How Students Try to Avoid AI Detection

Mid-terms are just around the corner and many students are looking to get ahead by using AI for papers, quizzes, and discussion assignments. But as schools and universities get better at catching cheaters using AI on assignments, students are get smarter at avoiding getting caught.

Is AI Editing Cheating?

Students may edit an AI generated essay in efforts to fool AI detectors, introducing grammar and punctuation errors, removing certain words and phrases, and paraphrasing whole sentences and paragraphs. They test against free online AI detectors like Copyleaks or ZeroGPT until it’s been edited enough that it’s no longer recognized as AI writing. It’s altogether time-consuming—time a student probably could have spent writing an original paper. But where there’s a will, there’s a way.

Fooling the detector with editing isn’t completely foolproof. Pangram’s AI detector tool is still triggered even after the text is completely rephrased. That’s because it detects tiny markers in structure, organization, and tone, not just words. These flags cannot be completely removed through editing. Students certainly try, but it doesn’t work—modern AI detection is accurate against edited AI text.

Humanizers

Students also use AI programs called humanizers, such as Grammarly, Quillbot, and Undetectable AI, which reword, rephrase, and reorganize AI generated text to make it sound more natural and more human.

The good news is that humanized text can be recognized by the human eye. You may see unnatural and out of place synonyms, like “snipping edges” instead of “cutting corners.” Also indicative are extra or missing spaces and non-standard characters, tricks used by humanizers to bypass detectors.

In addition, humanizers are not perfect against AI detectors. Pangram is adept at detecting humanized text, performing with an accuracy rate of above 90% for all tested humanizer programs.

Version History

Version history in Google Docs displays all edits made to a document with timestamps, which can reveal if text was copied and pasted. Pangram’s Writing Playback feature also allows users to view Google Docs history.

When academic misconduct is suspected, students may be asked by their professor or an academic integrity office to show proof of their writing process and style, including outlines, notes, previous drafts, and Google Docs history. To pass this common evidence check, students who used AI may fabricate the writing process. Instead of copying and pasting, they can manually type an AI generated essay into Google Docs and pretend to edit it, then use it as evidence it was human written. They may also write false notes or fake an outline. Moreover, programs called auto typers will simulate the human writing and editing process in a Google document to make it appear like the student did it themself.

While Google Docs history can be used as evidence as to whether a student’s essay is original or plagiarized, it should not be used in isolation, especially when determining the outcome of a formal academic misconduct case. In addition to clever students falsifying their writing process, a student on a tight deadline may not have an outline or notes at all if they hurriedly typed the final draft in one sitting. Other methods for verifying if a student used AI are more reliable.

Submission Strategies

AI content detectors are often integrated into Canvas or Google Classroom where students turn in papers. To avoid the detector entirely, a student may try to submit an assignment by email or in-person instead. They may come up with an excuse, like “Canvas stopped working,” “my WiFi went out,” or “I handwrote it because my dog ate my computer.” They might try to turn in a corrupted file or an alternative file format so the detector can’t read the text. At the very least, it buys them some time while they troubleshoot or negotiate with the instructor.

Communicating clear expectations is key in preventing academic misconduct. Teachers should specify the requirements for submitting essays in the class syllabus and in assignment instructions, including the location, acceptable file formats (e.g. .pdf, .docx), and the deadline. Explicitly state you will not accept essays that don’t follow these requirements.

On the other hand, students do have trouble with technology sometimes. Being open to discussing academic concerns may also help prevent dishonesty. Refer students to Canvas tech support, your IT department, and other school resources for help with software and hardware issues. Computer labs and libraries are great places for students to get access to a device and an internet connection. If applicable, work with your school’s disability office to help a student receive appropriate accommodations that work for your course.

Common methods for avoiding AI detection are all either preventable, detectable, or subvertible. Schools are still a step ahead of students in the AI arms race… at least for now.