You tried it once. Probably a Sunday night, buried in essays, with a deadline Monday morning. You pasted a student's paper into ChatGPT, asked it to "grade this essay out of 100 with feedback," and waited.
What came back was… fine. Technically coherent. A little generic. The feedback was encouraging in a way that didn't match the student. The score felt arbitrary. And you had a nagging feeling the model had missed the thing this specific assignment was supposed to teach.
So you closed the tab and kept grading the old way.
That experience killed AI grading for a lot of teachers. It shouldn't have. The model wasn't what failed. What you wrapped around the model was.
What you actually tried
When you paste "grade this essay out of 100" into a general chatbot, here's what's missing.
Your rubric. The model doesn't know what criteria you grade on, how they're weighted, or what a 4 looks like versus a 3 in your classroom.
Your standards. "Proficient" in your department might mean something very different from what the model averaged across the internet.
The assignment itself. The model is grading the response without seeing the prompt, the unit context, or what you taught leading up to this.
Your voice. Your feedback has a particular tone. Pushing this student, encouraging that one, asking a real question of a third. A chatbot writes feedback that sounds like a well-meaning stranger.
Classroom judgment. You know that Marcus wrote this at 11:47 PM. You know Aaliyah just came back from a week out sick. You know Period 3 got the mini-lesson and Period 5 didn't. None of that is in the prompt.
What you got back was the model's best guess without any of it. So yeah, it was generic.
The part nobody names: prompt engineering is work
"Just tell the AI what you want" sounds simple. It isn't.
Writing the instructions that turn a general model into a rubric-aware grader that sounds like you takes, by our count, several thousand words of careful prompting. There are dozens of edge cases to handle: blank submissions, wrong file formats, rubric criteria that don't quite fit a student's approach, late work with partial credit rules, students who answered a different question than the one you asked.
You have to teach the model to cite evidence in the student's work instead of hallucinating a quote. You have to teach it to be strict on the criteria that matter and generous on the ones that don't. You have to teach it not to write a 400-word paragraph when a two-sentence comment is what you wanted.
It's a craft. Software people call it prompt engineering, and it's most of the work in making AI grading useful. It is not something a busy teacher should have to do at 11 PM on a Sunday.
The quality of AI grading is almost entirely determined by the prompt wrapper around the model, not by which model you pick. Nobody warns you about that part.
What a good wrapper does
Between you saying "grade this assignment" and grades showing up in Google Classroom, here's what has to happen.
- Pull the assignment prompt so the model grades the response against the actual question.
- Pull your rubric, or build one from the criteria you entered, and weight it the way you specified.
- Apply your strictness setting. Harsh grader, encouraging, standards-based, whatever you picked.
- Handle each submission's format. Google Docs, PDFs, images of handwritten work, mixed uploads.
- Write feedback in your voice, at the length and tone you set, not the chatbot's default.
- Check its own work for consistency across the pile, so student 47 isn't graded harder than student 3 because the model got tired.
- Flag edge cases for you to look at manually. Blank submissions, possible plagiarism, responses that don't match the prompt.
- Give you every grade to review before anything goes to students, because you are still the teacher.
A thousand small decisions, packaged. Each one is the difference between "generic and slightly wrong" and "this is how I would have graded it." Miss a few and you get what you got the first time you tried ChatGPT.
Why ClassLens exists
ClassLens is the packaging.
Steven Swanson built it because he'd been in the classroom for 22 years and knew, from the inside, what goes into grading a stack of essays well. He also knew that asking a general chatbot to do that work, with no rubric logic, no strictness calibration, no feedback voice handling, no edge-case flagging, no Google Classroom integration, and no teacher review step in the middle, was never going to produce something teachers would trust.
So the product is not "AI for grading." It's:
- Every submission graded against your rubric and your prompt
- Feedback in your voice and at your length
- Flagged for anything that needs your eye
- Delivered to a dashboard where you review every grade before a single student sees it
- With a Knowledge Gap Report at the end telling you what your whole class actually misunderstood
The teacher-review step is the one that makes this safe to use. Nothing auto-returns. You are always the last set of eyes on every grade.
If you quit on AI grading, try again
Most teachers bounced off their first attempt because they tried a general chatbot on a task that needs specialized packaging, and then concluded "AI can't grade."
AI can grade. A general chatbot, used raw, can't, because it's missing the teacher's context, the rubric logic, and the classroom judgment that turn a language model into a grading assistant.
The base model with no wrapper is one thing. The model with 22 years of classroom instinct wrapped around it is a different thing.
Try it
ClassLens works with Google Classroom. Sign in, pick an assignment, set your rubric and strictness, and let it grade. You review every grade before it reaches a student. When you're ready, you release the batch in one click.