> daily_signal(2026_06_22)

Amazon called three engineers into HR the week after they testified against a data center, while Britain prepares to point face-scanning AI at children at its border.

PickBits Daily Signal · Monday, June 22, 2026

By Mark Pickering · 9 min read · June 22, 2026

// tl;dr

Three Amazon engineers were called into HR one by one after they testified at a Seattle hearing in favor of pausing new data centers. Darius Irani, Patrick Schloesser, and Liesl Wigand, all members of Amazon Employees for Climate Justice, spoke in early June; the council passed a one-year pause 9-0 on June 9. Days later all three were pulled into separate HR meetings over their testimony, one warned it could end in firing. Their advocates filed a civil-rights complaint with the city on June 17.
62 rights groups asked Britain to abandon plans to use AI facial age estimation at the border to catch asylum seekers it suspects are adults posing as children. Amnesty International, Human Rights Watch, Liberty, and the EFF among them wrote to asylum minister Alex Norris. Even the best systems are off by about 2.5 years right at the 16-to-18 line the Home Office wants to police, and the Home Office concedes accuracy varies by ethnicity. Rollout is slated for 2027.
Microsoft is testing a fine-tuned version of China's DeepSeek-V4 as a cheaper option inside Copilot Cowork, its Anthropic-built agent for Microsoft 365. The draw is price: Anthropic's flagship runs about $50 per million tokens against roughly $0.87 for DeepSeek, a 57-fold gap. The model would be hosted on Azure so data stays in Microsoft's cloud, arriving "in coming weeks" alongside a move to usage-based billing.
The FAA is paying Palantir nearly $4 million to point an AI at runway close calls, and it has already changed how planes land at San Francisco. The system reads hundreds of thousands of incident records to find recurring danger; after it flagged one, the FAA banned side-by-side parallel landings at SFO. The work started after the January 2025 D.C. midair collision that killed 67.

Three Amazon engineers went to a Seattle city hearing, said their piece about data centers, and went home. The next week, each of them got a call from HR. That scene — someone speaks up, and an institution with power over them decides what to do about it — runs through the whole signal today. It plays out at a national border, where Britain is about to let an AI guess whether an asylum seeker is a child. It plays out inside your own software, where Microsoft is quietly choosing which model does your work based on what it costs. And it plays out in the skies over San Francisco, where an AI just talked a regulator into changing how planes land. Four rooms, one question in each: who is the AI pointed at, and who gets to decide.

Today: Amazon pulled three engineers into HR after they testified against a data center, a coalition of 62 rights groups asked Britain to drop AI age-scanning of asylum-seeking children, Microsoft moved to swap a far cheaper Chinese model into Copilot, and the FAA handed runway safety to a Palantir AI that already changed how planes land at SFO.

1. Three Amazon engineers testified against a data center. The next week, HR called them in one by one.

The work beat, where speaking as a citizen turned into a meeting that could cost you the job.

Three Amazon software engineers — Darius Irani, Patrick Schloesser, and Liesl Wigand, all members of the employee group Amazon Employees for Climate Justice — testified at Seattle City Council hearings in early June, urging the city to rein in new AI data centers. On June 9 the council voted 9-0 for a one-year emergency moratorium on new data centers above 20 megavolt-amperes, while it studies the strain on the power grid, water supply, and utility bills. Days later, all three were called one by one into separate video meetings with an Amazon HR investigator looking into their public testimony; one was told the inquiry could lead to termination. On June 17, the employee group filed a complaint on their behalf with the Seattle Office for Civil Rights, alleging unlawful retaliation. Amazon's position, through a spokesperson, is that the three "may have been speaking in their capacity as Amazonians and not as private citizens."

This is the data-center backlash reaching the people who build the data centers. The local fights over the power and water these sites consume have been intensifying for a year, but the new development is who got pulled into HR: a company's own engineers, for speaking at a public hearing in the city where they live. Whether Amazon crossed a legal line will hinge on Seattle's specific ordinance and on whether the three identified themselves as Amazon employees when they spoke — which Amazon says they did. The chilling effect, though, does not wait for a ruling. Every other employee who watched three colleagues get investigated for civic testimony now has a clear sense of what showing up at the next hearing might cost.

Why this matters: The legal question is narrow and local; the pattern is not. Almost everyone who works for a large company has, at some point, a view their employer would rather they not air in public, and the line between "private citizen at a public hearing" and "employee speaking for the company" is one most people never check until it is too late. That is exactly the line Amazon is leaning on, and where it lands will tell every tech worker whether civic speech is safe. Action this week: Spend twenty minutes knowing your own footing before you need it. Pull up your employer's external-communications or political-activity policy and actually read it; look up whether your state or city has an off-duty-conduct or political-activity protection law (many do, including Seattle's, which is the one in play here); and if you ever speak publicly on something your employer cares about, say plainly that you are doing so as a private citizen, not on its behalf, because that single sentence is often the whole legal distinction. If you manage people, decide where your team's line sits now, before a hearing forces the question.

cnbc.com: Amazon investigating engineers who criticized AI data center expansion (June 18, 2026)
tomshardware.com: Amazon workers who testified against AI data centers say they were intimidated and face possible termination (June 2026)
engadget.com: Amazon is investigating three employees who spoke out against building more AI data centers (June 2026)

2. Britain is about to point face-scanning AI at the border to decide which asylum seekers are lying about being children.

The border beat, where a margin of error becomes a decision about a child.

A coalition of 62 organisations — among them Amnesty International, Human Rights Watch, Liberty, the Electronic Frontier Foundation, Foxglove, and the Open Rights Group — sent an open letter to Alex Norris, the UK's border security and asylum minister, urging the Home Office to abandon plans to use Facial Age Estimation at the border. The technology, slated for rollout from 2027 with testing this year, scans a face and estimates an age, to flag asylum seekers the government suspects are adults "pretending to be children." The catch is in the numbers: even the best-performing systems are off by roughly 2.5 years right around the 16-to-18 boundary — which is precisely the boundary the Home Office wants to police.

An age estimate is not a neutral fact here; it decides whether someone is treated as a child or an adult, with everything that follows for housing, safeguarding, and the asylum process itself. The signatories note that the Home Office concedes the tool's accuracy varies by ethnicity and skin tone, that trauma, malnutrition, dehydration, and long journeys can make a child look older than they are, and that it is unclear which images the system was trained on or on what lawful basis a vendor could have gathered photos of asylum-seeking children to build it. A 2.5-year error sounds abstract until it is the difference between a frightened sixteen-year-old being placed with other children or sent alone into the adult system.

Why this matters: Facial age estimation is arriving everywhere — app stores, adult sites, and social platforms are all under pressure to verify ages — and the border is where the technology meets the people least able to contest a wrong answer. How a government handles the error bar on its most vulnerable group sets the template for how it will handle yours. "The computer says you're over 18" is a sentence a lot more of us are going to hear, and the asylum case is the stress test for whether anyone can argue back. Action this week: If you build or buy age-verification or any biometric-classification tool, treat this as the spec to copy in reverse: write down the error rate at the exact threshold that triggers a consequence, the demographic breakdown of that error, and the appeal path for a person the model gets wrong — if you cannot fill in those three, you are shipping the Home Office's problem. If you are in the UK, the open letter and Foxglove's write-up are public, and reading them is the fastest way to understand what your government is about to deploy in your name.

eff.org: EFF joins 60+ groups urging the UK to halt face estimation at the border (June 2026)
foxglove.org.uk: Open letter to the Home Office on facial age estimation for asylum-seeking children (June 2026)
lbc.co.uk: AI facial recognition to detect asylum seekers posing as children in clampdown on those gaming the system (June 2026)

PickBits Daily Signal is free. If it lands in your inbox every day and it is worth something to you, the best way to support it is to forward it to someone who would read it. Subscribe at pickbitsai.substack.com.

3. Microsoft wants to swap the AI inside Copilot for a Chinese model that costs 57 times less.

The tooling beat, where the model quietly running your work could change underneath you.

Microsoft is testing a fine-tuned version of DeepSeek-V4, the latest open model from the Chinese lab DeepSeek, as a lower-cost option inside Copilot Cowork — the background agent, built with Anthropic, that runs multi-step tasks across Microsoft 365. The reason is money. Anthropic's flagship model costs Microsoft about $50 per million tokens; DeepSeek-V4 Pro runs about $0.87 — a roughly 57-fold gap that, across millions of enterprise users, is the difference between a feature that loses money and one that makes it. Microsoft says the open model would be optional and fully hosted on Azure, so customer data stays inside its own cloud under its compliance and data-residency controls, and that it expects to name its low-cost choice "in coming weeks," alongside a shift to usage-based billing and a new $99-per-user tier.

Two things are true at once. The cost gap is real and enormous, and "fully hosted on Azure" genuinely answers the obvious fear that using a Chinese model means sending your data to China — the weights would run on Microsoft's servers, not DeepSeek's. But "which model is doing my work" is about to become a question with a quietly shifting answer. The same Copilot task could be handled by a top-tier American model one month and a fine-tuned Chinese one the next, chosen on a cost curve you never see, with quality and behavior that are not identical. For most memos that will not matter. For the document where it does, you will want to know what wrote it.

Why this matters: This is the open-model endgame arriving in the most mainstream software there is. For two years the frontier labs could charge a premium because nothing else was close; an open model at one-fiftieth the price, good enough to run real work, changes the math for every company reselling AI — and Microsoft is the biggest reseller of all. The takeaway for anyone buying AI is that "powered by [famous lab]" is becoming a label to verify, not a fact to assume, because the vendor's incentive is to quietly swap in the cheapest model that clears the bar. Action this week: If your team uses Copilot, Cowork, or any AI feature you pay for, ask your vendor a plain question in writing: which model handles my data, where does it run, and will you tell me when that changes? For Microsoft 365 specifically, watch for the Cowork model announcement in the next few weeks and check whether the low-cost option is on by default or opt-in. And if you have been assuming an open model is too weak for production, this is your cue to actually test one on your own workload — the price gap is too large to take on faith in either direction.

the-decoder.com: Microsoft's Copilot Cowork moves to usage-based billing and may tap DeepSeek (June 2026)
axios.com: Microsoft explores DeepSeek for Copilot Cowork (June 16, 2026)
cryptobriefing.com: Microsoft shifts Copilot Cowork to usage-based pricing, considers DeepSeek model for enterprise AI (June 2026)

4. If you fly, an AI you've never heard of just changed how planes land at San Francisco.

The safety beat, where AI did the unglamorous thing you actually want it to do.

The FAA is paying Palantir nearly $4 million to point its Foundry data platform at runway safety, FAA Administrator Bryan Bedford said. The system ingests hundreds of thousands of incident records and looks for patterns in the close calls between aircraft on and near runways. It has already produced one concrete change: after the AI flagged a recurring risk, the FAA banned parallel landings, two planes descending side by side onto adjacent runways, at San Francisco International. The agency began the work after the January 2025 midair collision over Washington, D.C. that killed 67, and against a run of near-misses since, including a plane striking a fire truck at LaGuardia and another clipping a light pole on approach to Newark.

This is the unflashy, genuinely useful version of "AI in government": not a chatbot, but pattern detection across a pile of safety reports too large for any human team to read, surfacing one specific risk that a regulator then acted on. It is also worth watching with clear eyes. Foundry is a Palantir system whose workings the public cannot see, and the SFO ban was made on the AI's reading of risk in records the rest of us cannot audit, so there is a fair question about how much a safety regulator should defer to a model's pattern-match. But on what we can see so far, the AI did the job you actually want it to do here. It read everything, found the recurring danger, and handed a human the decision to ground a risky kind of landing before it became the next headline.

Why this matters: If you fly, a recurring danger at one of the country's busiest airports was caught and acted on before it caused a crash, which is a rare piece of good news in a year of aviation near-misses. The wider lesson is the method, not the magic: the win here was concatenating every incident report and asking a machine to surface the patterns a human reading them one at a time would miss, then putting a person in charge of the call. Action this week: If you manage operational risk of any kind — a clinic's incidents, a warehouse's injuries, a fleet's near-misses — the move Palantir made for the FAA is now doable with off-the-shelf AI on your own logs; you do not need a $4M contract to dump a year of incident reports into a model and ask "what keeps almost happening here, and where." And if you just want to see the data behind aviation's safety record, NASA's Aviation Safety Reporting System at asrs.arc.nasa.gov is public and searchable, the same kind of raw report pile the FAA's AI is now mining.

aol.com (AP): FAA is turning to AI to reduce the number of close calls between planes at the nation's airports (June 2026)
cryptobriefing.com: FAA partners with Palantir to enhance runway safety using AI (June 2026)
thenextweb.com: Palantir, Thales, and a startup are competing to build the FAA's predictive air traffic AI (June 2026)

» What to watch this week

Whether the Seattle Office for Civil Rights acts on the Amazon complaint, and whether the investigations are dropped or escalate. The case will turn on a narrow city ordinance, but the precedent (can a company investigate employees for speaking at a public hearing) reaches every large employer and every worker who has ever considered showing up to one.
Whether the Home Office pauses the facial-age-estimation rollout, and whether it ever publishes the error rates and training data behind the tool. The science the rights groups cite is not in dispute; the open question is whether a government deploys a tool it admits is biased and imprecise on the people with the least power to appeal a wrong answer.
Which low-cost model Microsoft actually picks for Cowork, and whether it is default or opt-in. The bigger tell is whether other AI resellers follow Microsoft in quietly trading frontier models for cheap open ones, and whether any of them commit to telling customers which model is running.
Whether the FAA expands the Palantir model to more airports, and whether the SFO landing ban holds. One concrete change at one airport is a pilot; the question is whether AI pattern-detection on safety data becomes standard practice, and whether a ban made on a model's reading of the records survives the inevitable pushback from airlines.

Tomorrow's signal lands here.