Immigration and Customs Enforcement is looking for AI software that can monitor the phone and video calls and the movements of immigrants held in detention centers. The agency posted a document on a federal government contracting website last Friday, and The Washington Sun reported on it on October 13, 2026, crediting the watchdog group Project Salt Box with first reporting the document.
This is a request for information, not a contract. It signals what ICE wants to buy, how it describes the problem, and which AI techniques it expects: transcription, voice recognition and video analytics running around the clock.
ICE detainee AI monitoring at a glance
| Question | Answer |
|---|---|
| What was posted? | A request for information on a federal contracting site, last Friday |
| What is wanted? | Software, called the "Suite" in the document, running 24/7 |
| Inputs? | Detainee phone, messaging, tablet, video-visitation, movement, incident and camera data |
| Outputs? | Leads on safety, security and potential criminal activity for an employee to review |
| Voice features? | Transcription, voice recognition, stored detainee caller profiles |
| Contract awarded? | No, this is a preliminary market-research step |
| Who first reported it? | Project Salt Box, per The Washington Sun |
What the document reportedly says
The document states that ICE facilities "generate authorized telephone, messaging, tablet, video-visitation, detainee management, movement, incident, and camera data at volumes that cannot be reviewed and correlated manually in a timely manner," according to the report. ICE wants software that can integrate information it already collects and produce leads, with a human employee reviewing them.
Three capabilities stand out.
Call content and contacts. The system would use transcriptions and voice recognition to monitor who detainees talk to and what the calls contain.
Caller profiles and "receiver-side deception." The software would keep profiles of detainees and track whether the person on the other end of a call is the person it associates with the dialed number. The document labels a mismatch "receiver-side deception."
Camera analytics. Using video feeds inside facilities, the software would need to flag fights, falls, signs of aggression and self-harm.
ICE already monitors detainees' calls that are not related to legal matters. What would change is automation: voice recognition and correlation across data streams, instead of staff listening and reading.
What people are asking
Is this a done deal? No. A request for information is market research. Agencies use it to learn what vendors can offer, to shape requirements and to decide whether to issue a formal solicitation. Many RFIs never become contracts, and the final requirements often differ from the first draft.
Does the AI make decisions about detainees? The document, as reported, describes software that generates leads for an employee to review. That wording puts a human at the end of the chain, but it does not say how much weight the lead carries, whether reviewers see the underlying audio, or whether a flag can affect housing, visitation or case outcomes. Those details decide whether the human check is meaningful or a rubber stamp.
Why call it a "Suite"? The document uses the word for software that would combine information ICE already collects, from calls to cameras, into leads. Combining streams raises the stakes, because a voice flag could be cross-checked against movement and incident data, producing a profile richer than any single stream.
How is this different from existing call monitoring? Today staff and recording systems capture non-legal calls, and review is limited by human time. The RFI's own justification is volume that "cannot be reviewed and correlated manually in a timely manner." Automation removes that limit, which is the point for ICE and the worry for critics: surveillance that was once constrained by effort becomes cheap to run on every call.
What is confirmed and what is not
- Confirmed by the reporting: the RFI exists, its stated purpose, the inputs and the features above, and that it is a preliminary step in finding a contractor.
- Not known: which vendors might respond, what accuracy ICE would require, whether a human must approve each flagged lead, how long data would be retained, how legal calls with attorneys would be excluded in practice, and what the budget is. We could not find the solicitation text itself independently, so the details above rest on The Washington Sun's account of it; check the federal contracting listing directly if you need exact wording.
- Context: the report describes the proposal as part of an expansion of AI use by the Department of Homeland Security to track immigrants and citizens during the second Trump administration. It also cites an Associated Press investigation finding that suicides in ICE detention have occurred at an unparalleled rate during the current term.
Why experts are worried
Todd Schulte, president of the immigrants' rights group FWD.us, told the outlet it is "basically bringing AI into the game of superpowering ICE even more." Steven Hubbard, a senior data scientist at the American Immigration Council who tracks ICE's AI use, warned that with biometric data, errors can enter a system and then become "a black box" where nobody knows how the AI reached a determination.
Immigration policy experts also raised concern about how the data could be used to target undocumented families outside detention. David Bier of the Cato Institute and Schulte were skeptical that the software would protect detainees' wellbeing. Bier said people die by suicide because they feel hopeless and cut off, and that feeling unable to speak freely with the outside world could make that worse.
These are the experts' views as reported, not findings about a system that does not yet exist.
The technical problem: voice matching is error-prone
Speaker recognition compares a voice with an enrolled profile and returns a score. Accuracy drops with phone-line compression, background noise, accents, illness and stress, and a fixed threshold trades false matches against false misses. A "receiver-side deception" flag is therefore a probabilistic guess about identity, and a false flag could implicate a relative who simply handed the phone to someone else.
Research on jail-call surveillance by the Surveillance Technology Oversight Project, in its report This Call Will Be Recorded, Monitored, and Used Against You, argues that automated speaker recognition can misidentify who is on the line and that call data is shared across agencies. That report is advocacy research, but it describes the same failure mode that the experts in the RFI coverage point to.
Video analytics has its own tradeoffs. Detecting a fall or a fight is a well-defined task, but self-harm is subtle and context-dependent, so a system tuned to catch it will also generate alerts that staff must triage. Alert fatigue can undermine the very safety benefit being claimed.
How this fits the wider AI surveillance picture
The proposal follows a pattern we have covered elsewhere on explainx.ai: AI turning existing data into automated leads. See the Graylark Raven interior-photo geolocation tool used by police, the backlash and layoffs at surveillance firm Flock Safety, and the debate over Palantir's role in Sankar's Face the Nation claims versus the record. AI labs are also drawing lines: Anthropic's 2026 usage policy restricts surveillance uses, which matters if a contractor wanted to build on a hosted frontier model. And in Congress, senators have pressed the White House on AI standards for testing and auditing, the kind of oversight that has not yet reached this use case.
A dome camera with a green indicator, representing AI video analytics that would flag falls and fights in detention facilities
What to watch next
- Responses and the next step. An RFI usually precedes a solicitation. Watch the contracting site for a formal request for proposals and any named vendors.
- Accuracy and audit requirements. Whether ICE specifies error rates, human review rules and independent testing will say a lot about how it expects to manage mistakes.
- Legal-call protections. Attorney-client calls are privileged. How the system would reliably exclude them is an open question.
- Data retention and sharing. Stored voice profiles are biometric data. Who can access them, and for how long, matters as much as the detection features.
- Oversight and litigation. Immigrants' rights groups and lawmakers may demand records or challenge the plan.
What to do if you work on or around this
- Builders and vendors: read your model and data terms. Many providers ban surveillance uses, and speaker identification on detainee data carries legal and reputational risk.
- Researchers and journalists: request the solicitation text and any privacy impact assessment. The details determine whether this is incident triage or broad profiling.
- Families of detainees: assume calls other than legal ones may be recorded and analyzed, and use attorney channels for privileged matters.
- Teams deploying AI agents inside organizations: guardrails matter whenever automated systems act on sensitive data. AgentBeam, from the explainx.ai team, is an agent security platform that stops AI agents before they take dangerous actions.
Based on The Washington Sun's October 13, 2026 report and an advocacy research report on call surveillance. The RFI is preliminary; details may change in any later solicitation.