R0053/2026-03-31-02/C002/H1¶
Statement¶
The claim is accurate — negative constraints ("must not") are necessary for AI compliance and positive instructions alone are insufficient.
Status¶
Current: Eliminated
Supporting Evidence¶
| Evidence | Summary |
|---|---|
| SRC02-E01 | Instruction hierarchies do fail, suggesting enforcement is needed |
Contradicting Evidence¶
| Evidence | Summary |
|---|---|
| SRC01-E01 | Negative instructions produce worse output |
| SRC01-E02 | Anthropic recommends positive over negative framing |
Reasoning¶
The evidence directly contradicts the claim's specific mechanism. While enforcement is needed (partial support), the claim's assertion that you "must tell the AI what it is not allowed to do" is shown to be counterproductive in many cases.
Relationship to Other Hypotheses¶
H1 and H2 agree that enforcement is needed but disagree on mechanism. H1 and H3 are fully contradictory.