Venture capital in education technology is concentrating in two product categories: AI tutors and anti-cheating tools. Both bets share an assumption, that instruction is the bottleneck and enforcement is the cure. The evidence points the other way. The cheating crisis is an assessment design problem, and the industry is paying to defend a test format that no longer measures what it claims to measure.
Detection is failing on its own terms
The enforcement play looks like this: browser lockdowns, webcam proctoring, keystroke analysis, and AI-writing detectors. The track record is poor. OpenAI shut down its own AI text classifier in 2023, citing low accuracy. Independent testing of commercial detectors has repeatedly shown false positive rates high enough to flag honest students, with non-native English writers disproportionately accused. Simple paraphrasing tools degrade detector performance further.
The economics favor evasion, though not for the reason usually given. A detector only needs one confident catch to trigger discipline. The problem is that confidence is exactly what these tools lack. Every accusation an institution acts on carries legal and reputational risk, so an unreliable detector is unusable at precisely the moments it matters most. Meanwhile, evasion improves as a free byproduct of every new model generation, while detectors chase a moving target with error bars too wide for high-stakes decisions. That disadvantage is structural, and it is permanent.
What the tests were measuring
Frontier models have passed bar exams, medical licensing exams, and AP tests. That fact should change how we read a standardized test score.
If a general-purpose model can produce a passing answer, then a passing answer no longer demonstrates that a human internalized the material. It demonstrates that a passing answer was producible. Those were the same thing for a century. They are not the same thing anymore.
Recall still matters. Foundational knowledge is often a prerequisite for higher-order reasoning, and a student who has internalized nothing has little to reason with. The point is narrower: a proctored recall test used to be decent evidence that the knowledge lived in the student's head, and it no longer is. Access to a correct answer became separable from possession of the underlying knowledge, so instruments that only sample answers lost most of their evidentiary value. We still need to verify internalization. We need different instruments to do it.
The inversion
Stop asking how to keep AI out of the exam room. Ask what an exam should measure when producing a correct answer is cheap.
The cheating framing assumes the test is fine and the students are the problem. Invert it: students are behaving rationally inside a broken measurement system. When the assessed skill and the valuable skill diverge, gaming the assessment is the predictable outcome.
What defensible assessment looks like
We already have working models for assessing skill in environments where reference material is assumed to be available:
- Medical education uses OSCEs, structured clinical encounters where students respond to a live patient scenario. You cannot paste a chatbot answer into a conversation with a standardized patient.
- Aviation uses check rides. The examiner introduces failures mid-flight and grades the response. The pilot has full access to checklists, which is the point: the test measures application under changing conditions, not recall.
- Strong engineering interviews grade code review and design discussion rather than memorized puzzle solutions, precisely because the puzzles leaked years ago.
Translated to general education, the pattern is multi-stage formative assessment:
- Grade the delta between a draft and a revision after feedback, instead of only the final artifact.
- Introduce a novel constraint mid-task and evaluate how the student adapts.
- Use short oral defenses where students explain and extend their own submitted work.
- Treat process artifacts, like version history and documented reasoning, as first-class evidence.
None of these are cheat-proof, and for high-stakes credentialing that caveat is not trivial. Institutions answer to accreditors, courts, and the public, and systematic cheating at scale is a genuine liability even in interactive formats. But these formats shift the measured quantity from output, which AI produces cheaply, to iteration and reasoning under live interaction, which is far harder to delegate and much closer to what employers actually need.
The honest objection: cost
This is the strongest counterargument, and it deserves a straight answer. Every format above is labor-intensive. OSCEs require trained standardized patients. Check rides require an examiner in the seat. Oral defenses consume faculty hours that a Scantron never did. The standardized test won historically precisely because it was cheap at scale, and that constraint has not disappeared.
Two things can reduce the cost without eliminating it. First, not all assessment is high-stakes. Frequent, low-stakes formative checks carry most of the pedagogical load, and AI assistance for scenario generation and first-pass feedback looks economically plausible there today. Second, high-stakes assessment can concentrate scarce human attention where it matters, using process evidence gathered cheaply along the way to decide who needs deeper live evaluation. Whether AI-administered assessment can ever meet the reliability bar for credentialing decisions is unproven, and nothing in the structural argument depends on it. Even if interactive assessment stays expensive, an expensive instrument that measures something real beats a cheap one that measures nothing.
Where the money should go
The assessment market, estimated in the billions, is currently funding the defense of an obsolete instrument. The larger opportunity is a better question: instruments that measure how a student iterates, reasons, and applies concepts when yesterday's answers are free. That work is harder than shipping another proctoring plugin. It is also the only version of this market with a future.