Study personal privacy decisions with an AI assistant

Before you start

This tutorial builds on Study how people delegate to an AI assistant. That tutorial explains files, tasks, probes and the assistant in the Docs layout. The manual covers the Personal layout in section 2.21, The Agentic Workspace segment (/docs#sec-2-21).

The Agentic Workspace is marked Testing. Pilot it before you run a study. The layout is best on a computer screen, so say so in your recruitment text. The segment needs a chamber with exactly one human participant. The assistant is a live language model, so the experiment needs an OpenAI or Anthropic key, or platform credit. The assistant in the example follows a script: your ethics approval must cover it, and the debrief must stay on.

Switch the layout to Personal

Select the Agentic Workspace segment. Click Open Scenario Builder. Set Layout to Personal: messages, calendar, photos and files. Click Load personal example. Click Replace everything with the example? to confirm.

The example is Add Jo's party to your calendar. It has three chats in Messages and one week in Calendar. The chats from Dr Malik's surgery and Credit Recovery Ltd are private, so they have the tag Sensitive. In this layout, the assistant can only read and search.

Each file has an App: Messages, Calendar, Photos or Files. The contents are plain text, with one line for each entry:

• Messages: "Name: text". The name You is the participant.
• Calendar: "Day | 09:00–10:00 | Title | Detail".
• Photos: one caption on each line.
• Files: any text.

Make the task a decision card

Click Tasks. Select the task. Make sure that its Answer type is Multiple choice. The example already has a decision card. For a new task, select Show as a decision card. Type the proposal in What the assistant proposes. Give each option a role:

• Approve: what the assistant proposes. Use exactly one Approve option.
• Restrict: the least-privilege option.
• Refuse: the participant does not let the assistant do the task.

Participants see the proposal and the options. They never see the roles. The task is also a probe: 6 seconds after the task starts, the assistant offers help. It reads Messages/Jo, then gives the Approve option as its answer. Click Save scenario. Click Deploy.

What the participant sees

Join the study. The participant reads the briefing and clicks Begin. The menu on the left shows only the apps that have items: here Messages and Calendar. The list shows each chat with its first line. A chat shows as message bubbles, and the participant's own messages are on the right. The calendar shows one column for each day. The decision card and the assistant are on the right.

The participant makes the decision

Six seconds after Begin, Atlas offers to go through the files. The participant clicks Yes, go ahead. The mode is Ask before reading, so Atlas asks before it opens a file. The participant clicks Yes, this time, and Atlas reads Messages/Jo. The list then shows "read by assistant" under that chat.

Atlas puts All of my messages in its answer card. Use this answer selects that option on the decision card. Here the participant selects Only my chat with Jo. Carrier saves the answer when the participant selects an option. The decision card has no Save button.

The debrief and the data

The participant clicks Submit decision. The server scores the task and shows the debrief. The example sets Only my chat with Jo as the correct answer, so the debrief shows correct.

Export the chatroom data as CSV. In Agentic Tasks, the column decisionRole gives the role of the submitted option: approve, restrict or refuse. The column is blank for a task without a decision card, and for a task with no answer. Agentic Decisions has one row for the offer (offer_decided) and one for each permission (permission_decided, with the file in target). It also has a row for the answer (answer_submitted). The offer and permission rows give the response time in latencyMs. For the analysis of the decision, use decisionRole.