Since December 2023, the Texas Education Agency has used an automated scoring engine as the first reader for written answers on the State of Texas Assessments of Academic Readiness (STAAR), with humans scoring at least 25 percent of responses and every response the engine flags as unusual or low confidence. The agency says the change was needed to control cost after a 2023 test redesign added six to seven times as many written responses, and it reports that the engine agrees with human scorers about as often as two humans agree with each other. In October 2026, rescoring requests from Texas districts reached 64,489 for reading, which keeps the question of trust open.
Key takeaways
- The engine scores every STAAR written response first and sets the score of record for roughly 75 percent of them; humans set it for the rest.
- Texas Education Agency (TEA) documents say hand scoring the redesigned test would have cost an extra USD 15 million to USD 20 million per year.
- Agreement figures come from TEA's own studies, and no independent audit of them was found.
- Reading rescoring requests rose from 2,758 in spring 2024 to 64,489 in spring 2026, and more than 13,300 scores went up after review.
- Spanish-language STAAR responses are scored only by people.
Why a redesigned test needed a machine reader
In 2023 Texas redesigned STAAR so that more questions look like the open-ended questions teachers ask in class. TEA's March 2024 presentation on hybrid scoring says the redesign meant six to seven times more constructed responses to grade each year, and that keeping full human scoring would have cost USD 15 million to USD 20 million more annually. The commissioner told a House committee in August 2022 that automated scoring was a necessary step to control costs while keeping scores reliable.
The test has two kinds of written items. Short constructed responses are scored on a zero-to-one or zero-to-two scale. The extended constructed response in reading and language arts uses a five-point rubric, with up to three points for development and organization of ideas and up to two for language conventions.
How a response is routed
TEA's scoring process document dated December 2023 describes the sequence. The engine is trained on student responses and human scores from field tests, about 3,000 hand-scored field-test responses according to the March 2024 presentation. It also assigns a confidence value to each score. Where a person scores a response, the human score replaces the engine's score as the score of record.
| Response type | Who sets the score of record | Reason given by TEA |
|---|---|---|
| Typical response the engine scores with confidence | The engine (about 75 percent of responses) | Cost and turnaround after the redesign |
| Random sample, at least 25 percent of each grade and subject | Two trained human scorers | Monitoring the engine, like a second reader |
| Low confidence, often a response on the border between two score points | Human scorers | Borderline answers should get a person's judgment |
| Condition codes: very few words, mostly copied or duplicated text, another language, off-topic, or vocabulary unlike the training set | Human scorers | Unusual responses may not resemble what the engine learned |
| Spanish-language STAAR | Human scorers only | The engine is not used |
TEA's own diagram labels the 75 percent share approximate and the human sample as at least 25 percent.
What the agency's accuracy studies show
Before adopting the engine, TEA re-scored spring 2023 responses, which humans had already graded, and compared results across five groups: male, female, Black, Hispanic or Latino, and White students. It concluded the engine met its performance criteria. The agency requires the engine to agree with human scorers at the same rate human scorers agree with one another, and to produce a similar spread of scores.
The March 2024 hybrid scoring presentation gives examples from the first operational run in December 2023. On one grade 9 reading extended response, exact agreement on the conventions trait was 69.3 percent between two humans and 73.9 percent between a human and the engine. Exact agreement on the ideas trait was 66.5 percent and 70.5 percent. On a grade 10 item the human-to-human figure for conventions was 78.4 percent and the human-to-engine figure was 87.9 percent. Exact agreement measures how often two readers give the identical score, and it says nothing about whether the rubric rewards the right qualities.
The zero-score dispute
Districts such as Lewisville reported more zeros on extended responses in December 2023, and the timing coincided with the engine. TEA's presentation separates two changes. The redesigned rubric, in use since spring 2023 under human scoring, allows a zero for a response that has no coherent structure or evidence, which earlier rubrics did not. Among retesters on the two English end-of-course tests, no more than 11 percent received a zero on the extended response in any 2022 administration. The figures were 50 percent and 62 percent in spring 2023, still under human scoring but with the new rubric, and 72 percent and 84 percent in December 2023 under hybrid scoring. TEA attributes the jump to the rubric change and to the retester population, not to the engine. For students who received a zero in December 2023, the agency reported that 14 percent had left the response blank, 30 percent were flagged by a condition code, and 55 percent received a zero under the rubric.
That breakdown does not settle every district's concern. It does show why a rise in zeros cannot be read as a measurement of engine error by itself.
The rescoring numbers of 2026
The Texas Tribune reported on October 7, 2026 that requests to rescore open-ended reading responses grew from 2,758 in spring 2024 to 21,620 in 2025 and 64,489 in spring 2026. More than 250 districts and charter networks filed requests. More than 13,300 reading scores improved after human review, and the newspaper reported that humans had originally scored 42 percent of those and the engine 58 percent. Dallas ISD submitted more than 5,100 reading tests, and 34 percent received a higher score. TEA told the newspaper that humans drive scoring in the hybrid model and that the engine does not learn from one response to the next. A University of Texas at Austin professor quoted in the story read the surge as a sign of declining trust in how scores are assigned.
The same report put the share of open-ended reading responses that changed at under 0.5 percent. The base for that percentage is not defined in the article, so it cannot be reconciled with the 13,300 figure here, and neither number is an overall error rate. Rescoring is also a selected sample: districts request it for scores they doubt.
Questions a district or parent can ask
- Which items were scored by the engine, and which went to people?
- What are the agreement figures for the specific item and student group, not only statewide?
- Does a rescore go to a person? In Texas it does, and the fee is waived when the score changes.
- Does the test affect a school's accountability rating, a student's grade, or promotion?
- Who audits the engine, and are the results public?
This article describes published agency documents and news reporting. It does not assess any student's score and is not legal advice.
FAQ
Is the STAAR scoring engine generative AI?
TEA says it is not. The agency describes a closed system trained on human-scored responses that does not learn from one response to the next. The Texas Tribune describes it as natural language processing, a related field of computing.
Can a family ask for a human to look at a STAAR written answer?
TEA's process document says district testing staff submit rescore requests when a district, parent, or guardian has concerns, and that rescores go to human scoring. A fee applies unless the score changes.
Do the rescoring numbers prove the engine is inaccurate?
No. They show that districts disputed more scores and that many of the disputed scores rose. Some originally came from humans, and the requests are a self-selected group.
Is similar scoring used in Canadian provincial tests?
No Canadian equivalent was verified for this article. Provincial assessment bodies publish their own scoring methods, and those documents are the place to check.
Related reading
- documented limits of AI in education
- the potential of AI in education
- AI and academic integrity
- how AI is used in learning
More in AI in education.
Last verified: October 2026
Sources
- Scoring Process for STAAR Constructed Responses, Texas Education Agency, December 2023
- STAAR Hybrid Scoring Key Questions, Texas Education Agency, March 2024
- Texas will use computers to grade written answers on this year's STAAR tests, The Texas Tribune, April 9, 2024
- STAAR rescoring requests from Texas schools triple in one year, The Texas Tribune, October 7, 2026