Confidence, not consistency, characterizes flashbulb memories

Psychological Science 14(5): 455–461, DOI 10.1111/1467-9280.02453. Duke University. © American Psychological Society. Read as the published PDF posted on the first author's publication page (sites.lafayette.edu/talaricj). The Duke repository copy is behind an anti-bot check and was not used. Provenance: papers/carry_on/talarico2003_flashbulb_confidence.provenance.json.

What was read

All 718 lines of pdftotext output in reading order: abstract, method, results, discussion, acknowledgments and references.

• The text layer drops "=", "<" and minus signs, and scrambles Table 1. Pages 457–459 were therefore rendered and read: Figure 1 (three panels), Table 1, Figure 2, and the correlation paragraph, whose signs were read from the image.
• Values marked "(fig)" are plot readings.

Question

Compared with an everyday memory from the same days, are flashbulb memories of 9/11 more consistent over time? Or do they only feel more accurate? What part does emotion play?

Method

• Participants and design. 54 Duke students wrote down, on 12 September 2001, how they heard about the attacks, and an everyday event from up to 3 days before.
  • They were randomly assigned to a single retest after 7, 42 or 224 days (18 per group). Retesting only once avoids rehearsal from repeated testing.
• Recall. Six open questions per memory: who told you, when, where, others present, what you were doing, other distinctive details.
  • Two coders counted details and scored them consistent or inconsistent (directly contradictory, or not the same entity). Inter-coder agreement was 96–97%.
• Ratings (Autobiographical Memory Questionnaire, 1–7):
  • recollection, belief in accuracy, remember/know, vividness;
  • story coherence, fragmentation;
  • valence, intensity, visceral response, same emotion;
  • rehearsal, field versus observer perspective.
• PTSD symptoms. The PCL-S at the retest.
• Analysis. Mixed ANOVAs (memory type × session × group), plus correlations of day-1 measures with delayed measures after removing group means.

Results

• Consistency declines alike for both kinds of memory.

  	Day 1 (fig)	Day 224 (fig)
  Consistent details, flashbulb	≈ 11.5	≈ 7.4
  Consistent details, everyday	≈ 12.5	≈ 8.4
  Inconsistent details, both	0	≈ 2.6–2.7

  • The decline is roughly logarithmic.
  • No memory type × time interaction: F for consistent details .02 to .45; for inconsistent details .24.
  • "Other distinctive details" supplied 42% of the inconsistencies. People listed different details at different times, rather than keeping odd details.

• What differs is how the memory feels. For flashbulb memories, recollection, belief, remember/know and vividness stay high and flat. For everyday memories they fall (memory × session interactions F = 5.13–10.67).

  Rating	Flashbulb, day 1 → 224 (fig)	Everyday, day 1 → 224 (fig)
  Belief	≈ 5.55 → 5.35	≈ 5.4 → 4.45
  Remember/know	≈ 5.7 → 5.55	≈ 5.35 → 3.4
  Vividness	≈ 5.35 → 5.05	≈ 4.95 → 3.6

• Other properties of flashbulb memories:

  • more story-like (4.68 against 3.61) and less fragmented (3.29 against 4.53);
  • more negative (valence −1.89 against 0.61);
  • rehearsed more (5.23 against 2.46);
  • the field (own-eyes) perspective stays for flashbulb memories. Everyday memories drift to an observer view (memory × session F = 29.98; Figure 2).

• Emotion. Intensity and visceral ratings fall over time for both kinds.

  • No day-1 rating, emotion included, predicted later consistency.
  • Later belief in accuracy correlated with day-1 visceral response (r = −.31), fragmentation (r = −.35), remember/know (.45), personal specificity (.33) and rehearsal (.29).
  • Day-1 visceral response predicted PTSD symptoms at retest (r = .48); valence did too (r = −.28).

Limits

• 54 students at one university, one flashbulb event, one everyday event each, and 18 per delay group.
• "Consistency" is agreement with a report written the next day, not accuracy.
• The correlation analysis includes many tests (only significant ones reported) at about 54 people.
• Inconsistencies found:
  • Consistent details differ by memory type. The abstract says consistency for the two kinds "did not differ", but Table 1 shows a significant main effect of memory type (F = 4.26, p < .05). Figure 1 has everyday memories about one detail higher throughout.
    • Only the time course (the interactions) does not differ. The difference favours everyday memories, which supports the paper's point, but the sentence as written is not what the table shows.
  • Visceral response and belief. The abstract and discussion present visceral response as the emotion measure that predicts later belief in accuracy, implying it supports confidence. The printed correlation is negative (r = −.31): more visceral response on day 1 goes with lower later belief. The direction is never discussed. This could be a sign error or an unremarked result; it cannot be told from the paper.
  • The acknowledgments thank instructors for access on "September 12, 2002"; testing was on 12 September 2001.
• Checked:
  • 3 × 18 = 54 participants.
  • Error degrees of freedom 51 = 54 − 3, as the table note says.
  • The delay-correction constants (1.78, 1.93; 2.67, 3.15) match the plotted gaps in consistent details between the 7-day group and the 42- and 224-day groups: ≈ 1.8 and 2.7 (flashbulb); ≈ 1.9 and 3.3 (everyday) (fig).
  • Every F in the text matches Table 1.

What it means for Kurisutina

• Q3: a human-like memory drifts in detail while confidence stays. For an emotional memory, people lose and change details on a logarithmic course, just as for ordinary memories. Their sense of vividness and certainty does not drop. For ordinary memories, both fall and the viewpoint drifts from first person to observer (verified).

  • This dissociation is a signature a replica could show or not show. A retrieval store with verbatim episodes shows neither the drift nor the fading confidence.

• Where drift belongs (proposal, consistent with the memory-architecture summaries). The replica's record of what happened should stay immutable (the content layer).

  • If human-like recall is wanted, the drift belongs in how episodes are retold, in the interpretive layer. It should be logged against the record so it can be measured and never overwrites it.
  • The replica's confidence should follow the person's pattern: high for emotional or central memories, fading for ordinary ones. It should not follow the record's actual fidelity.

• An evaluation recipe, directly reusable (proposal).

  1. Ask the replica about the same episode at different times, fresh session each time.
  2. Score consistent and inconsistent details with this coding scheme, plus its 1–7 ratings of vividness, belief and perspective.
  3. Compare against the human curves here (about −35% consistent details by 32 weeks; flat confidence for emotional memories).

  Peripheral "distinctive details" are where people drift most, so they are the sensitive probe.

• Emotion is not what keeps details. No day-1 emotion rating predicted consistency. Emotion goes with confidence, rehearsal and later symptoms, not accuracy. A replica that treats emotional salience as a reason to store details more faithfully would be less human, not more (inferred).

Cross-references

• summaries/carry_on/hirst2009_911_memory.md: long-term 9/11 flashbulb and event memory, with consistency and confidence over years.
• summaries/carry_on/coman2009_forgetting_conversation.md: conversation-induced inaccessibility of 9/11 memories.
• summaries/carry_on/diamond2020_real_world_recall.md: accuracy and amount of real-world recall.
• summaries/carry_on/wu2024_longmemeval.md: knowledge updates and abstention in LLM memory.