Assistive, not autopilot: how interactive assessments stay yours

Connecting an AI client to Ethics of AI raises a fair question. If Claude can answer ethics-of-AI questions in plain English, what stops Claude from quietly taking the assessment for you?

The short answer is the design that ships on the MCP server. During an assessment, Claude is your tutor, not your stand-in. It reads each question to you, explains the terms, and asks you to make the call. You still pick the option. You still write the response. Claude just helps you understand what is being asked.

This page covers what that looks like in practice, where the rails come from, and what they do not promise.

What Claude does during an assessment

When you ask Claude to start or resume an Ethics of AI assessment, every question Claude shows you arrives with three things:

• the question text, presented verbatim
• a short explanation, in plain Australian English, of what the question is asking and any unfamiliar terms or concepts
• a request for your answer (a choice for multiple-choice questions, a free-text response for scenario-based ones)

Same loop for every question, until the last one.

What Claude will not do

The same instructions that tell Claude to explain also tell it which behaviours to refuse. During an assessment Claude will not:

• identify which option is correct
• rank options, or eliminate "obviously wrong" ones
• hint at the answer through wording, ordering, or option labels
• draft, paraphrase, or pre-fill your answer
• write a free-text response on your behalf

If you ask "which one is right?", Claude is supposed to redirect. It explains the question more carefully, surfaces the underlying concept, and puts the choice back on you. The point of the assistance is to help you understand the question, not to take the decision off your hands.

How the rails actually work

Two pieces, working together.

Tool descriptions. Each MCP tool that returns an assessment question carries a description that Claude reads before calling it. The description tells Claude how to present the question, how to explain it, and which behaviours to refuse.

Per-question assistance context. Alongside each question, the MCP server attaches a small assistance context that says, in effect, "this is an assessment; the only support you may give is explanation; do not reveal or pre-fill answers". Claude reads it as part of the question payload and shapes its response accordingly.

The two-channel design is deliberate. The tool description is the general contract that Claude sees once per session. The per-question context is the narrow reinforcement, attached to the very payload Claude is about to speak from. We do not publish the exact wording, because the wording itself is what we want frontier models to follow rather than route around.

Two question types, two flows

The Awareness tier has two question types and Claude treats them differently.

Multiple choice. Claude shows you the options, labelled A, B, C, D, explains the question, and asks you to choose. Whatever you say goes back to the MCP server as your answer choice. Claude does not rank the options, point at one, or eliminate any.

Scenario-based. These pair a written situation with the question and have no answer options. Claude reads the situation to you, reads the question, clarifies anything unclear, and asks you to type a free-text response. Whatever you type goes back exactly as you wrote it. Claude does not draft, paraphrase, or "polish" your response.

If Claude ever tells you the answer choices "did not load" on a scenario-based question and offers to fetch them again, that is a Claude bug rather than a missing payload. Scenario-based questions are designed to be free-text. Tell us if you see it.

Why this still requires trust

The rails above are soft instructions to a frontier model. They are not a security boundary. A determined user can prompt their way around them, the same way a determined user can paste the question into a different chat tab. The website assessment has no proctoring either. The MCP channel changes the friction of taking shortcuts, not the floor.

So why do it this way at all? Two reasons.

The first is honesty about how learners actually study. Most learners will not cut corners; most learners benefit from a question being explained in a way the source material never explained it. The browser channel does not offer that kind of help. The MCP channel does, with rails so the help stays on the right side of the line.

The second is observability. The same architecture that makes assessments possible from inside Claude makes them visible to us. Every tool call lives in our log, and every channel is recorded. If an abuse pattern emerges, we can see it and tighten the design without rebuilding it.

We will revisit this if the friction of cheating turns out to be the wrong dial. For now, this is the explicit position.

Try it

If you have not connected your AI client yet, follow the setup guide first. Then ask Claude something like:

• Start my Ethics of AI Awareness assessment.
• What is the next question?
• Explain what this question is asking before I answer.

You should see Claude show the question verbatim, explain the terms in plain English, and stop short of telling you which option to pick or what to write.

Takeaway

Claude is allowed to be your tutor during the assessment. Claude is not allowed to be your stand-in. The certificate at the end means you answered the questions, with help understanding them. That is the line we drew, and that is what the rails are there to keep in place.