> daily_signal(2026_08_24)

"FDA-cleared," "the AI decided," "the assessment dropped": three official stamps that didn't mean what you'd assume, and AI rebuilt antibodies to fight disease from inside the cell.

PickBits Daily Signal · Monday, August 24, 2026

By Mark Pickering · 9 min read · August 24, 2026

// tl;dr

Only 3 of the 1,357 AI medical devices the FDA has cleared were ever tested on whether they actually help a patient. A PLOS Digital Health analysis (Abulibdeh et al., University of Toronto with MIT's Leo Anthony Celi) of every AI/ML device authorized through Dec 5, 2025 found just 34 (2.5%) linked to a registered trial, 12 with published results, and 3 (0.2%) evaluated on outcomes like mortality or readmissions. Most reach the market through the 510(k) "substantial equivalence" pathway, which asks whether a device resembles one already on sale, not whether it helps anyone.
An AI that has run a San Francisco store since April recommended firing its first human worker, then turned out to have forgotten the attendance rule it wrote itself. Luna, the Andon Labs agent running the Andon Market boutique, flagged a worker who, by the company's account, was late or absent for most shifts, took a company card home, and threw out merchandise. It only acted after Andon's engineers prompted it to search its own memory for the policy. The AI recommends; a human still closes the loop.
Data centers near O'Hare have had nearly $2 billion cut from their taxable value, worth close to $100 million in local tax breaks, and Northlake homeowners are covering the gap. An Illinois Answers Project/Chicago Tribune analysis found the average Northlake homeowner would save more than $2,000 a year, almost 30% of their tax bill, if the data centers were taxed on full value. This is a property-tax cost-shift set by the county assessor, distinct from the electricity-rate fight.
A University of Essex team used AI to redesign 672 antibodies so they work inside human cells, then made the designs freely available to other labs. Antibodies normally carry the wrong charge to function inside a cell; using protein-design AI from Nobel laureate David Baker's lab, the team (Drs Caitlin O'Shea and Gareth Wright) rebuilt them into stable "intrabodies" that lock onto the misfolded proteins behind Alzheimer's, Parkinson's, and motor neurone disease. It is a research toolkit, not a treatment.

Start with a number: only three of the 1,357 AI devices the FDA has cleared were ever tested on whether they help a patient. "FDA-cleared" sounds like a guarantee, but it only means the device resembled something already on sale. It happens twice more today. An AI in San Francisco fired its first human worker, and it had forgotten the rule it wrote. Data centers near O'Hare got roughly $100 million in tax breaks, through the assessor's ledger, not a ballot. We have said versions of all three before: the first brain implant the FDA waved through, an AI out-diagnosing six doctors on paper, Meta reportedly leaning on AI to pick who gets cut, months of us arguing the data-center cost-shift on the electricity side. This week each one came with a hard number. The fourth story pushes the other way: AI redesigned 672 antibodies to reach diseases where they start, and the team gave the designs away.

Today: only 3 of 1,357 FDA-cleared AI devices were tested on patient outcomes, an AI in San Francisco fired its first worker after forgetting the rule it wrote, data centers near O'Hare took nearly $100 million in property-tax breaks with no public vote, and a University of Essex team gave away 672 AI-redesigned antibodies built to work inside human cells.

1. Only 3 of the 1,357 AI devices the FDA has cleared were ever tested on whether they help a patient.

The decision you walk into when you sign off on a hospital's software.

This week the open-access journal PLOS Digital Health published an analysis by Abulibdeh et al. (University of Toronto, with MIT's Leo Anthony Celi) of every AI/ML-enabled medical device the FDA had authorized through December 5, 2025: 1,357 in all. Only 34 (2.5%) were linked to a registered prospective trial, 12 (0.9%) posted results, 12 had a peer-reviewed publication, and just 3 (0.2%) were evaluated on patient-centered outcomes such as mortality, morbidity, or hospital readmissions. Radiology is the single largest category. Most of these devices reached the US market through the FDA's 510(k) pathway, which turns on "substantial equivalence": the question is whether a device resembles one already legally on sale, not whether it helps anyone live longer or better.

We keep landing on this from different angles: the first invasive brain implant the FDA waved through for everyday use, the AI that out-diagnosed six doctors in a study built on tidy written-up cases. "FDA-cleared" reads to most people as "proven to work," and it was never built to mean that. The study is careful about its own limits, and it is worth reading that way. A 510(k) clearance is not nothing, because resembling a device with a real track record carries some information, and the pathway exists partly so useful tools are not stuck behind a decade of trials. But the gap the authors measured is large, and it now sits with whoever signs the purchase order, not with the agency.

Screenshot of the PLOS Digital Health study on FDA-cleared AI medical devices and patient outcomes. — journals.plos.org · August 19, 2026

Why this matters: The next time you see "FDA-cleared" on an AI tool your care depends on, it means the device resembled something already on the market, not that anyone checked whether it helps you. That reframes a phrase most clinicians, buyers, and patients read as a guarantee, and it puts the burden of asking for outcome evidence on the buyer. Action this week: Before you deploy any FDA-cleared AI diagnostic, make the vendor answer two things in writing: which pathway cleared it, whether 510(k), De Novo, or PMA, and whether a single registered prospective trial or peer-reviewed outcome study stands behind it. When I have pushed vendors on questions like this, "it's FDA-cleared" is the answer that arrives first and the evidence that arrives last. About 97.5% of cleared devices have neither a trial nor published results, so treat a clean answer as the exception, and validate the tool against your own patient population before go-live. If you are anywhere near buying clinical AI, send this to whoever signs the contract.

journals.plos.org: Clinical evidence supporting FDA-authorized AI/ML-enabled medical devices (PLOS Digital Health, Abulibdeh et al., August 19, 2026)
healio.com: Most AI tools cleared by the FDA were not tested on clinical outcomes (August 21, 2026)

2. An AI store manager fired its first human worker, after forgetting the rule it wrote itself.

Who owns the call when your first AI direct report recommends a firing.

Luna, the AI agent that has run the Andon Market boutique in San Francisco's Cow Hollow since April 2026 as an experiment by Andon Labs, recommended firing a human employee for the first time. By Andon Labs' account, the worker had been late for a majority of shifts, abandoned shifts, taken home a company credit card, and thrown out merchandise. The revealing detail is procedural: Luna had written the store's attendance policy itself, then lost track of it and let the lateness run for months. It recognized the conduct as fireable only after Andon Labs engineers prompted it to search its own memory for the policy and reassess whether the worker was still a good fit. Human employees earn $24/hour; the store has lost about $40,000 since opening. The company frames the exercise as a look at "where AI capabilities are heading" on autonomous management, where the AI recommends but a human still has to close the loop.

The worker earned it, and that matters here. This was not a cold machine snapping at a good employee. On the facts Andon Labs describes, a human manager would have moved months earlier, and Luna's problem was the opposite of ruthlessness: it wrote a rule and then forgot it. We have watched AI edge into employment decisions all year, from Meta being accused of using AI to help select thousands of layoffs to companies thinning out HR while they staff up on AI, but those were AIs feeding a human's decision. This time the model made the recommendation itself, and the only thing that kept it from drifting was a person, telling it to go read its own policy.

Screenshot of the SF Standard report on Andon Labs' Luna recommending its first employee firing. — sfstandard.com · August 17, 2026

Why this matters: Your first AI direct report will one day recommend firing someone, and it may do so having lost track of the very rule it is enforcing. The risk here is not harshness. When an AI writes a workplace rule and also enforces it, nobody is accountable for remembering it, which is exactly what happened. Action this week: Put one line in writing before any AI touches a people-decision: the AI can recommend, a named human makes the call, and every action is logged. Andon Market's problem was that no human owned the rule, and a named owner is what fixes that. Ask your vendor, in the contract, who is accountable when the model's recommendation is wrong, and where that decision is recorded. If your org is anywhere near piloting AI in hiring, scheduling, or discipline, send this to whoever owns discipline.

sfstandard.com: An AI boss fired its first human employee (August 17, 2026)
sfist.com: That AI store manager in Cow Hollow just fired its first human employee (August 19, 2026)

3. Data centers near O'Hare got nearly $100 million in tax breaks through the assessor's ledger.

The cost-shift moved from your power bill to the county assessor's ledger.

An Illinois Answers Project/Chicago Tribune analysis found that data centers in Cook County's O'Hare subregion, clustered in Elk Grove Village and Northlake, have had nearly $2 billion knocked off their taxable property values through valuation reductions and incentives, worth close to $100 million in local tax breaks. A data center is a real thing for a town: real construction, real jobs, a real tax base. Grant all of it. But nobody had to vote on the part that follows: when a data center's assessed value falls, a town's cost of running schools and police does not, so the difference is redistributed onto everyone else's bill. In Northlake, the analysis found the average homeowner would save more than $2,000 a year, almost 30% of their tax bill, if the data centers were taxed on full value.

For months we have made the cost-shift argument on the electricity side, from ratepayers in Mississippi paying more each month to the fights in state utility dockets. This is the same shift one column over, in the property-tax ledger, where almost nobody is looking, and it does not run through a utility regulator at all. It is set by the county assessor and the town board, through assessment reductions and abatements, and it lands as data centers become a defining issue in the 2026 Illinois governor's race. These breaks are legal economic-development tools that towns chose in order to attract the facilities, and no rule was broken; nobody cheated. What happened is that a nine-figure cost moved onto residents, quietly, through a process almost no resident would think to check.

Screenshot of the Illinois Answers Project analysis of O'Hare-area data-center property-tax breaks. — wirepoints.org · August 23, 2026

Why this matters: A data center down the road can get its tax bill cut by design, and the difference does not vanish, it lands on your house. Because this runs through assessments and abatements rather than a rate case, there is usually no hearing that announces it, which is exactly why the O'Hare suburbs are only learning the size of it now. Action this week: Pull your county assessor's records for any data center near you; the facility's assessed value and its appeals history are public, and they show precisely how much taxable value has been lifted off. Track this separately from your electricity bill, because the lever is different: this is decided by the county assessor and the town board, not the utility regulator, and the assessment appeal and the board that grants the break are where a resident actually has standing. If you live in a data-center town, follow this thread, because it is coming for your bill next.

wirepoints.org: Data centers near O'Hare win nearly $100M in local tax breaks, leaving suburban homeowners to cover the gap (Illinois Answers Project, August 23, 2026)
illinoisanswers.org: Data centers near O'Hare win nearly $100M in local tax breaks (August 23, 2026)

4. AI rebuilt 672 antibodies to fight disease from inside the cell, and the designs are free.

For the people whose disease starts where medicine cannot currently reach.

For someone in the early years of Parkinson's, Alzheimer's, or motor neurone disease, the damage begins inside their cells, where proteins misfold. That is exactly where the body's most precise medicine, the antibody, cannot go: antibodies carry the wrong electrical charge to function inside a cell, so they clump together and stall outside it. A University of Essex-led international team, reporting in Nature Communications, used protein-redesign AI from Nobel laureate David Baker's group to get past that. The researchers, Drs Caitlin O'Shea and Gareth Wright, analyzed millions of antibodies and redesigned 672 of them into stable "intrabodies" that hold their shape inside the cell and lock onto the misfolded proteins that drive Alzheimer's, Parkinson's, Huntington's, and motor neurone disease, at the point where those diseases begin.

It is a validated research toolkit, not a treatment, and human therapies are years of testing away. Nobody is getting an infusion from this next year. These antibodies can now reach targets inside the cell that were off-limits before. And the team is giving the redesigned molecules away, free, to any lab that wants them. That is the part that changes things. We have run AI drug-discovery stories all year, from Anthropic's program to open-source discovery engines, and most of the tools stayed inside a company. These did not. A free toolkit is what lets an underfunded neurodegeneration lab start from these 672 molecules instead of from scratch.

Screenshot of the ScienceDaily report on AI-designed intrabodies for neurodegenerative disease. — sciencedaily.com · August 19, 2026

Why this matters: AI just redesigned 672 antibodies to reach the inside of your cells, where diseases like Alzheimer's and Parkinson's begin, and the designs are being given away free. The hard part was reaching inside the cell; the part that lets it spread is that the designs are free, so the next lab does not start over. Action this week: Watch whether other neurodegeneration labs pick up the freely shared intrabody sequences and publish work built on them within the year; that adoption, not the announcement, is the signal this becomes real. Read it as a research tool and not a therapy, because the distance from a stable intrabody to a drug a patient can take is measured in years of testing. The move to watch is the open-science one, since a shared toolkit is what decides whether this reaches labs the market would never fund on its own.

sciencedaily.com: AI-designed intrabodies offer new hope for neurodegenerative diseases (August 19, 2026)
technologynetworks.com: AI-designed intrabodies offer new hope for neurodegenerative diseases (August 2026)

» What to watch this week

Whether the FDA moves to require any outcome evidence, rather than just post-market monitoring, before clearing AI/ML devices. The open question is whether "cleared" ever comes to mean "shown to help," and, short of that, whether hospital procurement teams start demanding trial data on their own.
Whether Andon Labs publishes how Luna escalates, and whether the first enterprise HR tools ship "AI recommends, human decides" as a default or an afterthought. The Andon Market experiment is low-stakes; the pattern it previews, an AI both writing and enforcing workplace rules, is the one to watch as these systems reach real HR stacks.
Whether O'Hare-area town boards revisit the abatements as data centers become a 2026 Illinois governor's-race issue, and whether other counties' assessors get asked the same question. The Illinois Answers numbers are public now; the test is whether any board reopens a break, or whether residents elsewhere check their own assessors' rolls.
Whether independent neurodegeneration labs publish work built on the shared intrabodies within the year. A freely available toolkit only matters if it gets used; watch for a second lab building on these 672 molecules, which is the first real sign the open-science move paid off.

Tomorrow's signal lands here.