case studies
How new context changes an AI allocation decision
Sequential vignettes reveal how artificial agents revise a fixed $1,000 allocation as information about a recipient unfolds.
Exploring the updating that a one-shot decision can conceal.
The question
How does an artificial agent revise a resource allocation when it learns more about a person’s circumstances, actions and motives? Contextual Resource Allocation Vignettes (CRAV) hold the allocation problem fixed while progressively revealing information about one recipient.
Study design
- 10 vignettes · 5 stages
- Each vignette introduced a sequence of contextual revelations about a single recipient.
- 500 trajectories
- 50 stochastic GPT-4o conversational runs per vignette generated 2,500 allocations in total.
- −19.6 to +16.4 points
- The reported range of stage-to-stage reallocation changes, expressed in percentage points.
What it revealed
Culpability-related information sometimes produced large penalties. Later information about constraints, need, mitigating motives or corrective effort could offset or reverse those penalties.
How to interpret it
These are configuration-specific findings about artificial agents, not a benchmark of human judgement. The method makes sequential narrative updating visible for behavioural auditing and further research.
Working paper
Context-Sensitive Resource Allocation: Evidence from Sequential Vignettes Using Artificial Agents.
Steve J. Bickley, Ho Fai Chan, David Stadelmann, Massimiliano Tani and Benno Torgler. ARC BITA Working Paper Series.