AI & Immigration Zero Percent. The company behind your state’s AI assistant also holds a $5.7 million contract with ICE. And nobody can prove your data stays where they say it does.

This week, the Senate passed its version of the PROTECT Act, and I support every word of it. It bans ICE from making warrantless arrests at courthouses, schools, hospitals, daycares, and churches. It prohibits new 287(g) agreements, which are the contracts that allow local agencies to enforce federal immigration law. Three thousand of us stood on Beacon Hill in January and demanded this. The House passed it 134-21, and the Senate delivered. That fight mattered, and I am proud of everyone who showed up for it.

This newsletter is about what I found on the other side of the wall.

In December, the Department of Transitional Assistance launched a pilot that transcribes SNAP eligibility calls in real time using AI. The system listens to the conversation between the caseworker and the caller, recording everything: their name, their address, their situation. It generates a summary and saves it directly into the caller’s benefits record, where it becomes part of the case file that follows them through the system. DTA says callers are notified and can opt out, but no law required the disclosure, and the notification process was never subject to public review.

That pilot is one of at least 40 state agency functions, including MassHealth claims processing and school complaint intake, that now use AI to process resident data, according to an investigation by The Shoestring. The tools are live or being rolled out across the executive branch.

The company that set all of this up also holds a $5.7 million contract with ICE.

You may remember my February newsletter. I pulled the procurement documents the administration did not want to release (I asked three times before they produced the actual records). The scoring rubric weighted AI safety at zero percent, meaning no one evaluated the system for bias, error rates, or whether it actually produces accurate results. Thirteen vendors applied. Nine were eliminated before scoring. Both finalists came through a single reseller called Carahsoft Technology Corporation, a company the FBI raided in September 2024 as part of a bid rigging investigation. I published the full breakdown here.

MA AI Procurement Scoring Rubric — Safety Weight
AI Safety Evaluation Bias, error rates, accuracy, data protection	0%
Source: Massachusetts AI procurement documents obtained via public records request.

The procurement was the first problem. The vendor is the second.

In September 2025, while Massachusetts was finalizing the AI contract, ICE signed a $5.7 million, five-year deal with Carahsoft for a product called Zignal Labs, an AI platform that scans over eight billion social media posts per day. Zignal uses machine learning and computer vision to build what it calls “curated detection feeds,” which are automated watchlists that flag accounts for investigators to review. The ACLU’s Patrick Toomey called it black box technology deployed without any accountability.

Zignal Labs also works with the Israeli military. A company pamphlet from 2025 advertised its use for tactical intelligence for operators on the ground in Gaza.

The company that brokered your state’s AI assistant is the same company that provides ICE with its social media surveillance platform. That does not mean the two systems are connected, but it does mean that the company operating in both spaces has no independently verified wall between them, and neither does the state.

I asked the administration for the enforcement mechanism, the independent audit, the technical verification, anything that would let the state prove its promise that your data stays “within the four walls,” and they could not point to one.

The PROTECT Act was written to close gaps like these. It closed some of them.

The Act draws a line around physical spaces, from courthouses and schools to hospitals and houses of worship. It addresses law enforcement cooperation: the badge, the warrant, the arrest.

It does not address the digital infrastructure underneath, the systems that process your benefits calls, your health records, and your case files every day.

The PROTECT Act bans new 287(g) agreements. The existing Department of Correction agreement stays in place. Under that deal, corrections officers can notify the federal government about people being released from custody who do not have lawful immigration status. The legislature voted to preserve it.

The Carahsoft contract, the SNAP call transcription, and the forty-plus AI functions touching your data also remain in place. None of it was subjected to the oversight you would expect: no independent safety evaluation, no vendor background check that would have flagged an active FBI investigation, and no democratic input of any kind.

Nobody planned it this way. The legislature focused on the physical side of enforcement, and the digital infrastructure went unchecked.

The cost of that gap is already visible. Immigration detentions in Massachusetts increased 336 percent from 2024 to 2025. Nearly 80 percent of people picked up by ICE have no criminal history. School enrollment has declined more than five percent in districts with high populations of immigrant students. Families are afraid to go to work, run errands, or send their children to school.

In a state where families are already pulling back from the services that keep their children fed and healthy, deploying AI into those systems through a vendor that also sells ICE surveillance tools, under a contract that no one can independently verify, demands more than a promise.

You do not need a data breach to cause harm. You need a family that does not call about their child’s benefits because they are not sure who is listening. The chilling effect is the damage, and it is happening now.

Critics will say there is no evidence that data has flowed from the state AI tool to ICE, and they are right. But that is not the point. A contractual promise from an administration that chose a vendor under FBI investigation, scored safety at zero, and cannot produce an enforcement mechanism is not the same thing as a safeguard.

This is not a new principle, and your community already knows it.

In Medford, the city council passed an ordinance requiring democratic approval before any government agency buys surveillance technology, including social media monitoring software. Cambridge passed the same kind of ordinance. Somerville was the first city on the East Coast to ban government facial recognition entirely.

Those ordinances say something clear: the people affected by surveillance technology have a right to decide how it is used.

If your city requires a public vote before police buy a license plate reader, why doesn’t the state require one before deploying AI into your benefits calls?

The state deployed AI into your benefits calls, your health records, and your case files, through a vendor that serves ICE, scored safety at zero percent, and asked nobody. Your city would not allow that for a surveillance camera. Your state did it for the systems that process your family’s most sensitive information.

I am writing legislation that extends the principle your city already established to the state level: independent safety evaluation before AI touches resident data, vendor vetting that includes active federal investigations, and democratic oversight from the people whose data is at stake rather than a hand-picked task force with financial ties to the winning vendor.

This is not a ban and it is not a moratorium. It applies at the state level the same principle your city already enforces locally.

This is what oversight looks like: public records requests, line-by-line contract review, months of asking the same questions until the administration produces actual answers instead of general assurances. I am running for State Senate to bring it there, and this is how you put me there.

Bring This Fight →

The PROTECT Act is a victory and it is the right wall. I am proud of the families who showed up, the organizers who mobilized, and the legislators who voted yes. But walls need to cover every side, and if we believe the state should not cooperate with ICE at the courthouse, we should also ask whether the company processing your SNAP call should be the same company building ICE’s surveillance platform.

We should be able to verify the answer, not just take someone’s word for it. I checked, and what I found is not reassuring, so here is every document I have.

Share this with your neighbors. Every source is linked. And if you want this kind of oversight in the State Senate, put it there.

Follow on Instagram · Follow on Bluesky