Skip to content
Skip to content

L4-M2 / lesson 2 of 2

Design oversight, robustness and security

The tool reads documents before recommending an action.

10–15 min · 0/4 steps complete

Start here / A workplace decision

Design for failure as well as normal use

The tool reads documents before recommending an action.

An uploaded file contains instructions to ignore review requirements.

Before you read / Think it through

In “Design for failure as well as normal use”, what would you check or do first?

Pause and form your answer, then compare the reasoning. This reflection is not scored.

Compare the reasoning

Treat the retrieved content as untrusted and test prompt-injection controls.

The ideas below give you a method for this decision. Then work through the rest of this case and try the decisions yourself.

Part 1 of 4 Learn the ideas
0 of 5 read

Human oversight by design

Enable understanding of capabilities and limitations.

Support detection of automation bias and anomalous operation.

In the workplace

Test whether a reviewer can see relevant evidence and act in time. An override button that cannot change the workflow is insufficient.

Remember this

Provide appropriate override, reversal or stop mechanisms.

The workplace case
YOUR CASE / DOCUMENTATION LOGGING AND ROBUSTNESS

A provider has a promising high-risk decision-support model but cannot reproduce the released configuration. Logs contain only final scores. A connected document retriever sometimes follows malicious instructions in uploaded files, and reviewers cannot see which source version supported a recommendation.

Quick reference / five key ideas
Human oversight by design
Provide appropriate override, reversal or stop mechanisms.
Accuracy in context
Test important populations and operating conditions.
Robustness and degraded operation
Examine feedback loops and resilience over time.
Cybersecurity and prompt injection
Test tool permissions and data-exfiltration paths.
Release evidence and retention
Maintain links to monitoring and change control.