Batch Scanning Essays for AI – What’s Your Workflow at Scale?
what works for 25 students may not work for 1,500. that’s the problem I’m facing as the person our department asked to figure out our AI detection workflow.
We have 8 English teachers each handling 3-4 classes of 28-32 students. Running individual essay checks through any detection tool is not scalable. and if we’re being honest, the results of individual checks aren’t reliable enough to use anyway.
I’ve tried three approaches this term:
1. Random sampling (5-10% of submissions): low overhead but you miss things. not really a policy, just a chance.
2. Turnitin batch submission: works at scale if you already have the subscription. adds AI score alongside plagiarism check. most practical option we have.
3. Manual triage first: teachers flag essays that feel off, then test only those. reduces volume but introduces selection bias.
None of these are satisfying. What are large schools doing?
4 Replies
Join the discussion.
Log In to Reply
the random sampling approach is not a policy its a lottery. and if you only scan some essays, which ones you flag is more about luck than cheating. that creates its own equity problem.
my workflow: batch scan first pass with flagging anything over 60%, then manual review of flagged essays for context. the batch scan saves me from reading 30 essays carefully up front - i only do close reading on the ones that flag. cuts review time roughly in half even accounting for manual follow-up.
does any batch tool actually handle PDFs well? every tool I've used has formatting issues when essays come in as PDFs through our LMS. the text extraction is messy and i'm not confident the scores are based on clean input.
The honest answer from a curriculum policy perspective: at scale, detection is not a viable primary integrity strategy. Turnitin batch works logistically but the accuracy limitations don't improve at scale - you get the same 15-20% false positive rate across 1,500 essays, which means potentially 225-300 false flags. The workflow for those conversations would overwhelm any department.
What scales better: assessment design that doesn't require detection. Process documentation, oral components, assignment-specific prompts. These require upfront work but don't generate the false-flag handling overhead.