Coverage for narrative_harm_classifier/classifier/counter_narrative.py: 100%

1""" 

2classifier/counter_narrative.py — General counter-narrative guidance per harm mechanism. 

3 

4Deliberately general, templated guidance grounded in the public 

5"acknowledge -> redirect -> inform" counter-messaging framework (the 

6methodology behind projects like Moonshot CVE and the Redirect Method), not 

7auto-generated bespoke rebuttal text for the specific input. Generating 

8custom rebuttal text is a materially harder, riskier text-generation problem 

9(prone to tone-deaf or factually wrong output) and is out of scope here — 

10this gives a human moderator/responder a starting frame, not a script to 

11paste verbatim. 

12""" 

13 

14from typing import Optional 

15 

16_GUIDANCE: dict[str, str] = { 

17 "animalization": ( 

18 "Dehumanizing comparisons (to animals, vermin, disease) are a documented precursor to " 

19 "real-world violence against targeted groups. Effective counter-messaging typically: " 

20 "(1) avoids repeating or amplifying the dehumanizing frame itself, (2) affirms the " 

21 "target group's humanity with specific, concrete facts rather than abstract appeals, " 

22 "(3) redirects toward the underlying grievance being exploited, if any." 

23 ), 

24 "demonization": ( 

25 "Framing a group as evil or supernaturally malevolent shuts down empathy and factual " 

26 "engagement. Counter-messaging typically works better by naming the rhetorical move " 

27 "explicitly (\"this frames an entire group as inherently evil\") and redirecting to " 

28 "specific, falsifiable claims rather than arguing the moral framing directly." 

29 ), 

30 "objectification": ( 

31 "Reducing a group to objects or property strips away agency and interiority. " 

32 "Counter-messaging typically centers first-person voices from the affected group and " 

33 "concrete examples of agency, rather than abstract arguments against the framing." 

34 ), 

35 "criminalization": ( 

36 "Blanket criminal framing of a group is usually contradicted by the actual data. " 

37 "Effective counter-messaging leads with accurate, sourced statistics and specific " 

38 "counterexamples rather than general appeals not to stereotype." 

39 ), 

40 "direct_call_to_violence": ( 

41 "Explicit calls to violence should typically be escalated for human review and, where " 

42 "applicable, reported to platform trust & safety or law enforcement channels rather than " 

43 "countered by automated messaging alone. If counter-messaging is used at all, prioritize " 

44 "de-escalation and directing at-risk viewers toward support resources over debating the " 

45 "claim itself." 

46 ), 

47 "false_attribution": ( 

48 "Claims about a group's supposed hidden agenda are typically unfalsifiable by design. " 

49 "Counter-messaging tends to work better by asking what specific, checkable evidence is " 

50 "being offered (usually none) rather than trying to disprove an unfalsifiable claim " 

51 "directly." 

52 ), 

53} 

54 

55_DEFAULT_GUIDANCE = ( 

56 "Consider whether a direct rebuttal, a redirect to factual information, or escalation to " 

57 "human review is the most appropriate response before responding automatically." 

58) 

59 

60 

61def guidance_for(harm_mechanism: Optional[str]) -> Optional[str]: 

62 """Return general counter-narrative guidance for a harm mechanism, or None if harmless.""" 

63 if not harm_mechanism: 

64 return None 

65 return _GUIDANCE.get(harm_mechanism, _DEFAULT_GUIDANCE)