The prose is unusually polished. The citations feel generic. The argument is competent but strangely impersonal.
You suspect AI was involved, but suspicion is not evidence.
A detector appears to offer an answer. Upload the paper, receive a percentage, and decide what to do next.
Unfortunately, the percentage does not solve the problem.
Why Are AI Detectors Unreliable?
AI detectors attempt to infer whether a piece of writing resembles machine-generated text. They do not directly observe who wrote it, which tools were used, or how the work was produced.
That distinction matters.
Human writing can resemble AI-generated writing, especially when it is formal, predictable, heavily edited, or written by a multilingual student. AI-generated writing can also be revised until it resembles ordinary human prose.
As a result, detection tools can produce both false positives and false negatives.
A false positive can subject an honest student to an accusation they may struggle to disprove. A false negative can give faculty false confidence that an AI-generated submission is authentic.
Research on assessment in the age of generative AI argues that current detection methods cannot provide the reliability needed to enforce prohibitions consistently. When institutions depend on these tools, they risk promising faculty a level of certainty that the technology cannot deliver.
The Deeper Problem Is Not Detection
Suppose a detector were perfectly accurate and confirmed that a student used AI.
You would still need to answer several questions:
- Was AI permitted?
- What did the student use it for?
- Did the student understand the final answer?
- Did AI replace the learning the assignment was designed to develop?
- Can the student evaluate or defend the output?
The presence of AI does not automatically establish academic misconduct. Its absence does not automatically establish learning.
A student may use AI within the rules but rely on it so heavily that little intellectual development occurs. Another may use AI as a legitimate editing or brainstorming tool while retaining responsibility for the substantive reasoning.
That is why the distinction between cheating and over-reliance is so important. Cheating violates rules or trust. Over-reliance interferes with development and can happen even when no policy was broken.
A detector is not designed to distinguish between those situations.
What Should Professors Do When They Suspect AI Use?
Start by avoiding an accusation based solely on style or a detector score. As we've written before, AI didn't kill cheating — it killed visibility, and detectors don't restore it.
Instead, gather evidence about the student's understanding.
Ask the student to explain:
- the central claim,
- why they selected particular evidence,
- how they reached a conclusion,
- what alternatives they considered,
- and what they would change under a different set of conditions.
These questions should be tied to the student's actual submission.
A generic question such as "Tell me about your essay" gives a prepared student room to repeat memorized language. A specific question such as "You describe this factor as the primary cause. Why is it more important than the second factor you mention?" provides better evidence.
The goal is not to catch the student in a contradiction. It is to determine whether the submission represents capabilities the student actually possesses.
Replace AI Detection With Capability Verification
Capability verification asks a different question.
Instead of:
"Did AI generate this?"
ask:
"Can this student demonstrate the knowledge, reasoning, and judgment this assignment was intended to measure?"
That may involve a brief oral follow-up, a live application question, a revision completed under observation, or a comparison between the student's process and final work.
Researchers describe this as a structural change to assessment. Rather than relying only on instructions that students may follow or ignore, the assessment itself requires additional evidence of capability. Interactive oral questioning is specifically identified as one example of this kind of structural redesign.
How to Make Existing Assignments More Reliable
You do not need to eliminate essays, projects, or take-home work.
Instead, add one or more lightweight verification points.
1. Ask for a process artifact
Require a short outline, decision log, annotated draft, or explanation of how the student approached the task.
Do not assume the artifact proves authorship. Use it to create better questions about the student's decisions.
2. Add a short oral defense
A five-minute conversation can reveal whether the student can explain the work, respond to a variation, and connect the submission to course concepts.
The conversation should not simply ask the student to summarize what they wrote. It should probe the reasoning behind it. This is the idea behind oral defense software: every submission gets a short, individualized defense.
3. Ask students to critique an AI-generated alternative
Give students a plausible but flawed answer and ask them to identify its weaknesses.
This assesses whether they can evaluate AI output rather than merely produce polished content.
4. Separate permitted tool use from required human judgment
Be explicit about what students may delegate and what they must still demonstrate.
"You may use AI to generate possible approaches. You remain responsible for selecting the approach, checking its assumptions, and defending your decision."
That is more useful than a broad statement that AI is either allowed or prohibited.
5. Use repeated, lower-stakes checkpoints
A single final submission places too much weight on one artifact.
A sequence of short checkpoints creates a more credible picture of learning over time.
Researchers describe this as an evidentiary chain: validity comes from several connected demonstrations rather than one supposedly AI-proof task.
What AI Detectors Can and Cannot Do
AI detectors may sometimes provide a signal worth investigating. They should not be treated as proof.
They cannot reliably establish:
- who completed the work,
- what the student understands,
- whether AI use was educationally appropriate,
- or whether the learning outcome was achieved.
Those are assessment questions, not detection questions. We keep asking the wrong question — and detection is the wrong answer to it.
The safest approach is to redesign assessment so that the grade does not depend entirely on the authenticity of an unobserved final artifact.
That may include an AI-resistant assessment, an oral defense, a process checkpoint, or a live application of the same concepts.
The goal is not to win an arms race against AI-generated text.
It is to gather better evidence of learning.