Research map / Large language models / AI and society
Selective Response Strategies for GenAI
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes "selective response," where GenAI strategically withholds or degrades answers to emerging-topic queries so users shift to human forums, generating data that raises GenAI's long-term revenue and user welfare, with approximate revenue optimization and welfare-regulation results.
Problem
GenAI answering every query can starve human Q&A forums (e.g., Stack Overflow) of the data GenAI needs to stay accurate, creating a negative feedback loop that harms long-term revenue and user welfare.
Approach
Model GenAI as choosing a per-round response proportion x_t; lower responses push users to a human forum that generates data, which raises future accuracy and user proportions. Shows selective response can Pareto-dominate always-answering, gives an approximately optimal DP algorithm (ASR) for revenue, a welfare-constrained variant, and sufficient conditions for regulator interventions to raise wel
Key results
- Selective response can Pareto-dominate always-responding: in a worked example revenue is about 5% higher and welfare about 7.6% higher than full response.
- The ASR dynamic programming algorithm achieves an additive O(eps T^2) approximation of optimal revenue in O(T^2|A|/eps) time.
- Revenue price of always responding (RPAR) is unbounded: for every M there is an instance with RPAR > M.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use selective response by answering no queries in rounds 1–4, then all queries | GenAI · 10-round model example | revenue: roughly 5% higher than U(\Bar{\bft{x}}) | Helped |
| Use selective response by answering no queries in rounds 1–4, then all queries | GenAI · 10-round model example | user welfare: about 7.6% higher than W(\Bar{\bft{x}}) | Helped |
| Reduce the response level in one round while keeping other rounds unchanged | Any selective response strategy · Future rounds after the modified round | GenAI user proportion | Helped |
| Use a selective response strategy instead of always responding | GenAI · Instances in the model | revenue and user welfare | Mixed |
| Never answer any query | GenAI · 5-round instance with a(D)=1-e^{-0.4D}, γ=1, β=3, w^s=0.7, and r(p)=p | revenue: U(x)=0 < U(\Bar{x}) | Hurt |
| Never answer any query | GenAI · 5-round instance with a(D)=1-e^{-0.4D}, γ=1, β=3, w^s=0.7, and r(p)=p | user welfare: W(x) = T w^s = 3.5 > W(\Bar{x}) | Helped |
| Answer only in the final round | GenAI · 5-round instance with a(D)=1-e^{-0.4D}, γ=1, β=3, w^s=0.1, and step-function revenue | revenue: U(x) = r(p_T(x)) = q(4, 1) > 0.89; U(\Bar{x}) = 0 | Helped |
| Answer only in the final round | GenAI · 5-round instance with a(D)=1-e^{-0.4D}, γ=1, β=3, w^s=0.1, and step-function revenue | user welfare: W(x) < 1.122; W(\Bar{x}) > 1.17 | Hurt |
| Use a non-binary response schedule (0.04, 0.97, 1) | GenAI revenue maximization · 3-round instance with a(D)=0.7(1-e^{0.4D})+0.3, γ=1, r(p)=p, β=41, and w^s=0.66 | revenue: U(x) - U(\Bar{x}) > 9.71 · 10^{-6}; U(x) - U(x^1) > 9.71 · 10^{-6}; U(x) - U(x^2) > 7.9 · 10^{-6}; U(x) - U(x^3) > 7.89 | Helped |
Surprising findings
- The optimal strategy can require non-binary response levels, so step-function (0/1) strategies are not always optimal.
- Withholding responses can increase welfare and revenue simultaneously, challenging the always-answer default.
Limitations
- Assumes a single GenAI platform, which is unrealistic; competition among platforms is left to future work.
- Forum is assumed non-strategic and results rely on a Lipschitz accuracy assumption (L_a <= 4/beta).
- Accuracy and user utility are stylized abstractions; actual GenAI performance evaluation is not modeled.
Experimental setup
Metrics: revenue cumulative user welfare approximation guarantee
Baselines: always-responding strategy
Builds on: Price of Anarchy PARS-MDP information design
Closest papers by meaning
- Braess’s Paradox of Generative AI (AAAI 2025 · similarity 0.90)
- Human vs. Generative AI in Content Creation Competition: Symbiosis or Conflict? (ICML 2024 · similarity 0.84)
- Position: AI Competitions Provide the Gold Standard for Empirical Rigor in GenAI Evaluation (ICML 2025 · similarity 0.84)
- Generative Social Choice: The Next Generation (ICML 2025 · similarity 0.84)
- Diverse Yet Biased: Towards Mitigating Biases in Generative AI (Student Abstract) (AAAI 2024 · similarity 0.84)
- Position: Evaluating Generative AI Systems Is a Social Science Measurement Challenge (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.