AI detector alternative
Two answers to the same submission. Only one of them is gradable.
What a detector returns
A probability. No audit trail.
An AI writing detector hands back a likelihood that the text was machine-generated. There is nothing under it: no record of the student’s reasoning, nothing to read, nothing the student can answer to.
What a defense returns
A transcript. A rubric score.
Tolus runs a short oral defense and returns the student’s own words alongside a rubric-based score, each sub-score citing a quoted moment. You read the receipt before you grade it.
If the question is whether the student understands the submission, measure that directly. Detection guesses at authorship; a defense shows the understanding, in evidence a teacher can stand behind.
Detection vs verification
Detection answers the wrong question.
The detection workflow asks whether the text was AI-generated. The classroom workflow needs to know whether the student understands the submission. Only the second question is gradable, and only the second one survives a parent meeting.
Why a probability falls short
A detection score, sometimes broken down by paragraph, carries no audit trail and no explanation the student can respond to. The teacher is left to investigate or, more often, ignore. Even a high probability says nothing about whether learning happened.
Why a defense holds up
A live conversation in the student’s own voice, scored against the teacher’s rubric, with each sub-score citing a quoted moment. Defensible, auditable, and immediate. A student who explains the work well simply earns the score.
Detection vs defense
What teachers and admins ask first.
Do AI writing detectors work?
AI detectors estimate the probability that a passage was machine-generated, based on statistical features of the text. The accuracy varies, and false positives are a documented concern. Even when the probability is high, it does not establish whether the student understands the work.
What does Tolus measure that an AI detector cannot?
Tolus measures whether the student can explain and defend their own submission in their own voice. It produces evidence (a transcript and citations), not a probability.
Can Tolus be fooled by a coached student?
Defenses are live and the AI generates follow-ups based on the actual submission. A student who has memorized a script struggles when the AI presses on a specific claim or asks for transfer to an unprompted example.
Is Tolus safer than detection from a fairness standpoint?
Detection produces a probability that can be wrong about authorship. Tolus only ever scores what the student explicitly says during the defense. There is no inference about origin, no false-accusation risk: a student who explains the work well receives a high score.
What about students who use AI to learn?
Students who use AI as a study aid and genuinely understand the material defend their work fluently. Tolus is built around the assumption that the use of AI is increasingly inseparable from study; what matters is whether learning happened.
Stop guessing
Swap the probability for a transcript.
Bring a real submission. We will run a defense end to end and show you the report a teacher receives: the student’s own words and a rubric score, in place of a detection guess about authorship.