The Short Answer
ProPublica and The Texas Tribune reported on 1 October 2026 that 21% of students passed state tests at an Austin charter using Alpha School's AI program. The lesson is not that AI fails in schools. It is that the growth evidence districts saw came from the vendor's chosen test and self-published paper, while the independent measure reported a year late. Districts should own the evaluation data, which on ibl.ai means you own all the code and the data.
The circulating version of this story is that an AI school replaced teachers and students failed. That reading is tidy, partly wrong on the facts, and gives a superintendent nothing to do differently on Monday.
The usable version is narrower and harder. Two measurements of the same instructional model disagreed by an enormous margin, and the favourable one was produced by the party selling the model.
What did ProPublica actually find at the Austin school with the 21% pass rate?
The 1 October 2026 investigation, reported by Lexi Churchill for ProPublica and The Texas Tribune, is about how the Texas Education Agency helped a private AI education company reach public and charter schools.
The number at its centre belongs to a specific campus. "Morath's agency knew about results in at least one of those charters, Texas Preparatory School in Austin," the piece reports. "Last year, state test scores compiled by his agency showed just 21% of students passed."
The agency gave that campus its third consecutive F rating, mostly based on those scores.
The reporting also establishes the promotion that ran alongside it.
Commissioner Mike Morath responded positively to the company in September 2025, visited its campuses with agency staff, offered personal introductions to Houston ISD superintendent Mike Miles, and had staff help draft materials explaining the program for distribution to districts.
This happened after the company's own charter application was rejected 10 to 3.
Three districts are taking it on, and the selection is worth noting.
Aldine will deploy the platform to roughly 300 students across four middle school campuses that need support in math. Fort Davis, a district enrolling around 800 students in total, will use it in K-8 classrooms for students needing extra help in reading and math.
Houston ISD plans the opposite selection. Its press office said the district will use the technology for students with strong test scores, "to evaluate whether this instructional approach can accelerate academic growth."
At least five other districts were approached and declined, including Ector County, Fort Bend, Irving, Pecos, and Pflugerville.
We covered the company's own account of its model when its leadership presented at ASU GSV 2026, and the Texas policy context in the same session series.
Did the AI platform cause the 21% pass rate?
No source establishes that, and the claim should not be made. This is the first place the story gets reported badly in the summaries now circulating.
The reason is an absence, not a counter-fact. The reporting never says when Texas Preparatory School adopted the program. It places the program in the year that produced the 21%, and says nothing about the two earlier years that made this the campus's third consecutive F.
So the record supports no causal direction in either direction. It does not show the program caused the failure, and it equally does not let the company argue the school was already failing on its own before the platform arrived.
What the record does support is narrower and still damaging: the program did not rescue an already-failing school, and the agency promoting it to other districts held the scores showing that.
The newsrooms report the same pattern in a second jurisdiction. Students "performed significantly worse on state-mandated exams than Alpha leaders predicted" at the one charter the company got approved, in Arizona, and at the existing Texas school where students used the program.
That is a prediction-versus-outcome failure, which is a cleaner and more serious finding than a causal one. The company forecast specific results on tests it did not control, and the results came in materially below the forecast in two states.
It is also worth being precise about "replaced teachers." Alpha retains adults, which it calls guides, and describes core academics as completed in two hours daily with guides focused on motivation and support. The design reassigns the adults rather than removing them.
Why do a vendor's "2.6x growth" claim and a state's 21% pass rate disagree?
Because they are different tests, administered by different parties, with different consequences. This is the Edge of the whole story, and it is checkable in about ten minutes.
The company's public claims are specific.
Its program page states that "On average, Alpha students grow 2.6 times faster than peers on nationally normed MAP tests" and that "Our best students achieve up to 6.5x growth." The same page carries a 99th percentile heading above the separate sentence "The majority of students consistently outperform national averages."
Its FAQ adds the gate: "Students achieve 90% mastery in each concept before they move on."
Those figures come from MAP, a nationally normed adaptive assessment, not from the state accountability test.
Then look at where one of those numbers was published.
The 99th-percentile claim circulated through a page labelled "Sponsored Content" whose own footer states it was produced "with information solely provided by the local business as part of their 'sponsored content' purchase through our advertising team."
The reporting adds that when asked for evidence, the company "referred to a self-published paper." The company declined interviews and did not answer detailed questions.
David DeMatthews, an education professor at the University of Texas at Austin, told the newsrooms it was "disheartening" and "painful" that the agency was promoting the product.
He put the evidentiary problem plainly: "It'd be one thing if they were approved as a charter and they did have a track record and they could show evidence. We don't have any of those sorts of things."
So the evidence available to a district buyer divided cleanly:
| Evidence | Instrument | Administered by | Reported |
|---|---|---|---|
| 2.6x growth, 99th percentile, up to 6.5x | MAP, NWEA's nationally normed adaptive test | Chosen and proctored by the program | Continuously, in the vendor's framing |
| 21% passed, third consecutive F | State accountability test | The state | Once a year, after the cohort was taught |
Both can be literally true at once. A mastery-gated adaptive platform can move a student up its own ladder quickly while that student still fails a state exam with different content coverage, different format, and no retries.
The point for a buyer is not which number is honest.
It is that the favourable number was produced on an instrument the seller chose, under conditions the seller controls, and reported in the seller's framing, while the unfavourable one arrived from the only party with no stake in the outcome, twelve months late.
What should a K-12 district require before an AI pilot touches a classroom?
The governance layers a district needs do not ship with the software, a gap we have written about for open-source agents in K-12. The evaluation layer is the one this story indicts, and it reduces to five requirements.
1. Name the instrument before the pilot starts. Written into the agreement: which assessment decides whether this worked. If the answer is the vendor's internal mastery metric, the pilot cannot fail.
2. Require alignment to the test that carries consequences. A district is graded on its state exam. An adaptive platform normed nationally is not automatically covering the same standards, and no reporting documents a specific Texas standards mismatch here, so do not assume one. Verify coverage against your standards, in writing, before signing.
3. Run an interim independent check. Annual state results reporting after the fact is the structural defect. A district needs a mid-year measure it administers, so a failing cohort is detectable inside the pilot rather than after it.
4. Hold the raw data. Not a dashboard view. Exportable per-student, per-session records in the district's possession, so an outside evaluator can recompute the vendor's growth figure.
5. Keep the adults in the loop with something to act on. Reassigning teachers to motivation only works if they receive interpretable signal about who is stuck and why.
Note what is not on this list.
The circulating version of this story alleges that the campus had no teacher training, no curriculum alignment, no feedback loops, and no human governance checkpoints. The reporting establishes none of those four things.
They may be true and they are not established, so they are not a basis for anyone's procurement policy.
Who should own the evaluation data in a district AI pilot?
The district. This is the part that is an architecture question rather than a contracting question, because a vendor cannot hand over evaluation data it never gave the buyer custody of.
On a self-hosted stack the measurement apparatus is district property by construction.
Evaluations run scoring and human annotation inside the district's own deployment and export. Transcripts retain what every agent session actually did with a student.
Data Reports generate on demand and download as CSV for an independent evaluator. The audit log records which user took which action, searchable by user and date range, inside the district's own deployment.
None of that is an argument that the technology in this story would have worked under different governance.
It is a narrower claim: a district running its own stack can answer the question "did this work" with data a vendor cannot curate, and can answer it in month three instead of month fourteen.
That capability follows from the deployment model.
ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
The one sentence worth keeping
A pilot whose only evidence lives in the vendor's dashboard is not a pilot. It is a testimonial that took a year of instruction to produce, and in Austin the bill for it arrived as a third consecutive F.