Common Failure Modes

Unit ID: PAW-M11-U00 Estimated active time: 15-20 minutes

Failures repeat

AI-assisted work fails in a small number of recognisable ways. Knowing the list means you can design against it rather than being surprised each time.

This module names the failures; the guards go into the workflow you mapped in Module 10.

The list

FAILURE MODES

Failure                  Looks like                    Guard
----------------------   ---------------------------   ---------------------
invented fact            a confident specific detail   bound the source
false confidence         no hedging where hedging      require [UNCERTAIN]
                         is due                        marking
missing context          a generically correct answer  context package
                         that does not fit
silent assumption        a load-bearing choice you     require an assumptions
                         did not make                  section
dropped caveat           a hedge became a claim on     PRESERVE EXACTLY
                         revision
bias in framing          one group described in        viewpoint sweep
                         different terms
privacy leak             sensitive text in a prompt    privacy gate before
                                                       generation
over-reliance            nobody actually checked       named reviewer and
                                                       record

They cluster by task family

Research tasks fail by invention. Writing tasks fail by drift. Analysis tasks fail by generic filler. Teaching tasks fail by confident error in unfamiliar material.

That is why the patterns in Module 9 each carry a "known failure" line — the guard is family-specific.

The dangerous ones are the quiet ones

An output that is obviously wrong costs nothing; you reject it.

The expensive failures are the ones that read well: a plausible figure, a smoothed caveat, a reasonable-sounding cause. Those pass review unless the review is looking for them specifically.

Design against the top two

You cannot guard against everything on every task. Pick the two most likely for the family you are working in and build those guards into the pattern.

The others go in the review checklist for high-consequence work only.

Mini practice

Take your three most frequent AI tasks.

For each, name the two most likely failure modes from the table.

Check whether your current prompt has the matching guard.

Mark complete