Citizens for Safe AI

Work Test

Enter the access code from your invitation email to begin or resume.

Access code

That code isn't right. Check the email we sent you.

Private assessment. Do not share this page or your code.

Citizens for Safe AI

Before you start

Thank you for making the time for this! Before you begin, please read these instructions carefully. If you have any questions, please email (with “Work Test Help” in the subject) before you start Part 1. If something breaks, email Antony with “Work Test Help” in the subject, then carry on as best you can.

You have access to an AI assistant throughout the assessment.
The assessment has several parts, each separately timed. Once you start a part, its clock runs until it ends — it does not pause, even if you close this tab. When a part's time is up, it locks with whatever you've written.
Breaks between parts are untimed. Finish a part, stand up, come back when you're ready, and start the next one.
Your work saves automatically to our server every few seconds. You can refresh or switch devices; your code brings you back to exactly where you were.
Don’t put your name anywhere in your answers or in the assistant chat. Submissions are graded blind — the review panel reads your work before seeing who wrote it. Where a draft email needs a sign-off, use a placeholder like “[Your name]”.
When you finish the last part, your answers and assistant transcripts are already saved for the review team — nothing to download or email.

Use the provided assistant — and only the provided assistant. The right-hand panel connects you to Claude. Use it however you like: ask about the scenario, brainstorm, pressure-test a draft. Note that we have intentionally given you access to Sonnet 4.6, which is far from the smartest model on the market. That means that you should not assume the AI is smarter than you, and should use it carefully. The full transcript is recorded and read by the review panel alongside your answers — how you use the assistant is part of what we're evaluating, since we expect employees will use LLMs in their job (we certainly do!). A transcript showing heavy, skillful use of the assistant is a positive signal, but copying the LLM's writing is a negative signal. Note that AI tools other than the assistant are not permitted.

Internet research is allowed, and you may look things up in other tabs freely. But nothing can be pasted into the answer box. All answers must be typed, in your own words. Paste attempts and tab switches are recorded with your submission. Final submissions are also run through Pangram, AI-detection software that is quite good, and if it flags your responses as AI-generated, you'll be disqualified. Responses will also be read by a human who works with LLMs every day. Basically — please don't try to cheat.

Citizens for Safe AI

Part complete

You're on a break — nothing is being timed. The next part's clock starts only when you press the button.

Citizens for Safe AI

Before you begin

The clock has not started. It starts the moment you press Start.

Citizens for Safe AI

Session complete

Every part is locked. Your answers and assistant transcripts are saved for the review team — there is nothing more you need to do. Thank you; we read every submission closely.