Everyone has heard a big claim about ICE. Almost nobody has been shown how to check one. This page takes twenty-one common talking points, from both political directions, and answers the quieter question behind each: what source would actually settle it? Same rigor for both sides. It is a map, not an argument.
Every number here was checked live on 2026-10-17, and each carries its source and its dates. Fair warning about this subject: the official sources break often. That is why the first stop below is a status board of government data series that have simply gone dark.
Immigration enforcement runs on words that sound alike and mean different things. That is not an accident; it is where most bad statistics hide. Tap any term for the plain-English version.
The single sharpest fact in this inventory: the best current data on ICE is adversarial, not official. The agency’s own congressionally required detention report, due every two weeks, has come out twice since February 2026. The records still flowing come from FOIA lawsuits (the Deportation Data Project[src]) and an independent clearinghouse (TRAC[src]). Of the 11 series this page rests on, 7 had a problem on the day we checked.
Status as verified 2026-10-17. Check again before you reuse anything; nothing in this space stays live for long.
Most fights about ICE numbers are not fights about facts. They are fights about counting rules that nobody says out loud. Four rules do most of the damage. Before you trust any claim, ask which side of each rule it stands on.
Here is the whole page in one interaction. Each panel below holds a claim steady and changes only the counting rule. Both options in every panel are defensible. That is exactly the problem: pick the rule, and the verdict follows. When a claim flips like this, the honest answer is not one side or the other. The honest answer is that it flips.
One more thing about the bottom of every rate. Nobody counts the unauthorized population; two research groups model it. For the same year, 2023, Pew says 14.0 million[src] and CMS says 12.2 million[src]. That is a 1.8-million-person gap. Every per-capita figure you have ever seen about this population inherits it.
Twenty-one talking points, each opened with a plain-English answer and backed by the full workup one tap deeper. The register holds 9 claims critical of enforcement, 9 favorable to it, and 3 used by both sides. Every one gets the same three steps: compute the number from the primary source, stress-test it under other definitions, and record the strongest point against it. The symmetry is a rule you can check, not a vibe.
Every figure on this page traces to one of these 82 sources. Each tier says how close the number sits to the original record, not whether we like the publisher. Advocacy groups appear on both sides, and both sides get the same T4 label. One rule above all: a think tank's chart points to a source. It is never the source.
Definitions and reporting systems changed on these dates. Comparing two figures from opposite sides of a line below, without a footnote, is how most viral statistics go wrong.
Some questions have no public answer, and saying so is a feature of this inventory, not a failure. Here is what the record cannot currently tell you, and what would change that.
Every figure here is a constant in a downloadable dataset, tagged with its source and two dates: the date it describes and the 2026-10-17 date we verified it. The derived numbers (the 71.7% arrest figure, the 17.6% interior share, the composition percentages) come from small pure functions, and our tests re-run those same functions in CI. So the arithmetic on this page cannot quietly drift from what the register says. Verdicts come from a fixed list of six labels. Claims about what the government *should* do get the "not a factual claim" label and are left alone, because data does not settle values. And in both directions: aggregates only, never a dossier on any person.
Behind this page sits a three-layer corpus: a source inventory, an independent fact-check of 58 of its assertions, and a resolution addendum that closed most of the fact-check's open items and reversed it twice. All three live in the site's repository at docs/research/ice-claims-*.md, with corrections applied. Refresh cadence: quarterly, or immediately when ice resumes the semimonthly detention report or publishes an fy2025 annual report.
This is the counting-rules companion to Counting the Killings, which found the American state unable to count the people its police kill. Here the state can count, and in 2025-2026 it increasingly declined to publish. The opinion I drew from the October 2026 refresh, labeled as opinion and kept off this page, is I built the ICE register to be fair. Both follow Classification as Infrastructure: the category system (interior vs border, removal vs return, conviction vs charge) does the argumentative work while staying invisible. The field-data method comes from The Legibility Gap and Pingree Grove Studies. The full catalog is at /research.
I build the AI tools and automations a team adopts, and the full-stack apps and data pipelines under them, and ship them to production. Tell me the problem in a sentence and I'll give you an honest read on fit within a day.
Work with me →