If you use LLMs to generate recommendations, they can favor or disfavor user groups; auditing with a generative-aware fairness test prevents reputational and regulatory risk.
ChatGPT shows measurable unfairness on movie recommendations when measured by pairwise ranking agreement (PRAG*@20).
Numbers: Movie PRAG*@20 SNSR up to 0.2191; SNSV up to 0.0828 (Table 1)
FairPy makes bias audits repeatable and faster across multiple metrics and models, but mitigation effects are metric-dependent so teams must validate fixes with several tests before deployment.
FairPy collects common bias metrics and mitigation methods into one toolkit.
PCN-Rec turns persuasive LLM outputs into auditable, machine-checked recommendations so platforms can guarantee and log policy compliance without trusting LLM explanations.
On MovieLens-100K, selecting W=80 gave 551 of 943 users with at least one compliant slate inside the window.
Numbers: 551 / 943 users (W=80)
Decentralized agent systems can amplify bias and create unfair outcomes; adding fairness checks and incentives early reduces legal, reputational, and customer-risk without obvious short-term performance loss in toy tests.
Applying a demographic-parity fairness adjustment narrowed the final group reward gap from 45 points to 5 points in the 10-agent simulation.
Numbers: With fairness: Group A 375 vs Group B 370; Without: 390 vs 345 (gap 5 vs 45).
Pooling audit queries across teams cuts the number of queries needed to detect bias and improves fairness estimates; avoid heavy pre-coordination when many teams audit the same platform because it can increase error.
Collaboration reduces average DP estimation variance versus independent audits.
Numbers: Empirical DP error reduced by 17.4%–24.6% (Fig.3, Section 6.1).
Task-vector editing is a low-cost way to tune subgroup parity without full retraining; it can be used as an operational knob to reduce worst-case demographic gaps while keeping accuracy near existing adapt methods.
Uniformly scaled task-vector merges can reduce group disparities while keeping accuracy close to FFT/LoRA.
Numbers: Civil Comments (DistilBERT): Task Addition accuracy ≈0.9395; worst-DPD 0.0454; worst-EOD 0.3358 (Table 2)
How agents phrase decisions affects cooperation and task success; monitoring and nudging tone and explanations reduces coordination failures and builds trust in agentic workflows.
Respectful tone and clear justification increase proposal acceptance even when resource splits are identical.
Numbers: High-High (5:5) acceptance = 1.0 (Table 3)
