Week 6 Session: Design a test that isn't fooled by the resource
Sort three fictional resource cases against Clark and Chalmers' four conditions, then design a content-change test for a real resource you use.
29 March 2027
Where Week 5 left us#
Week 5 built a seven-rung ladder, but Recall and Reconstruction assume that removing a source makes the evidence stronger. This week tests that assumption by varying permitted resources. We move from the lock to a fictional courier; the examination question continues even though the setting changes.
Purpose and preparation#
This self-guided activity applies Week 6’s lesson. Allow 20 minutes. All required information is below. Work alone or with a partner; no navigation app, live delivery, paid tool or AI account is required. The courier case is fictional. You are examining what a resource-supported performance is evidence of, not running a technical certification.
If you already work with one of the resources described below professionally, record that familiarity rather than pretending the material is new.
1. Recap the four conditions — 2 minutes#
Clark and Chalmers (1998), §4, give four conditions under which an external resource functions like part of someone’s own memory rather than an outside reference: it is a constant presence consulted whenever relevant; the information is available without difficulty; it is automatically endorsed on retrieval; and it is there because of the person’s own past endorsement — they put it there. They treat the fourth as more arguable than the first three.
2. Sort three cases — 5 minutes#
For each case, decide whether the resource looks closer to an integrated part of the person’s own competence, or closer to a stranger’s advice consulted for the first time. Write one sentence of reasoning per case, naming which condition(s) it passes or fails.
- A hiker checks a trail map app for the first time, at a junction, on a trail they have never walked before.
- A long-haul truck driver checks a route-planning app used on every shift for years, trusted without re-verifying, configured with her own preferred settings.
- A student pastes a homework question into an AI tool opened minutes earlier and copies out whatever it returns without reading it.
Compare after sorting
The hiker’s map fails constancy and past endorsement: first use, not assembled from prior choices. The truck driver’s app meets the strong conditions: constant, effortless, trusted from repeated correct use, self-configured. The student’s AI use fails constancy and past endorsement in the same way as the hiker, and additionally fails automatic endorsement in the way that matters — the output was never checked before being copied, so what happened was not trust earned by reliability but an absence of checking at all.
3. Design a content-change test — 8 minutes#
Pick a resource you actually use for something familiar: a recipe app, a calculator, a reference chart, a translation tool, a spellchecker, a GPS.
Write:
- The task, and why the resource ordinarily meets most of the four conditions for you.
- One change to the resource’s content — not whether it is present — that would create a conflict between what the resource says and what a person actually engaging with the task would notice.
- What a correct response to your changed version would demonstrate, in one sentence.
- What it would still not demonstrate, in one sentence.
Preserve your original design before reading the comparison.
Compare your design with an example
Resource: a spellchecker used while writing in your first language, checked constantly, trusted, self-configured — meets the four conditions well. Content change: insert a correctly spelled word used in the wrong sense (e.g. “there” for “their”) that a spellchecker will not flag. A person editing for meaning, not just accepting the checker’s silence, catches it; a person who treats “no red underline” as “no error” does not. This demonstrates whether the writer is reading for sense rather than trusting the tool’s silence as proof of correctness. It does not demonstrate that they could write the passage from nothing, with no draft to edit at all — a different task would be needed for that.
4. Compare a presence test against a content test — 3 minutes#
Write two short task descriptions for the same skill: one that tests by removing the resource entirely, and one that tests by changing the resource’s content while keeping it present. Name one thing each version risks mismeasuring.
For example: removing a translation app entirely from someone learning a new language risks testing memorised vocabulary from a single recent lesson, which not everyone would have had reason to memorise yet. Leaving the app present but feeding it a sentence with a false friend or an idiom it mistranslates risks nothing comparable — a correct catch shows engagement with meaning regardless of how new the vocabulary is.
5. Keep a short record — 2 minutes#
| Field | What to write |
|---|---|
| Resource | What it is, and how habitual is it for you or your example person? |
| Fit against the four conditions | Which are met, which are not? |
| Version used | Presence/absence, or content change? |
| What it supports | The specific claim the response backs up. |
| Limit | What further claim would still need a different task. |
Working alone, you cannot independently verify your own account of how you produced an answer. Working with a partner adds observation, not access to their thoughts. Neither arrangement makes the conclusion certain.
For the underlying argument, read Clark and Chalmers (1998), “The Extended Mind,” §§4–5. Ask which condition is doing the most work in your own sorted cases.
The spec
What you bring to the session. Written so a reader can tell whether it has been met without asking you.
- sort three cases against the four conditions and justify each
- design a content-change test for a resource you actually use
- compare a presence/absence version of a task against a content version