
...
By Markus Goldbach
Founder
0
Share
Newsweek is a Trust Project member
See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.
I spent the first decade of my business career building plagiarism-detection tools after a career in neuroscience.
The logic was simple. If a student copied a paragraph, the source existed, the date existed, and you could prove the misconduct. The evidence held up in a hearing. Nobody could argue with a matched sentence and a publication date.
Generative AI ended that. There is no source to match. There is no paragraph to trace. So-called AI detection software can produce a number, but it cannot explain what it found or why it flagged the text.
Read More on Leadership Insights
A student can be accused on the strength of that number alone, and the university can lose in court when it tries to defend it. In the end, only the scope and consequences were discussed, not the evidence.
The evidence of failure is now substantial. According to HEPI, when seven widely used detectors were tested on 91 essays written by non-native English speakers, they wrongly flagged 61 percent of that genuine human writing as AI-generated.
The error is not random. It falls hardest on the students least equipped to contest it.

...
The problem goes beyond detection accuracy. The BBC reported that in November 2026, England’s higher education regulator, the Office for Students (OfS), and Advance HE published research drawing on the views of 244 higher education staff and 91 students from at least 127 institutions.
The research found that "a significant proportion" of assessments are unsupervised and many "can be completed to a high standard using AI tools with limited engagement from students," sometimes in minutes, without effectively assessing the knowledge and skills they are designed to measure.
It also concluded it was "frequently possible" for students to use AI "in ways that cannot be reliably detected."
The implication is that universities cannot reliably establish what a student has actually learned by examining the final submission alone.
Documenting the writing process addresses this challenge by preserving the assessability of extended written work without banning AI.
Institutional Response
Institutions worldwide are responding. According to Inside Higher Ed, for instance, Indiana University's Kelley School of Business now states that detection tools are not approved for use because they are "highly unreliable" and can produce false positives on human writing.
Others have switched the feature off. That is a sensible retreat, but it is only a retreat. Walking away from a broken instrument does not tell a professor what the student actually learned.
Here is what the detector cannot see. A student who writes an essay over six weeks leaves a record. The first draft is confusing. The second draft is still confusing but differently. Somewhere in the fourth draft, the argument collapses and gets rebuilt.
That sequence is the learning. It is also the one thing a paraphrasing tool cannot fake, because the tool does not produce a history of thought. It produces a finished product.
This is also why AI humanizers make things even worse. They can make AI-generated text appear more natural, vary sentence structures, and reduce the likelihood of detection.
A system that evaluates only the finished text can always be played by a better tool. One that examines the writing process asks a different question: how did the work develop, who shaped it, and can the student demonstrate ownership of the final result?
The current system ignores that record. It asks faculty to judge a single PDF at the end of the line, then hands them a probability score to help. The score is unreliable, and the PDF is silent about how it came to exist.
So the professor is left with two bad options. Trust the number and risk accusing an honest student. Ignore the number and risk missing a dishonest one.
A Potential Solution
There is a third option. It requires universities to treat the drafting process as part of the assessment. Not as surveillance, and not as a trap for students who use AI to polish their grammar. As evidence.
A student who used AI to check syntax and then rewrote the paragraph in their own words has done something different from a student who generated the essay from a prompt. The first leaves a work trail. The second, just a copy-and-paste event.
I am not arguing against AI. Students should use these tools the way earlier generations used a roommate, a parent, library services, or a writing center.
The problem is that the current system cannot distinguish between help and substitution, so it treats both as cheating. Students respond by hiding their AI use. Faculty respond by distrusting everything. The relationship that makes education work erodes from both ends.
The stakes extend past the campus. A degree is a claim about what a graduate can do. If employers stop believing that claim, the degree loses its value, and the students who worked honestly lose the most.
Countries that depend on intellectual capital rather than natural resources will feel this first. Innovation requires people who can stay with a hard problem for months. It does not require people who can write a prompt in thirty seconds.
Keeping Credibility
Some universities have moved to oral exams as a substitute. That solves one problem and creates another. Oral assessment rewards students who think quickly and speak comfortably.
It penalizes students who think slowly and write well. And while oral examinations are a valid option in certain cases, they also do not bring the learning goals of critical thinking, thorough research, and optimizing arguments, just like a writing exercise does.
A system that replaces one bias with a different bias is not a fix.
The honest answer is that assessing writing in the age of AI is harder than it was, and there is no tool that makes it easy again. Universities can keep buying detectors that do not work, or they can throw away the substantial learning effects of academic writing and working.
The first option is a waste of money, and the second option a waste of talent.
Detection was always a proxy for a question nobody wanted to ask directly. What has this student actually learned? The universities that answer that question honestly will keep their credibility.
The ones that decide to stay in the dark or throw away essential teaching content will not.
Markus Oskar Goldbach is the founder of Mentafy, which sells academic-integrity technology focused on writing processes. All opinions are his own.
View 0 Comments
Submit Correction
View Editorial & AI Guidelines