Data collection and channel separation

The primary channel contains 10,449 OECD AI Incidents Monitor records from the frozen 24 August 2026 snapshot. The secondary channel contains 1,751 AI Incident Database records from the 19 October 2026 snapshot. OECD AIM uses news-based collection, while AIID curates incidents editorially. The event-year ranges in these snapshots are 2020–2026 and 1983–2026, respectively.

Each graph retains every record in its own channel. Some AIM records are linked to AIID, so the two totals must not be added as independent events. The paper's pooled sensitivity set removes 1,083 linked AIM rows and retains all AIID rows, yielding 11,117 records. MIT Tracker classifications and CSET labels enrich AIID for external validation and do not form a third incident channel.

Hazards are excluded from these two incident graphs. The paper examines them separately as potential harms. The AIM collection change on 1 November 2024 is also examined separately rather than treated as a change in underlying risk prevalence.

Cleaning and non-circular card mapping

Source identifiers and channel membership are preserved. Linked identities are reconciled explicitly rather than merged through title resemblance alone. Each incident is embedded from its title, description, and AI-system text using the frozen BGE-M3 dense representation. Cosine similarity retrieves the three closest cards from the 622-card, 55-family inventory.

Source risk categories, harm labels, MIT classifications, and CSET labels are excluded from embedding inputs so they can be used independently for validation. An activated card is a retrieved candidate, not an expert-adjudicated attribution. The same model and retrieval rule are used in both channels.

Network calculation and display

An undirected edge joins two cards when at least one incident activates both. Three distinct cards produce three unordered card pairs per incident. Repeated pairs increment the co-activation count. Degree counts distinct neighboring cards in the full channel network. Node area increases linearly with observed incident counts above a small minimum visible area, with a larger AIID node-size display scale for readability. Cross-channel count comparisons should use the detail-panel counts. Edge widths use co-activations per 1,000 incidents, with one scale shared across channels.

The overview starts with the 25 cards with the highest observed incident counts in each domain, explicitly includes the control-transfer card, and retains the 240 strongest ties among those candidates. Displayed isolates are removed. The full view contains every activated card and every observed co-activation edge. Filtering changes the display, not the reported full-network metrics.

Both plots share a deterministic joint layout with seed 20261021. Positions are separated to reduce node overlap. Curved links follow the AI risk space convention. Every L4 node reads its exact color by card ID from the same semantic_space.json source as AI risk space. Domain legend anchors are blue, orange, and green; individual semantic shades can blend near domain boundaries. A red ring highlights automated-driving control-transfer failure without replacing its domain color. Positions are for comparison and do not measure causal or geographic distance.

Representation and statistical checks

For domain d in channel c, let a_dc count top-three card assignments, N_d count stated cards in that domain, and N = 622. The revelation ratio compares the domain's share of assignments with its share of the inventory.

Each incident contributes three assignments. Physical cards account for 10.1% of the inventory and 3.0% of AIM's 31,347 assignments, giving a revelation ratio of 0.30; AIID's corresponding ratio is 0.61. A ratio below 1 indicates lower representation relative to inventory size. It does not establish lower real-world occurrence or identify why records are missing.

The paper resamples incidents 2,000 times for confidence intervals and permutes embodiment labels 2,000 times within the closest card's home domain. Regression also adjusts for home domain, top-one similarity, and the first-to-third similarity gap. Network tests shuffle domain labels 1,000 times while retaining channel topology and domain counts. These within-channel nulls account for unequal domain sizes.

Bias-corrected Chao1 uses cards activated once and twice to estimate a lower bound on card richness under the sampling model. Its gap from the inventory is an extrapolation gap, not a count of risks that can never be reported. Incident absence should not remove an inventory-based scenario from RAI evaluation.

What the evidence supports

Reported incidents bring risks from different operating domains together, with automated-driving control-transfer failure linking General, Agentic, and Physical neighborhoods. This motivates evaluating the composed service and its human–automation handovers across domains.

Public reporting is selective. Disclosure, collection rules, language and geographic coverage, event composition, text detail, and mapping can all influence visibility. Network degree is connectivity, not importance; co-activation is a retrieved relationship, not a proven causal chain. The independent mapper and masking checks support parts of the paper's findings, while the cross-encoder reversal marks a boundary under stricter card selection.

Sources and reproducible aggregates

OECD AI Incidents Monitor · AI Incident Database snapshots · HEART AI risk space

• Channel networks and display coordinates · JSON
• All activated-card metrics · CSV
• All co-activation edges · CSV
• Input hashes and integrity checks · JSON
• Definitions and reproduction commands

Incident summaries are reproduced from the existing source snapshots with links and attribution. Source article full texts and human-adjudication sheets are not included.