Texas’ AI Grading System Faces Major Issues
State exams are designed to assess students’ abilities and overall grasp of the curriculum. But what unfolds when the grading process breaks down? Texas recently encountered this dilemma when an AI-based grading system resulted in numerous errors and a massive recount.
It’s widely known that AI can make mistakes. This is a cautionary note that accompanies nearly every interaction with leading chatbots today. Nevertheless, society seems inclined to rely on AI for increasingly significant responsibilities. In this instance, Texas assigned the grading of students taking the State of Texas Assessments of Academic Readiness exams to AI tools, and the outcome was somewhat predictable.
As people grow more skeptical of AI, their trust in results produced by AI systems wanes.
The AI grading system grabbed attention in 2024 when Texas claimed it would save $15 million by substituting human evaluators with “technology similar to ChatGPT” for a portion of the exams. This system combined human and AI efforts, utilizing humans for the multiple-choice questions while AI handled the more extensive open-ended written responses.
Whenever the AI expressed “low confidence” regarding a student’s score, those responses should have been flagged for human review. However, if the AI was confident—similar to how chatbots confidently mislead—no alert would reach the human graders.
During a trial run in December 2023, educators were surprised to see a “dramatic increase” in failing scores among their students. Parents and students were charged $50 if they wanted to appeal a score, unless the new score ended up being higher than the initial one, which would waive the fee.
Initially, Texas received 2,758 rescore requests by spring 2024—fairly notable given that appeals of this nature were uncommon at that time. Yet, this figure skyrocketed to 21,620 requests by 2025, and by 2026, it reached an astonishing 64,489 requests, translating to about $3 million in expenses faced by parents hoping to improve their children’s scores.
Future rescore requests are estimated to cost between $15 and $20 million.
After human review of the tests, the number of corrected scores increased steadily each year, with 13,300 scores—nearly 20%—improving on average. Remarkably, none of the grades were lowered.
This situation might reflect a clear failure of the AI grading system, but a deeper look reveals inaccuracies caused by both AI and human evaluators. Of the identified errors, 58% originated from AI, while 42% stemmed from human mistakes. So, eliminating AI might lead to fewer grading errors, but wouldn’t eradicate them altogether.
This scenario invites an intriguing question: despite the presence of both AI and human errors, the fact that AI graded the extensive written sections made students and parents doubt their final results, contributing to the surge in rescore requests. As distrust in AI grows, people are more inclined to question any outcome linked to it. Until such inconsistencies are resolved, they will remain a lingering consequence of relying on AI.
For the Texas education system, one must ponder whether the anticipated savings are worth the issues, distrust, and ongoing rescore requests likely to trouble the state for years ahead. Currently, Texas has no intention of halting the automated AI grading process.
