Tech's silence on ICE is not neutral
The accounts I follow can take eight paragraphs to argue about a system prompt. About what ICE bought from Palantir this month, they have written nothing.
An ICE officer types a zip code into a Palantir app called ELITE. A count of people with an “immigration nexus” appears on a map. The count, per court testimony, includes naturalized citizens. Somewhere in the model that produced the count there is a confidence threshold tuned by a person who is paid to think about thresholds. That person reads model cards on release day. That person is, in some sense, ours.
Here is what the accounts I follow have written nothing about.
The Washington Post reported on Tuesday that David Venturella has been named acting director of Immigration and Customs Enforcement, effective June 1. Venturella worked at GEO Group, the private prison company, from 2012 until 2023, including as Senior Vice President of Client Relations. Democrats on the House Judiciary Committee had already flagged a conflict-of-interest concern about him overseeing ICE contracts that flow to his former employer.
The day before, Lever News published a follow-up to its April investigation of Edge Ops LLC, a vendor that won a $12 million ICE contract for “Project SAFE HAVEN,” described in the contract as a “question-based AI interface” to track immigrants and categorize them as potential threats. After Lever’s first piece, Edge Ops scrubbed its website. The follow-up documents that the company appears to have used stock-photo executives, an absent computer scientist, and a “partner” that says it never worked with them.
And the same week, the ACLU published a roundup of four named Palantir products inside ICE. Not allegations — products with names, contract numbers, sam.gov entries, and a leaked user guide obtained by 404 Media. ImmigrationOS: a $30 million no-bid contract awarded April 2025, now over $145 million on USAspending.gov, running through September 2027. ELITE (Enhanced Leads Identification and Targeting for Enforcement): a targeting app ICE uses to map “target rich” neighborhoods, pulling data from HHS, USCIS, and a commercial data broker. ICM (Investigative Case Management): ICE’s case backbone since 2014, sole-sourced, and the DHS Privacy Impact Assessment confirms US citizens are included in the consolidated profiles. And an AI tip-processing tool the DHS AI Use Case Inventory describes in less specific terms than the contract documents do. In a separate agreement, ICE and CMS now share personal data on roughly 80 million Medicaid patients.
This is the documented, public stack. The contracts are on USAspending.gov. The PIAs are on dhs.gov. The user guide is in a 404 Media story. The GAO published a report in December 2024 — GAO-25-107435 — finding that of seventeen DHS sector AI risk assessments analyzed, none fully addressed the six activities required for effective risk assessment. Zero of seventeen. That finding is still standing.
I am not breaking any of this. Just Futures Law and Mijente have a long-running report called Automating Deportation. EPIC has been in FOIA litigation against ICE and Palantir since 2017. ACLU, 404 Media, The Intercept, The Guardian, WIRED, Fortune, and Lever News have each documented pieces of this stack across multiple years. The information is not hidden. It is just unread by most of the people I would expect to read it.
I have not posted about any of this either. The line tech’s silence on ICE is not neutral has been in my bio since I put it there months ago. The news that finally moved me to write came in a morning brief generated by an AI agent I built to find stories I would otherwise miss. The agent saw what I missed. I want to take that seriously rather than glide past it.
What it means, I think, is that the line was easier to write than to live. I had the conviction; the attention was the harder part. The platforms I read for AI news have organized themselves around things that are not this. After a workday of code and releases, the attention I have left is the attention those platforms shaped. Calling this cowardice misses what is actually happening. It is an attention economy doing what attention economies do, inside a community that calls that doing intelligence.
In the 2010s the same community I am writing to now made a critique of the Google and Facebook engineers who built systems whose downstream harms they did not have to see. The complaint then was that the architecture of not-looking was itself the harm. I think about that complaint less than I should, because the version of it I would have to make today implicates me.
I am also not going to say Palantir is purely doing harm, or that the government can be trusted on any of this. I have not read every line of Palantir’s code and I am not going to pretend I have. The contracts on USAspending.gov persist across administrations. The OIG and GAO write reports that go unread regardless of who is in the White House. Both things can be true.
I don’t want a post from you. I want a few minutes of the same attention you give a model card on release day. Open the ELITE user guide, the ICM Privacy Impact Assessment, the ImmigrationOS statement of work. These are AI systems whose outputs end in arrests.
The line has been in my bio for months. It was easier to write than to live.
Today I am trying to live it.
If this was useful, subscribe. I write here once a week, deep and slow, mostly about where AI is actually pointing and what it looks like from the inside. For lighter stuff — Codex tricks, Zo Computer walkthroughs, AI for people who feel locked out — find me at Zo Computer Toolshed on Facebook, or at jeffkazzee.zo.space.
P.S. If you want to try Zo Computer — the cloud workspace where most of my AI experiments live — here’s my referral link.