Separate the assessments
The DCI framework [B, chapter 5] examines how likely harm is, how serious it could be and how widely it could affect people. It then considers the capacity to manage that harm, including law, institutional capability, the data environment and public confidence. Its ratings are qualitative. A favourable assessment cannot compensate for missing minimum safeguards. Its deployment checklist excludes automatic decisions on eligibility, exclusion, sanctions and appeals.
SAFE AI [F, pp. 12–17] sets the level of oversight according to the use and its context. Tier 1 covers reversible advisory work with no direct influence on decisions about people. Tier 2 covers influence over operational choices. Tier 3 covers direct effects on assistance, protection or information, or a high cost of error. Where the tier is uncertain, use the higher one. Its Transparency Card records the system’s development and use. No tier is assigned to the examples in this draft.
AidGPT proposes keeping each assessment in a separate field, with its source reference. A CDAC tier is not equivalent to a DCI outcome. Do not average them or treat one framework as overriding the other. Where they disagree, the accountable organisation must explain the difference in writing and decide how to proceed.
Examine the humanitarian setting
In each setting, consider who the proposed process could miss, including displaced people, minority-language speakers, people without documents and people with disabilities. Could revealing someone’s location or an association expose them to harm? Specify another route for people who cannot use the system. These questions support review. They do not amount to a completed protection assessment.
Show how the use upholds humanity, impartiality, neutrality and independence [F]. Explain what the responsible person can inspect and correct. Before operational use, teams must provide tests justified for the local setting, arrangements for handling incidents and conditions for stopping the system. Review is required if the purpose changes or a result cannot be explained.
Outstanding source checks
The SAFE AI user guide, tier rules and card structure were inspected [F, pp. 12–17 and 88]. Check each field against the separate tools before operational use. DCI’s website and report number the companion tools differently. The report describes a maximum rule for combining ratings, but also raises the overall rating when every dimension is medium. This draft applies no scoring rule.
Sources
F. SAFE AI Governance Framework