Vishal Srivastava

Vishal Srivastava works on AI evaluation, model risk, and enterprise governance. His research examines how model errors become harmful actions and what evidence organizations need before deploying AI or continuing to fund it. He serves on the editorial board of Risk Analysis, the Society for Risk Analysis journal, which lists his affiliation as New York Life Insurance.

His writing on AI evaluability focuses on the evidence needed to justify those decisions. In his discussion of AI value, Srivastava argues that adoption counts and favorable anecdotes cannot establish what a system contributed on their own. Sustaining investment requires causal evidence: a credible comparison with what would have happened without the system. Two initiatives with identical value estimates may warrant different decisions if the evidence behind those estimates differs. Weak measurement can also lead an organization to discontinue a beneficial system because it cannot demonstrate the system’s contribution.

He proposes an Evaluability-Based Evidence Score as a way to reason about how strongly evidence supports a governance decision. He presents the score as a conceptual construct, with quantitative formulations left for future work.

• Deployment risk beyond model accuracy. Srivastava and Tanmay Sah’s automation-risk framework separates the probability of technical failure, the chance that a failure causes harm, and the severity of the consequences. Under its assumption that technical failure probability is independent of automation configuration, approval requirements and faster intervention can change deployment risk while the underlying model stays the same. This gives organizations a way to compare spending on oversight with spending on model improvement. The framework is theoretical; its numerical examples illustrate possible tradeoffs, with empirical validation left for future work.

• Verification costs and recovery. With Sah, Dolly Sah, and Kayden Jordan, Srivastava coauthored The Verifier Tax, a March 2026 preprint studying agents that use tools in simulated airline and retail workflows. The team distinguished task completion from completion without policy violations. In tests with GPT-OSS-20B and GLM-4-9B, adding planning and verification raised reported mean agent-token use per episode to 2.0–2.8 times the direct tool-calling baseline. Additional model calls and accumulated context increased computational overhead, while agents often struggled to find a compliant next step after an action was blocked. Some apparently successful runs relied on fabricated identifiers accepted by the simulated environment. The team proposed code checks that tie identifiers to earlier retrieval results, alongside verifier feedback that helps an agent recover after intervention. These remedies were proposals, with their effectiveness left to future testing.

• Limits of testing through outputs alone. In his solo-authored preprint Fundamental Limits of Black-Box Safety Evaluation, Srivastava studies models whose unsafe behavior is triggered by conditions rarely encountered during testing but common in deployment. He examines what evaluators can infer from inputs and outputs when they lack access to model internals and deployment conditions. The analysis concerns constructed worst-case settings, with conclusions dependent on the evaluator’s access and the assumptions about triggering behavior.

• Evaluating evaluation methods. Srivastava also examines the quality of evaluation pipelines. Generated tests, public benchmarks, sampled labels, agent traces, and model-based judges can introduce weaknesses into the evidence used to approve a deployment. He organizes this assessment around sufficiency, credibility, and auditability: whether tests cover the intended operating conditions, whether the methods producing evidence are valid, and whether an independent reviewer can reconstruct and challenge the assessment.

Srivastava and Sah are also coauthors of Evaluable AI, announced as forthcoming from Palgrave Macmillan. They describe its Risk-Based Evaluability Design framework as a way to build AI systems whose evaluation evidence can support consequential decisions and withstand independent scrutiny. In his writing about the book, Srivastava emphasizes that evidence loses relevance as models, data, users, workflows, and operating conditions change. He also argues that effective oversight requires authority to intervene, connecting monitoring to the people accountable for acting on its findings.