I built the ICE register to be fair. It left me convinced enforcement has gone too far

Some of the defenders' claims hold up. ICE's arrests of people with a conviction or a pending charge really did more than double, from 81,312 in fiscal 2024 to 167,651 in fiscal 2025, on ICE's own definition; the register rates that "supported by primary data." Southwest border apprehensions for fiscal 2026 (98,708, preliminary) are the lowest since 1968. Enforcing immigration law is a legitimate function of government, and I'm not arguing for open borders or against deporting people convicted of serious crimes.

Some of the critics' claims are weaker than they sound. "2025 was the deadliest year in two decades" holds on a calendar-year count and fails on a fiscal-year count or as a death rate, where 2004 was worse. Press the buttons below: the claim from my side of this argument fails on two of three rules, and the defenders' claim holds on the count and fails on the share. I'd rather lose that point than make it badly.

Numbers alone won't move a reader who thinks this is simply the law working, so here are the principles I'm measuring against. None of them belongs to one party.

Detention needs a reason, every time. Locking someone up is the most serious thing a government does to a person short of taking a life. Immigration detention is legally civil, not punishment; the Supreme Court has said it is justified to make sure people appear for their cases or to protect the public. In Zadvydas v. Davis (2001) the Court wrote that "freedom from imprisonment ... lies at the heart of the liberty" the Due Process Clause protects, and that the clause covers all "persons" in the United States, whatever their status. So the burden of proof sits with the jailer, and the composition numbers below test whether that burden is being met.

Process should grow with the stakes. Deportation can split a family or send someone back to danger. A system that holds more people owes each of them more care, and the section after next shows it giving less.

The means should fit the stated end. A program sold as removing dangerous people should mostly hold dangerous people. That test does not depend on how many deportations you want; it asks only whether the system does what its own defenders say it is for.

Equal treatment and an open record. A policing model halted once over documented racial profiling should carry a heavy burden before it returns. And a government acting in your name owes you the record of what it did, on schedule. Power the public can't see is power the public can't consent to.

If you share these four, the rest of this post is a test of whether the system meets them.

Here is the part that moved me. The case for a large detention system is that it holds dangerous people. ICE held 39,314 people on January 20, 2025, and 67,180 on July 30, 2026, a 71 percent increase in GAO's count, with a peak above 73,000 in January. The documented target is a daily average of 99,000, funded by a $45 billion appropriation that runs through 2029.

On ICE's own July 11 file, 70.6 percent of the people held had no criminal conviction. Count pending charges as criminal, the most generous rule available, and 39.8 percent still had neither a conviction nor a charge: about 26,000 people. The share of arrests leading to detention that involved anyone with a conviction hit 27 percent in July 2026, the lowest on record, and arrests of people without convictions rose more than eightfold between the last six months of the previous administration and January 2026.

So the doubling of criminal arrests is real, and it is the smaller part of the story. The system grew far past it. Whatever you think the right number of deportations is, a detention population where most people have never been convicted of anything is not the population the system is sold as holding. Toggle the chart to Share: the red is the people with no conviction and no pending charge.

A skeptic could answer that a bigger system is fine if it is accountable. Look at what happened to the accountability.

The courts first. Between fiscal 2025 and 2026, 279 immigration judges, about 38 percent of the 2024 bench, were fired or resigned. The 130 who stayed raised their own asylum denial rates by 33 points on average. Asylum grants as a share of decisions fell from 45.7 percent in fiscal 2024 to 7.5 percent through the third quarter of 2026. Sixty-eight percent of June 2026 removal orders were issued in absentia, to people absent from the hearing, and backlogged cases now wait an average of 940 days, up from 571. Applicants did not change that much in two years. Who decides, and under what rules, did.

Then the hearings themselves. Expedited removal, deportation with no court hearing for people who can't prove two years of residence, was expanded nationwide in January 2025 and has been in force since a July 2026 stay order while the litigation continues.

Then the deputies. Signed 287(g) agreements, which hand immigration powers to local police and jails, went from 135 in January 2025 to 2,688 on October 16, 2026. Of those, 1,933 (71.9 percent) use the Task Force Model, the street-level version that was halted in 2012 after documented racial-profiling findings and revived in 2025. Each of these is defensible in isolation. Together they mean more people move through the system with fewer independent checks at every stage.

ICE's own list shows 33 deaths in custody in calendar 2025, triple 2024's 11. Fiscal 2026 had 36 by August 23, more than any fiscal year in the JAMA series that goes back to 2004, and its partial-year death rate (88.9 per 100,000 person-years) is nearly double fiscal 2025's 47.5. Advocates count seven apparent suicides in the year after January 20, 2025, against one the year before.

The honest caveat, again: fiscal 2004's rate of 127.7 was higher still. But the direction since 2023, when the rate was 13.0, is unambiguous, and in June 2026 ICE stopped counting deaths within 30 days of release. When an agency changes how it counts deaths while its death rate is rising, I don't need to assume bad faith to object.

Everything above comes from the scraps that are still published. ICE's enforcement dashboard still shows December 31, 2024. The DHS monthly enforcement tables were last updated January 16, 2025. There is no fiscal 2025 annual report. The detention report Congress requires every two weeks has posted twice since mid-February, last on July 20: 89 days and about six missed releases ago. The one series updated to the minute is the 287(g) list, the measure of the program's own growth.

The freshest record-level data now comes out through Freedom of Information lawsuits, one round at a time. A program this large that cannot keep its own mandated reporting current is the strongest single sign, to me, that it has outrun its capacity to be run well.

Since I'm asking skeptics to trust the register, here is where it was wrong. The refresh found seven errors in my July run. The biggest was in the page's own counting-rule switcher: it set Obama's DHS-wide removals against an ICE-only figure for the first Trump term. On one series it is 3,062,466 against 1,196,427, about 2.6 to 1, not 3.3 to 1. Gallup's 2024 reading was 64 percent where I had 66, and the detention report I called frozen since April had published once more in July. All seven corrections are on the page, dated. One verdict moved, claim 06, the deaths claim above, and it moved against the critics' framing.

I'd like this view to be testable, so here is what would move it:

1. Publish the detention report on schedule, with a breakdown by most serious conviction. If most people held have violent or serious convictions, the composition argument above fails, and I'll say so.
2. Show the death rate falling on a count that includes deaths within 30 days of release.
3. Stabilize the bench. If asylum outcomes stay this low once the bench stops turning over, the drop reflects the law, and I'll drop that point.
4. Release a method for the 2.2 million self-deportations DHS has claimed since January, a figure that has not moved while the counted deportations beside it rose by about 400,000.

Any one of these would weaken my case. None of them is a big ask of an agency with a $45 billion appropriation.

The composition numbers could partly reflect how charges get recorded rather than who is held; ICE's files don't say, and the share with neither a conviction nor a charge fell from about 48 percent to 40 percent between ICE's February and April 2026 files. The fiscal 2026 death rate covers part of a year. Judges leave for many reasons, and some of the denial rise may follow precedent changes a court would uphold. And a reader who weighs sovereignty and deterrence more heavily than I do can accept every number here and still conclude the trade is worth it. That is a values disagreement, and the register tags it as one. My claim is narrower: on the evidence the government still publishes, the system has grown past what it can justify and past what anyone can check.

If you think the data points the other way, the register is built for that argument. Every claim has its sources, and the same method is behind Counting the Killings, on police shootings, and the legibility gap series.

So the burden sits with the government, and it is a small one. An agency asking for 99,000 beds should be able to tell us, every two weeks as the law already requires, who is in the 67,180 it fills now. It has not published that report since July 20. Until it does, every new bed is a bet placed with someone else's liberty.