and explain unfamiliar techniques such as browser-based Gaussian Splat rendering and scroll-driven animation with GSAP ScrollTrigger.
I also used AI tools as a sounding board during data analysis to debug API endpoints, troubleshoot data-pull failures, and discuss methodological trade-offs. The final interpretation of the data is my own.
The linguistic story is built on weighted collocation analysis. For each newspaper item containing the word "pigeon," I counted how often words from five hand-curated category lists - heroic / utility, sport / hobby, culinary, hunting, and nuisance / disease - appeared within a 15-token window of "pigeon." Each match was weighted by its distance from the target word, so adjacent words counted close to 1.0 and words 15 tokens away counted about 0.06. The category lists were built from frequent collocates of "pigeon" in the collected corpus and then manually grouped into themes; OCR artifacts and neutral filler words were excluded.
Per-article scores were normalized by article count when comparing across years, producing the "frame hits per article" metric used in the charts. The final public-facing time-series charts use the NYT rows from annual_collocation_tone_by_source.csv for the 1880-2000 arc, plus nyt_decade_collocations.csv for the vocabulary callouts. Chronicling America informed the broader archive, source selection, and historical context, but its per-article rates are not mixed with NYT rates on the same chart.
This separation matters because the corpora expose different kinds of text. Chronicling America records are longer and noisier OCR/page records, while the NYT API returns shorter metadata fields such as headline, abstract, snippet, and lead paragraph. The same collocation method can be run on both, but the final visual comparisons avoid treating the two text lengths as directly interchangeable.
The geospatial analysis aggregates eBird and GBIF Columba livia observations to the census-tract level for DC and NYC, computing pigeons per square kilometer and buildings per square kilometer for each tract. DC tract geometry is handled in EPSG:26985; NYC tract geometry is handled in EPSG:2263 and converted to square kilometers. NYC's building-density measure counts buildings 100 feet or taller from the NYC Open Data building-footprints pull; DC's measure counts all building footprints, matching the city's lower-rise built environment.
A note on the eBird and GBIF data: citizen-science observations reflect where observers go, not strictly where pigeons are. The signal is useful for comparing observed urban pigeon density, but it is biased toward parks, tourist areas, and well-traveled observation routes.
This project was developed for DSAN 5200 in Spring 2026. I would also like to acknowledge Professor James Hickman. His course materials are available at jfh.georgetown.domains/centralized-lecture-content (Student Access Only).