← Back to The IndexFinal ReviewGroup 5
Final Review

Review Nº 05

Causal-Intersectional Identity Harms & the Media Erasure Index
AuthorsAdd author names
28/30
Score
A theoretically grounded, methodologically careful and refreshingly honest causal-intersectional pipeline with a novel Media Erasure Index — but its quantitative headline claims rest on very small counts (key pairs n≤4) and a few internal numerical inconsistencies need reconciling.
AI Risks · Final Review

The Pros

+
Strong theoretical grounding (Crenshaw intersectionality, Collins' Matrix of Domination) tied directly to operational metrics rather than invoked decoratively.
+
The double-counterfactual causal gate (CQ1+CQ2) applied at the report level is a rigorous way to separate causally implicated identity from incidental demographic context, and it is validated against human coders (κ=0.81).
+
The Media Erasure Index is a genuinely novel contribution, and the class-invisibility finding (Lower Class: 104 inferred vs 10 explicit) is striking and well-argued.
+
Intellectually honest reporting: the paper presents not only amplified pairs but also the below-baseline (suppressed) pairs, and repeatedly stresses that observed co-occurrence is confounded by real-world demographic base rates.
+
Limitations are unusually candid about small N, instability of n<3 pairs, infeasibility of three-way intersections, and LLM coding risks.

The Cons

Statistical fragility is severe: the marquee 5.94× (Disabled + Older Adult) rests on n=4, Single Parent + Female on n=3, and most pairs on n<3; the amplification metric is essentially descriptive, which the paper concedes but the abstract still foregrounds.
Internal inconsistencies: the abstract reports “567 verified identity-harm links” while Results/Appendix consistently use 411 markers; Figure 1 text says “12 identity categories” but Table 2 lists 21.
Single-LLM extraction (Gemini Flash) for all causal coding.
The filtering criteria of the reports are described only briefly.
Privileged-group markers are only 8.2% of the corpus, so RQ4's privileged-vs-oppressed comparison is statistically underpowered.
Back to The Index
Final Review · Group 5The IndexAI Risks · 2026