Teachers use AI to grade student work. It’s harsher than they are.

Teachers are embracing new AI grading tools, saying the programs let them give students faster feedback and more chances to practice. Used properly, teachers say, AI helpers can provide consistency and remove bias from assessments of student work—although not everyone trusts AI to give out grades.

Education-focused AI startups tend to offer grading in writing-heavy disciplines like English and history, along with some in math and science. These bots generate a numeric score and offer up critiques on topic sentences, persuasive arguments and other elements. Teachers can choose to use the AI feedback as a guide, or pass the feedback directly to students. They say they typically tell parents and students when they use the programs.

“Does this make my life easier? Yes," said high-school English teacher Richard Vanden Bosch. “But that’s not what this is about. It’s about making the students better writers."

Others say the technology isn’t reliable enough yet to be used for something as high stakes as grading, which can affect college admissions and other life choices. Today’s tools can be glitchy or grade too harshly, critics say.

“It should not be used for grading," said Alex Kotran, co-founder of the AI Education Project, which teaches AI literacy. “It’s going to undermine trust in the education system."

Putting the bots to the test

To see how the graders work, we put a Wall Street Journal colleague’s 12th-grade English paper into three tools. The essay, on the oppression of the character Ophelia in Shakespeare’s “Hamlet," earned her a 97% in 2013.

One site, AutoMark, gave the paper 97%, and later 100%. The two others came down harsher, with Class Companion rating it 62%, and CoGrader judging it an 85%.

“Consider refining the focus of each paragraph to ensure that every sentence directly supports the main point," Class Companion urged. CoGrader advised: “Some sentences are quite long and could benefit from being broken down into shorter, more digestible parts," citing a passage quoting from “Hamlet."

The startups say the tools have gained traction over the last school year and are continually being refined. Class Companion co-founder Avery Pan said teachers can train the AI by overriding grades, or can hide the numeric score altogether. CoGrader offers a button to make feedback kinder.

The essay’s original grader, Scarsdale, N.Y., high-school English teacher Stephen Mounkhall, said some of the feedback sounded potentially believable as a teacher, but that the robot is incapable of viewing the nuances of a student’s progression.

“I think it’s really offensive that this has been invented," said Mounkhall, a teacher for three decades. “I don’t see how such a machine could consider the human element of what would make this person a better writer."

‘I don’t want to make them cry’

Sixth-grade English teacher Stephanie Galvani said she likes that the AI tools help keep her from bringing bias into her grading. She’s one of few early adopters in her small Massachusetts district, where other teachers have called AI dehumanizing.

“I feel a bit like I’ve been saying, ‘Hey, did you know I’ve been doing heroin and you might want to try it?’" Galvani said. “It’s verboten."

After asking students to write a mystery where a detective walks into a crime scene, Galvani told the grader to look for vivid verbs and multisensory descriptions. When an essay included blood “dripping" down the wall, the bot offered up “cascade" as a more interesting verb.

The AI can offer too much criticism, Galvani said, and she deletes suggestions about less important skills so students aren’t overwhelmed.

“They’re sixth-graders, I don’t want to make them cry," she said.

Colorado high-school English teacher Andrew Gitner said he’ll adjust AI-generated comments to sound more like something he would write. “I would never use the word ‘delve,’" he said, but the AI constantly does. Other teachers say they push the AI feedback straight through to students after becoming comfortable with the products.

Proceeding with caution

AI startup founders and industry CEOs say even as the tools become more trustworthy on individual assignments, teachers should always play an active role in final grades. “We are not at all trying to remove the teacher," said Andrew Goldman, co-founder of Writable, a writing and reading program now owned by Houghton Mifflin Harcourt.

One of the most popular AI companies aimed at teachers, MagicSchool AI, says it’s intentionally not developing ways for its technology to grade.

Founder Adeel Khan said he doesn’t think AI is ready to be relied on for grading, though he’s OK with AI giving writing feedback. He likened AI grading to self-driving cars; the technology may be possible, but not everyone is comfortable with it.

Incoming high-school senior Isaiah Green in California’s Central Valley said he grew to love the use of Class Companion in his Advanced Placement U.S. History class. He credits the grader with giving him confidence when he walked into the AP exam in the spring, because he’d spent weeks getting instant feedback after completing timed essay prompts.

The AI program let students challenge grades they thought were wrong. He said he saw classmates go from a 2 out of 7 to a grade from their teacher of 6 out of 7. “There could be a dramatic difference," he said.

Keri Rodrigues, the president of the National Parents Union, said parents can’t expect classrooms to be stuck in an analog past.

Rodrigues recently asked ChatGPT to grade an essay her sixth-grade son was about to turn in. The bot told her it would get a B minus.

The actual teacher’s grade exceeded expectations. Her son got an 87.

Write to Sara Randazzo at sara.randazzo@wsj.com