Why governance before adoption
The cost of skipping readiness, and why literacy comes before policy.
Try one before your team takes twenty.
AI policies fail through what people believe: that fluent output is accurate, that a tool can check itself, or that the safe move is never to use it. The Aegis literacy check names those beliefs, one scenario at a time.
This practice scenario is not one of the twenty and is not scored.
AEGIS LITERACY
CHECK · LIVE · 20 SCENARIOS · 5 COMPETENCIES
35 MISCONCEPTIONS · 6 FAMILIES
CURRICULUM · 8 MODULES · DRAFTS IN REVIEW
LITERACY THRESHOLD · HYPOTHESIS, NOT A FINDING
A score tells you how many. This tells you which.
Every wrong option in the check was written to catch one named belief. A team that scores 14 of 20 by over-trusting needs different teaching from a team that scores 14 by refusing everything.
Believing an output has been checked when it has not.
- Specific details, numbers and section citations, mean the output is real.
- If the AI gives the same answer twice, the answer is correct.
- Labeling output as AI-generated replaces checking it.
- If most of the output is right, the rest probably is too.
Misjudging what sharing data with a tool exposes.
- Removing names leaves no privacy risk.
- Deleting a chat afterwards undoes sharing the data.
- Public AI tools do not keep what users type.
- Without names or addresses, data cannot produce unfair treatment.
Accepting weak evidence that a tool works.
- Vendor-published figures are good evidence for our setting.
- A good overall number means the tool works well for everyone.
- The absence of complaints shows a tool works.
- Heavy use of a tool proves it helps.
Skipping the human step where one is needed.
- A human making the final call makes everything upstream low-risk.
- If only a small share of people are affected, the use is low-risk.
- Deploy first and assess the risk afterwards.
Treating caution as the same thing as care.
- Every AI output needs the same heavy review, regardless of stakes.
- When in doubt, stop using AI or discard everything it produced.
- Demanding perfection instead of checking fitness for the actual use.
Assuming an approval covers more than it does.
- Once a tool or vendor is approved, every new use is covered.
- A large, established vendor can be trusted by default.
- Review depends on who built the tool, not on what it decides.
The check is open now.
Twenty workplace scenarios, written for non-technical staff in public-sector and enterprise settings. Every scenario, organization, and number in them is invented for the assessment.
- No account, and nothing to install.
- Responses are not stored while the item bank is in review. Collection opens only with consent, in a pilot.
- The answer key stays on the server, so the page cannot leak it.
- LIT-WF-01Recognize when an AI output should be checked by a human before it is acted on.
- LIT-WF-02Identify a plausible-sounding but incorrect AI answer.
- LIT-DATA-01Explain why feeding sensitive personal data into an AI tool can create risk.
- LIT-GOV-01Decide whether a proposed use touches a rights-impacting decision that needs escalation.
- LIT-MEAS-01Recognize weak evidence, including an overall number that hides unequal results across groups.
At the end: results by competency, the misconception families your wrong answers point to, and a review of every scenario.
Eight modules, taught to the profile.
The curriculum is the readiness scan turned inside out: each dimension the scan measures becomes a module that teaches the literacy it assumes. It is built for the people who approve AI, not the people who build it.
- BeforeThe cohort takes the check. The profile shows which misconception families the group holds.
- Teach to the profileEach module maps those families to the sections that address them. Spend time where the misconceptions are, not evenly.
- AfterRe-assess. There is one form today, so some gain on a re-test is practice effect; a parallel form comes before anyone reports gains.
Knowing your data
No names is not anonymous. What sharing data with a tool actually exposes.
Mapping the work
What makes a workflow safe to automate. No check items assess this module yet.
People & judgment
Fluency is not accuracy. When a human has to be in the loop, and how to spot a confident wrong answer. The module the rest of the program leans on.
Build, buy, or wait
Reading vendor claims critically. Wait is a legitimate answer.
Rights, risk & oversight
Which uses touch people’s rights, and why an approval covers a use, not a tool.
Measuring what matters
Success metrics that mean something. Ask for the breakdown before trusting the average.
Mission & access
Pointing AI effort at real gaps. Accessibility is part of quality. No check items assess this module yet.
All eight modules are AI-assisted drafts. None is used with learners until a credentialed reviewer has checked its citations and released it. Practice scenarios in the modules are never the scored items, so teaching does not leak answers.
What is ready, and what is not.
Built and open
- Twenty scenarios, every wrong option tagged to a named misconception
- The check page, with server-side scoring
- Results by competency and misconception family
In review
- The scenarios’ wording and answer key, before any pilot collects responses
- All eight curriculum modules
Not claimed
- A literacy threshold for governance readiness: a hypothesis, reported only after a pilot of at least 30 people
- Pre/post learning gains, until a second form exists
Plan a literacy cohort.
Run the check with your team before an AI rollout, read the profile together, and teach the modules it calls for. Pilots with consented collection help validate the instrument itself.