AI wasn't in the room

The great misdirection

Here's what most assessment platforms will tell you: We use sophisticated AI detection to catch cheating.

Here's what they won't tell you: AI detection is a fool's errand. The tools are unreliable, easily circumvented, and create an adversarial relationship with students that poisons the learning environment. You're essentially saying, "I don't trust you, so I'm going to watch your every move and run your answers through a black-box algorithm that may or may not be right."

There's a better way. But it requires shifting the entire frame.

Conventional wisdom: Build better detection tools to catch AI-generated answers.

The counter-argument: Design questions that AI fundamentally cannot answer.

Why your multiple-choice questions are an AI buffet

Let's be honest about what most multiple-choice questions actually test.

Question: "What year did World War II end?"

Options: A) 1943 B) 1944 C) 1945 D) 1946

This isn't an assessment of understanding. It's a retrieval task. And retrieval is exactly what Large Language Models do best. These systems are probabilistic engines trained on the aggregate of human knowledge. Ask them for a definition, a date, a formula, or a standard fact, and they'll deliver it in milliseconds with near-perfect accuracy.

If your exam only tests retrieval, your exam is already compromised.

The problem isn't that AI is "too smart." The problem is that traditional assessment questions are too predictable. They tap into the exact patterns these models were trained to recognize and reproduce.

Context is Kryptonite

Here's the technical reality: LLMs are brilliant at synthesizing the average of human knowledge. They're spectacularly bad at the hyper-specific, the deeply contextual, and the contradictory.

They fail when the answer isn't in their training data. They stumble when the question requires real-time, localized, human-contextual grounding. They choke when you ask them to do something that demands actual cognitive friction.

This is your opening.

How to write an "un-AI" question

The strategy is simple: inject context that the AI cannot access.

Example 1: The Novel Dataset

Instead of: "Explain the principles of supply and demand."

Try: "Here's a dataset from the campus coffee shop showing sales figures from last week, including a sudden price increase on Wednesday. Using supply-demand theory, analyze why Thursday's sales dropped 40% despite the price change being announced Tuesday."

The AI has never seen this dataset. It wasn't in the training set. The student has to apply theory to fresh, localized information.

Example 2: The Contradictory Sources

Instead of: "Summarize the key arguments about climate policy."

Try: "Read these two abstracts from Tuesday's seminar, one argues for carbon taxation, the other for cap-and-trade. Based on our class discussion about implementation challenges in developing economies, which approach is more viable and why?"

The AI doesn't know what happened in Tuesday's seminar. It doesn't have access to the specific arguments your class debated. This question requires participation, not just comprehension.

Example 3: The Ultra-Recent Case

Instead of: "Describe the causes of the 2008 financial crisis."

Try: "A new banking regulation was announced last week [link to actual article]. Apply the framework we studied in Module 3 to predict two potential unintended consequences of this policy."

Training data has a cutoff. Real-time events don't. Questions anchored to recent developments force students to apply frameworks, not regurgitate facts.

The locked-down reality

Here's where we get real: great question design is only half the battle.

You can write the most contextually brilliant, AI-resistant question in the world. But if a student can copy-paste it into a chatbot mid-exam, you've lost. The question might be unanswerable by AI on its own, but the moment you give the model access to the prompt, it can attempt a response. Maybe it won't be perfect, but it'll be enough to blur the line between authentic and assisted work.

This is why environment matters.

You need two things working together:

• Questions that defeat AI's knowledge (through context, specificity, and cognitive demand)
• A delivery environment that defeats AI's access (through technical controls that block unauthorized tools)

One without the other is incomplete. Context-rich questions in an open browser are vulnerable. Locked-down browsers with generic questions are pointless.

How Online Exams makes this possible

This is where the platform's architecture becomes essential. Online Exams was built around a simple premise: assessment should be AI-resistant by design.

That means:

• Focus locking. Leave the exam screen, even once, and the attempt locks. No browser can stop a student opening another tab, so what this does instead is notice at once and end the attempt.
• Server-side grading. The answer key never reaches the device, so there is nothing in the page to read.
• Copy blocking on the question text. A deterrent rather than a wall: the words are on screen, so they can still be retyped or photographed.
• One attempt per email address, so an exam cannot simply be reloaded for a second go.

We don't claim to have a magic AI-detection algorithm. We don't pretend to read students' minds or scan their rooms with webcams. What we do is simpler and more reliable: we give you the vault.

You provide the un-crackable combination, the context-rich, human-grounded questions that demand actual understanding. We provide the conditions they are answered under.

Together, they create an assessment that measures learning, not workarounds.

The shift that matters

Moving to "Un-AI" exam design isn't about making questions arbitrarily harder. It's about making them smarter.

It's the difference between:

• Asking students to define a concept → and asking them to apply it to a scenario they've never seen
• Testing whether they can recall information → and testing whether they can synthesize it with new data
• Checking if they remember what you taught → and checking if they can use it in a context that matters

This approach does something else, too: it restores trust. When students know the exam is designed to measure their actual understanding, when the questions are clearly rooted in course context and recent discussions, the adversarial dynamic fades. You're not trying to catch them. You're trying to see what they've learned.

Where this leaves the question writer

You cannot out-surveil a probabilistic text generator. The technology will always stay one step ahead of detection tools.

But you can out-design it.

Write questions that demand human context. Anchor assessments in localized, recent, contradictory, or novel information that exists outside the training set. Then deliver those questions in an environment that prevents students from outsourcing the cognitive work.

That's the Un-AI exam. Not a detection game. Not an arms race. Just solid assessment design, backed by secure delivery.

The AI can have the facts. You keep the understanding.