Digital Planet · The Fletcher School, Tufts University · In collaboration with Oxfam

Register of AI Risks & Remedies

103 risks from AI: how much harm each can do, whom it reaches, how soon, and whether any remedy is ready to act on it.

Bhaskar Chakravorti Dean of Global Business, The Fletcher School, Tufts University Chair, Digital Planet Research conducted by Digital Planet in collaboration with Oxfam Version 21 · October 19, 2026

Which risks call for action?

Harm against readiness: every entry in the register

Where does harm outrun the remedies?

Gap between harm and readiness, widest first

Harm percentileReadiness percentile

What risks are in the register?

Entries by domain, split by task

Where in the AI lifecycle risks arise

A risk coded with several stages counts once in each, so bars sum to more than the entries shown.

How serious, and how soon?

Severity

Time horizon

Labels as recorded in the register; “Present / near-term” is kept as its own value.

What causes the harm

Causal entity and intent, coded with the MIT causal taxonomy. The coding is interpretive.

Who is affected?

Population groups named in each entry

Which remedies exist?

Entries each remedy family may reach

Remedies linked to each entry

Remedy approach coded for each entry

Second-level analysis

Cross-tabulations, root causes and the evidence base. All panels follow the filters above.

Remedy coverage by domain: entries each family may reach

Structural drivers and the entries they touch

How many of 105 published taxonomies name each entry

Domain by time horizon

Both fields hold one value per entry, so each entry sits in exactly one cell. The grid shows which domains hold the risks that have not yet arrived.

The evidence base: sources cited by the entries shown

Sources by type

Evidence records linked to the entries shown

The register

Methods note: how the four tasks are counted

The dashboard sorts entries into four tasks by fixed cut-offs for higher and lower harm and readiness. The counts shown throughout use every entry in the Register. Two further counts test how much that split depends on what is counted and where the cut-offs fall.

All 103 entries. Every entry, including the 31 cross-cutting overlays and the 2 aggregate rows. These are the counts used throughout the dashboard.
Distinct harms, same cut-offs. The same placement, counting only the 70 distinct harms. It shows how many separate harms sit in each task.
Distinct harms, recalculated cut-offs. The cut-offs recalculated on the 70 distinct harms alone. This is a sensitivity test, not the adopted ranking.