AI Collaboration Abandons the Advisor Model

LLMs are being deployed in matchmaking contexts where caste hierarchies have historically shaped marital decisions. Do these models reproduce or disrupt caste-based stratification?

Method: Controlled audit of five LLM families (GPT, Gemini, Llama, Qwen, BharatGPT) using real matrimonial profiles with varied caste identities (Brahmin, Kshatriya, Vaishya, Shudra, Dalit) and income levels. Same-caste matches received ratings up to 25% higher on a 10-point scale than inter-caste matches. Inter-caste matches were further ordered according to traditional caste hierarchy across all models tested.

Caveats: Tested on South Asian matrimonial contexts. Other cultural hierarchies may manifest differently.

Reflections: Can fine-tuning on counter-stereotypical training data disrupt these hierarchical patterns, or are they too deeply embedded in pre-training? · Do users perceive LLM-mediated matchmaking recommendations as more 'objective' than human matchmakers, thereby legitimizing caste bias? · How do these biases interact with other identity dimensions like religion, region, or disability status in multi-attribute matching scenarios?

IDE-integrated AI assistants operate conversationally within codebases, but empirical evidence of how developers actually use them in real workflows remains limited.

Method: Analysis of 74,998 developer messages from 11,579 chat sessions across 1,300 repositories using Cursor and GitHub Copilot. Programming work reorganizes around three patterns: progressive specification (iterative refinement rather than upfront task specification), cognitive redistribution (delegating diagnosis and validation to AI rather than engaging code directly), and active collaboration management (externalizing plans into persistent artifacts and constraining AI autonomy through context injection).

Caveats: Data from public repositories where developers committed chat logs. May not represent proprietary or sensitive codebases.

Reflections: Does cognitive redistribution to AI erode developers' ability to diagnose problems independently over time? · What happens when progressive specification fails—how do developers recover from conversational dead-ends? · Are there systematic differences in how novice versus expert developers manage AI autonomy and context injection?

The standard AI-as-advisor pattern (AI recommends, human accepts/rejects) suffers from advice underuse, overreliance, and skill deterioration. Can preserving judgment independence improve outcomes?

Method: Hybrid confirmation tree (HCT) elicits human and AI judgments independently. Agreement accepts the decision; disagreement triggers a second human tiebreaker. Tested across 10 datasets (medical diagnostics, misinformation detection) and four datasets with AI explanations. HCT outperformed AI-as-advisor in all datasets. Signal detection analysis reveals people cannot discriminate well enough between correct and incorrect AI advice to make the advisor pattern work.

Caveats: Requires access to multiple human judges. Cost-benefit depends on decision stakes and judge availability.

Reflections: Does the HCT maintain its advantage when AI accuracy significantly exceeds human accuracy? · How does the tiebreaker role affect the second human's judgment—do they anchor on knowing there was disagreement? · Can the HCT pattern be adapted for real-time decisions where independent elicitation isn't feasible?