RAI Risk Taxonomy — human audit

Audit tiers are snapshots of the granularity flow. Each consolidation step merges at the cohesion–attraction crossing of the current inventory and re-derives the boundary. Hierarchy order: General → Agentic → Physical.

Highest-similarity pairs in the source inventories

This is the problem the rest of the page answers. Both inventories carry pairs that are almost the same risk written twice, and the ten closest pairs by cosine similarity, before any consolidation, show what that near-duplication looks like. The question is not whether to merge such pairs, but how far down the similarity scale merging stays defensible.

Our Master inventory (1,612 cards)

#cosCard ACard B
10.9404RAI4-0506 Ownership uncertainty for AI-generated contentRAI4-1368 Uncertain IP status of AI-generated content
20.9252RAI4-0551 Objective misalignment with human intentRAI4-0591 Misalignment with human values
30.9231RAI4-1608 Reputational damage from use or misuse of AI systemsRAI4-1726 Reputational harm attributable to AI systems
40.9178RAI4-0523 Lack of system transparencyRAI4-0596 Lack of model transparency
50.9164RAI4-0228 Explicit hazard non-rejectionRAI4-0229 Implicit hazard non-rejection
60.9142RAI4-0682 Generation of content enabling nonviolent crimesRAI4-0691 Generation of content enabling violent crimes
70.9079RAI4-1055 Compromising privacy by leaking sensitive informationRAI4-1072 Compromising privacy by leaking private information
80.9031RAI4-0554 Active loss of controlRAI4-0856 Loss of control risks
90.9007RAI4-1732 Disinformation spread via generative AIRAI4-1733 Generative AI-driven disinformation spread
100.8994RAI4-0863 AI-generated advice influencing user moral judgmentRAI4-0870 Financial market instability from GPAI agents

MIT AI Risk Repository (1,795 entries)

#cosEntry AEntry B
11.0000Weidinger2023 Erosion of trust in public informationLi2025 Erosion of trust in public information
21.0000Weidinger2023 Unfair capability distributionLi2025 Unfair capability distribution
31.0000Vidgen2024 Self-harmGipiškis2024 Self-harm
40.9994Weidinger2023 Propagating misconceptions/ false beliefsLi2025 Propagating misconceptions / false beliefs
50.9963Weidinger2023 Toxic contentLi2025 Toxic content
60.9954Gipiškis2024 Reinforcement learning AI (Training design related)Gipiškis2024 Reinforcement learning AI (Training performance related)
70.9951Gipiškis2024 Supervised/unsupervised AI (AI training performance related - Accuracy)Gipiškis2024 Supervised/unsupervised AI (AI training performance related - Reliability)
80.9919Gipiškis2024 Supervised/unsupervised AI (AI training performance related - Robustness)Gipiškis2024 Supervised/unsupervised AI (AI training performance related - Reliability)
90.9919Gipiškis2024 IntentionalGipiškis2024 Unintentional
100.9904Gipiškis2024 Supervised/unsupervised AI (AI training performance related - Robustness)Gipiškis2024 Supervised/unsupervised AI (AI training performance related - Accuracy)
Entry labels prefixed by their source citation. Placeholder rows without a category label are excluded. 4 of the ten pairs are verbatim duplicates carried in from different source papers, which is the kind of redundancy the crossing boundary removes first; our inventory has no exact duplicates, so its closest pairs are paraphrases rather than repeats.

How the boundary is measured

The figure below is the flagship figure of the accompanying manuscript. It carries the whole argument in three panels: where the space breaks, where the last defensible merge boundary sits, and how that boundary moves once merging begins.
Two transitions, the cohesion-attraction crossing, and the granularity flow

Concepts1

Threshold graph. Cards are vertices; two cards are joined when the cosine similarity of their embeddings is at least τ. Connected components are the merge clusters, which makes the construction identical to single-linkage clustering read at one cut.
Cohesion Φcoh(τ). The pooled mean pairwise similarity inside clusters. It answers: how alike are the cards we have decided to treat as one?
Attraction Φatt(τ). The mean over clusters of the single highest similarity to any card outside the cluster. It answers: how strongly does the nearest outsider pull? Being an extreme-value statistic, it is the demanding half of the comparison.
Crossing τ*. The smallest τ at which cohesion still dominates attraction. Below it, the typical card inside a cluster is less like its own cluster than the cluster is like its nearest neighbour, so the partition has stopped describing the data.
Granularity flow. Merging at τ* changes the density of the space, so the boundary moves. Iterating merge, re-derive, merge again yields a trajectory of boundaries and inventories; a tier is one state along that trajectory rather than a separately designed release.

Measures2

τ1, chaining onset. The largest drop of Φcoh over a 0.01 window on a 10−4 grid. Below it, clusters grow by transitive chains rather than by mutual similarity.
τ2, worst-case collapse. The same changepoint statistic applied to the per-cluster minimum. Below it, the worst pair inside a cluster is no longer a merge in any defensible sense.
Uncertainty. Every curve is the mean of 1,000 subsample replicates at 80% of cards, with bands at two standard deviations; the changepoints carry the spread of those replicates.
Per-step null. Each merge group is compared with size-matched random groups drawn on the same inventory, scored as z = (observed minimum − null mean) / null s.d.
Validity rule. A step is admissible when at least 90% of its merge groups exceed the null by two standard deviations. The rule delimits the admissible range by itself, without an externally chosen stopping point.

Results3

a, Two transitions bound the usable range: τ1 = 0.818 ± 0.010 and τ2 = 0.690 ± 0.009. Between them lies a fidelity-oriented regime; below τ2 a giant cluster has absorbed the inventory.
b, The crossing is at τ* = 0.828 ± 0.004 across subsamples and 0.833 on the full inventory. It agrees with τ1 although the two are measured from unrelated statistics, which is what makes the boundary credible rather than merely convenient.
c, The flow runs 28 consolidations, carrying 1,612 cards to 32 and the boundary from 0.833 to 0.606. The validity rule first fails at step 15, so steps 1 to 14 are the admissible range; F1 to F5 are the first five of them. Merging in one pass at a low threshold instead of stepwise does not reach the same place: a single cut at τ = 0.7753 leaves 632 clusters with a giant cluster of 868 cards, against 901 well-formed cards with a largest group of 56 along the flow.
1 Definitions are stated in full in the Methods section of the manuscript. 2 Grids, windows, replicate counts and null specifications are fixed before analysis and reported with the code in analysis/. 3 Values quoted here are the released figures; the manuscript reports them with full derivations and the accompanying null tests.

How consolidation runs

The procedure in full, written as an algorithm. Everything turns on one quantity, the gap Δ(τ) = Φcoh(τ) − Φatt(τ): how much more a cluster resembles itself than it resembles the nearest card outside it. Merging is defensible while the gap is positive. Two percolation thresholds, measured independently of the gap, bound the region where the question is even meaningful.
Algorithm 1   Granularity flow by the cohesion–attraction gap
Input card set V, unit embeddings e, grid δ = 10−4, null level (90%, 2σ)
Output tier sequence F1, F2, …
 1  function Gap(S, τ)
 2      C ← connected components of G(τ) = (S, {(i,j) : cos(e_i, e_j) ≥ τ})
 3      Φ_coh ← mean of cos(e_i, e_j) over all pairs inside the same component
 4      Φ_att ← mean over components of max cos(e_i, e_j), j outside the component
 5      return Φ_coh − Φ_att                        ▷ the gap Δ(τ)

 6  function PercolationThresholds(S)
 7      τ₁ ← τ of the steepest fall of Φ_coh over a 0.01 window   ▷ chaining onset
 8      τ₂ ← τ of the steepest fall of the per-cluster minimum      ▷ giant cluster forms
 9      return (τ₁, τ₂)                              ▷ 0.818, 0.690 on the Master set

10  S₀ ← V ;  t ← 0
11  (τ₁, τ₂) ← PercolationThresholds(S₀)              ▷ above τ₁: isolated pairs
                                                    ▷ below τ₂: one blob, no partition left
12  repeat
13      t ← t + 1
14      τ*_t ← min { τ on the δ grid : Gap(S_{t−1}, τ) ≥ 0 }   ▷ last τ with a positive gap
15      assert τ*_t ≥ τ₁(S_{t−1}) > τ₂(S_{t−1})            ▷ the boundary sits above chaining
16      M ← { components of G(τ*_t) with more than one member }
17      for each g ∈ M
18          z(g) ← (min-similarity(g) − μ_null(|g|)) / σ_null(|g|)   ▷ size-matched random groups
19      if |{ g ∈ M : z(g) > 2 }| / |M| < 0.90 then break       ▷ validity rule; first fails at t = 15
20      S_t ← { medoid(c) : c component of G(τ*_t) }        ▷ representatives are original cards
21      emit F_t ← S_t
22  until S_t = S_{t−1}
23  return F₁ … F₁₄                                ▷ released for audit: F₁ … F₅
Why the gap and not a fixed threshold. Line 14 does not ask whether similarity is high in absolute terms; it asks whether the inside still beats the outside. That comparison is scale free, which is why the same line returns 0.8329 here and 0.8190 on the MIT repository without a parameter being touched.
Why percolation matters. Lines 7 and 8 locate the two thresholds at which the connectivity of the graph changes character: above τ₁ components grow by mutual similarity, between τ₁ and τ₂ they grow by transitive chains, and below τ₂ a giant component has swallowed the inventory (97% of cards by τ ≈ 0.70). Line 15 records the finding that the gap boundary lands at or above the chaining onset, so the two criteria agree although they are measured from unrelated statistics.
Why iterate. Line 20 changes the inventory, and a sparser inventory has a different density, so τ* must be measured again in line 14. The empirical dependence is τ* ≈ 0.576 + 0.035 ln n, roughly +0.024 per doubling of the inventory.
Why it stops by itself. Line 19 is the only exit, and it is a test rather than a preference. It first fires at t = 15, which is what makes steps 1 to 14 an admissible range rather than an arbitrary cut-off.

Granularity flow tiers

Two transitions and the granularity flow of the Master inventory
a, Within-cluster cohesion against the merge threshold on the Master inventory, as the mean of 1,000 subsample replicates with 95% bands. The pooled mean pairwise similarity breaks at τ1 = 0.818, where similarity chaining sets in, and the per-cluster minimum collapses at τ2 = 0.690, where the worst case within a cluster stops being a merge at all. Triangles mark the five crossing boundaries emitted by the flow. b, The flow itself: each consolidation lowers the crossing boundary τ*t (green, left axis) and the inventory size nt (purple, right axis), from 1,612 cards to 792 in five steps. Tiers F1 to F5 are these five states.
Tierτ*CardsMerge groupsAbsorbedG / A / PRole
Master1,6121,154 / 155 / 303canonical inventory
F10.83291,383109229906 / 140 / 337fidelity tier (crossing)
F20.80331,154121458734 / 128 / 292second consolidation
F30.79031,03881574653 / 112 / 273third consolidation
F40.775390182711568 / 92 / 241compression tier
F50.763479270820491 / 83 / 218extended compression tier
Merge groups are counted per consolidation step and absorbed cards cumulatively from the Master inventory. Highlighted rows are released for human audit; the F4 audit page reports 211 refined groups, the cumulative count over steps 1–4. The Societal Safety axis is one concept family applied at three scopes, so its cards are counted under General, Agentic or Physical (RAI3-G|A|P-SOC-nn share the same numbering and meaning).

Cross-corpus replication · MIT AI Risk Repository

The same pipeline applied to an independent inventory, with no parameter retuned. Entries are the repository's risk categories and subcategories, embedded with the same encoder.
MIT replication: transitions and granularity flow
a, Within-cluster cohesion against the merge threshold. The pooled mean pairwise similarity breaks at τ1 = 0.806 ± 0.004, where similarity chaining sets in, and the per-cluster minimum collapses at τ2 = 0.629 ± 0.009, as means over 100 subsample replicates with 95% bands. Triangles mark the five crossing boundaries emitted by the flow. b, The flow itself: each consolidation lowers the crossing boundary τ*t (green, left axis) and the inventory size nt (purple, right axis), from 1,810 entries down to 727 in five steps.
Tierτ*EntriesMerge groupsAbsorbedMedian zRole
Source1,810MIT repository inventory
M10.81901,4231893874.79fidelity tier (crossing)
M20.79081,2301045803.68second consolidation
M30.76961,0221157883.53third consolidation
M40.7545879879313.17compression tier
M50.7374727991,0833.00extended compression tier
Merge groups and median z are per consolidation step; absorbed entries are cumulative. Every step clears the validity rule, with at least 98.9% of merge groups exceeding the size-matched random-group null by two standard deviations. The domain columns are omitted because the MIT repository uses its own top-level scheme rather than the General / Agentic / Physical families. This replication is reported for reference only and is not released for human audit.

F1 · τ* = 0.8329

1,383 cards · the crossing boundary of the original inventory · L3 re-assigned by the seed-anchored hybrid EM.
Open audit

F4 · τ* = 0.7753

901 cards · four consolidations · consolidated labels and definitions rewritten for representativeness.
Open audit

F5 · τ* = 0.7634

792 cards · five consolidations · L3 propagated from the F4 representatives.
Open audit

Korean terminology reference

36 entries aligned with TTA, KISA, NIS, ETRI and the AI Framework Act — the basis for the Korean wording of card labels and definitions.
Open glossary

How to audit

Judge each card on three criteria — ① description adequacy (is the risk correctly described?), ② L3 mapping (is the assigned L3 appropriate?), and ③ redundancy (does it duplicate other cards?). Record the verdict, add a note where useful, and press Export judgments before closing; judgments are not saved automatically.