Texas graded millions more essay answers using an algorithm first

The classroom note

In December 2023, the Texas Education Agency published its own Scoring Process for STAAR Constructed Responses, describing a change already scheduled for the redesigned STAAR test: 'student responses to short constructed-response (SCR) questions and extended constructed-response (ECR) questions are scored using a hybrid scoring model,' meaning an automated scoring engine, or ASE, scores every response first, and at least 25 percent are then routed to trained human scorers as a check. For a Texas classroom, the redesign that made this necessary was itself large: TEA's own March 2024 presentation records that the redesign added roughly 13.6 million more written responses a year to grade, moving the state from about 2.2 million to 15.8 million scored constructed responses annually.

What the evidence says

Both documents are TEA's own account of its process, not an independent audit of scoring accuracy. The agency states that responses the engine flags as 'low confidence' or carrying a 'condition code' (unusual vocabulary, blank answers or off-topic language, for example) are routed to a human scorer, and that any human-scored result becomes 'the score of record'. Its separate Hybrid Scoring Key Questions document states the cost reasoning directly: maintaining full human scoring for the redesigned test 'would have cost $15-20M more per year' than the hybrid model, a figure TEA attributes to its own commissioner's August 2022 testimony rather than to an outside audit.

The implementation question

The mechanism has a built-in appeal: TEA's own document states that 'if district personnel or a parent or guardian has concerns about a student's score on a constructed-response question, district testing personnel can request that the student's response be rescored for a fee,' sent for human scoring, with the fee waived only if the score changes. That places the cost of doubting the machine on the family or district first, refunded only if they turn out to be right. The agency also states that Spanish-language STAAR responses are scored entirely by humans, not the ASE, an exception the document does not explain in cost or accuracy terms.

What holds and what fails

What holds, on TEA's own account, is a genuine second-reader structure for at least a quarter of responses and for anything the engine itself flags as uncertain, a real check rather than a rubber stamp. What is likely to fail, editorially, is treating the roughly 75 percent of responses never seen by a human as equivalently verified; the agency's own design accepts the engine's first score as final for most students, a trade it frames explicitly around the $15-20 million annual difference rather than around a stated accuracy target for the unreviewed majority.

• Would this district actually advise a family to request a paid rescore, and does it know the fee-waiver condition?
• What share of this school's students fall into the 25 percent routed to a human scorer, and is that known?
• Why is the Spanish-language exception handled differently, and does that affect confidence in the English-language model?

A hybrid scoring model is, by TEA's own description, a budget decision with a quality-control minority built in, not a claim that every score has been read by a person. That distinction is worth stating plainly to any family asking who graded their child's answer.