AI gatekeeping study: frontier models favor Global South, open models bias North
When AI professors decide who gets research access, geography matters more than merit.
Researchers Nouar AlDahoul, Hezerul Abdul Karim, and Myles Joshua Toledo Tan designed a controlled simulation where LLM-based 'professors' must grant access to only one requester for resources like paywalled articles, datasets, or CVs. Requesters varied only by global region (Global North vs Global South) and academic seniority (undergraduate, PhD candidate, postdoc, tenured professor), holding all other factors constant. This allowed the team to isolate how model behavior changes when only the requester's identity shifts.
The results revealed striking biases. For academic status, LLMs showed inconsistent preferences: some consistently prioritized PhD candidates, while others favored tenured professors—indicating no universal meritocratic standard. More importantly, global region bias split by model architecture. Frontier LLMs systematically favored requesters from the Global South, a pro-equity bias likely stemming from equity-focused safety training (RLHF). Conversely, open-weight and smaller models frequently preferred Global North requesters, reflecting unaligned geographic distributions in their base pre-training data. The study warns that these embedded normative assumptions can distort real-world gatekeeping decisions, urging systematic audits of AI systems for fairness before deploying them in academic or professional access control.
- Simulated LLM-based professors picking one requester among 4 seniority levels × 2 global regions
- Frontier LLMs exhibited pro-equity bias favoring Global South requesters due to safety alignment
- Open-weight and small models flipped to Global North favoritism, reflecting pre-training data imbalance
Why It Matters
As AI automates access decisions, hidden regional and status biases could systematically disadvantage Global South researchers and junior academics.