Prepare students for a machine-scored writing test the same way you would prepare them for a careful human reader: answer the exact prompt in their own words, explain the evidence instead of copying it, and write enough to show their thinking. Those habits match the rubric, and they also keep a response out of the categories a scoring engine is built to flag. Skip the tricks. Length padding, stuffed vocabulary and pasted passage text don’t raise a score, and some of them get an answer pulled for review or scored lower.
That’s the short version. The longer version matters because a lot of us now teach ECRs and SCRs knowing that a computer will be the first reader, and some educators who read their own students’ work are disagreeing with the scores that come back.
Why Are Teachers Questioning Machine-Scored Writing?
Texas is the clearest example. Since December 2023, the state has scored STAAR short and extended constructed responses with an automated scoring engine first, then sent at least a quarter of responses to human scorers. In spring 2026, districts asked for 64,489 rescores of written reading answers, up from 21,620 the year before and 2,758 the year before that, according to reporting from The Texas Tribune. More than 13,300 reading scores went up after people reread them, though fewer than half a percent of all open-ended reading responses changed.
Both numbers are true at once, and that’s the frustrating part. Most scores hold. But if you’re the teacher who read a student’s ECR in March and knows it was a 3, a 1 on the score report feels like a door slammed shut. Texas isn’t alone, either. Ohio publishes its own machine-scoring process, where the engine is trained on 2,000 human-scored field-test responses and a quarter of operational responses are also scored by hand.
So the questions teachers ask are fair. Do I teach differently now? Do I warn students? What do I do when a score looks wrong?
What Does the Scoring Engine Actually Look At?
The engine doesn’t “read” the way you do. In Texas, it’s trained on student responses that human scorers already rated during field testing, and the state requires it to agree with humans about as often as humans agree with each other. The TEA scoring process document also lists the kinds of responses the engine sends to a person instead of scoring on its own.
| What the engine flags | What it looks like in a student response | What to teach instead |
|---|---|---|
| Blank or too few words | Two sentences on an ECR, or a restated question | Write enough to prove the claim, every time |
| Mostly duplicated text | The same sentence repeated, or a pasted paragraph twice | Say a point once, then explain it |
| Mostly stimulus material | Long quotes from the passage with little commentary | Quote short, explain long |
| Vocabulary unlike the training responses | Off-topic words, heavy slang, or an answer about a different text | Stay on the prompt’s text and topic |
| Off-topic or off-task patterns | A personal story when the prompt asks for an argument | Answer the question that was asked |
| Low confidence | A response sitting between two score points | Make the claim and reasoning easy to find |
Responses with a condition code or a low-confidence score go to trained human scorers, and when a person scores a response, that score counts. In other words, the strange, borderline answers are the ones most likely to be read by a human. The flat, thin answers are the ones most likely to be scored by the machine alone.
That table is also just a list of things a good rubric already punishes. Nothing in it is a secret robot preference.
Should You Teach Students to Write for the Machine?
No. Teach them to write for the rubric, and say so out loud.
Picture an 8th-grade class two weeks before the spring test. A student has heard that “the computer likes long answers,” so her ECR practice runs 400 words. One hundred and fifty of those words are passage text. Her claim shows up in the last paragraph. She’s done exactly what she thought the test rewards, and her response would struggle with a human scorer too.
Here’s the trade-off worth naming for students. Length helps only when the extra sentences are explanation. Fancy vocabulary helps only when it’s precise. A machine trained on human scores is, at best, a copy of a careful human reader. Writing for that reader is still the right target.
Build Responses That Are Easy to Score Accurately
The goal isn’t a formula. It’s a response where the claim, the evidence and the reasoning are easy to find, so neither a scorer nor an engine has to guess. RACE, ACE and CER all do this if students use them as thinking tools rather than fill-in-the-blank frames.
Try this sequence on your next SCR or ECR practice:
- Answer in the first sentence. The student’s claim should respond to the prompt’s exact verb. If it says “explain how the author develops,” the first sentence names how.
- Quote short. One phrase or one sentence from the text, not a paragraph. Students who use evidence effectively usually quote less than they think they need to.
- Explain more than you quote. Two or three sentences of reasoning for every quoted line. This is the part engines and people both reward, and it’s the part students cut first under time pressure.
- Use the prompt’s key words. If the prompt asks about “the narrator’s conflict,” the response should talk about the narrator’s conflict, not “the main character’s problem” for three paragraphs.
- Proofread for conventions last. In Texas, the RLA ECR rubric gives up to 3 points for development and organization of ideas and up to 2 for conventions. Conventions are worth points, but they come after the ideas.
A quick before and after makes the difference visible to students.
Before: “In the story the narrator is sad. The text says ‘She stood at the window for a long time and did not answer when her brother called her name, because she was thinking about the letter.’ This shows she is sad.”
After: “The narrator’s silence shows that the letter has changed how she sees her family. When she ‘did not answer’ her brother, she is choosing to stay inside her own thoughts instead of reaching out. That choice sets up the conflict in the next scene, when she finally tells him what the letter says.”
The second version quotes less and explains more. That’s the whole lesson.
Calibrate Your Own Scoring Against the State’s Anchors
It’s easy to score practice ECRs a point off from the state and not find out until results come back. Fix that before the test, not after.
Pull the released scoring guides and anchor papers for your grade level. Score five anchor responses on your own without looking at the official scores, then compare. If you’re consistently off by a point on one trait, adjust your feedback now. The SCR and ECR rubric guide walks through how those trait levels usually break down.
If you use an AI tool for first-pass practice feedback, hold it to the same check. Grade Our Essays is aligned to 500+ state rubrics and drafts trait-level feedback, and you review and approve every score before students see it. That review step is the point. You stay the person deciding what a 2 looks like.
What Can You Do When a Score Looks Wrong?
Start with your district testing coordinator, because rescore requests usually go through the district, not the classroom. In Texas, district testing staff can ask for a constructed response to be rescored for a fee, the rescore goes to human scorers, and the fee is waived if the score on that question changes. Requests have to go in during a set window, so ask about the dates early.
Your job is to bring evidence, not a hunch. Keep a short file for students you’re worried about:
- two or three recent practice responses on similar prompts, with your rubric scores
- a note on which trait you expected to score higher, and why
- any accommodations the student had on test day
Be realistic with families and students too. Most scores don’t change on rescore, and a rescore can’t fix a response that didn’t answer the prompt. But a district that sees a pattern across several teachers’ students has something worth acting on.
Try This Tomorrow
- Show students the “What the engine flags” table and ask them to find one row their last practice response would trigger.
- Give a “quote short, explain long” mini-lesson: students cut a quote to its key phrase, then add one sentence of reasoning.
- Score five released anchor papers with a colleague before you score your next class set.
- Start a one-page evidence file for any student whose practice scores and test scores have not matched in the past.
- Ask your testing coordinator when this year’s rescore window opens, and write it on your calendar.
Frequently Asked Questions
Are state writing tests scored by AI?
Some are. Texas and Ohio, for example, use automated scoring engines for written responses, with a share of responses also scored by people. Check your state department of education for how your test is scored.
Does writing more get a higher score on a machine-scored test?
Not by itself. Engines are trained on human-scored responses, so length helps only when the extra sentences add explanation. Very short responses can be flagged, and padded ones still score on the rubric.
Should I tell students a computer will score their essay?
Yes, briefly, then move on. Students do better when they hear that the target is the same rubric a careful human reader would use, not a set of tricks.
Can a teacher request a rescore of a state test essay?
Usually the request goes through district testing staff, often during a set window and sometimes for a fee. Ask your testing coordinator how it works in your state.
Are Spanish-language responses machine scored too?
It depends on the state. In Texas, written responses on the Spanish-language STAAR are scored entirely by people.
