Financial firms using generative AI may need to test how prompts shape investment assessments, after a CFA Institute study found framing effects across nine OpenAI models in hypothetical scenarios.
Published Sept. 10, 2026, the report by James Tait and Brian Pisaneschi, CFA, covered 10 investment situations. Each model received 900 prompts. In one client-retention test, researchers described the same annual figures as either 77% retention or 23% churn.
The largest differences appeared when models assessed drawdown performance and client retention. The experiment measured model responses to hypothetical information, not actual investor behavior or portfolio performance.
In tests of ways to limit the effect, adding both interpretations of information to a prompt reduced framing effects the most. Explaining framing bias in the prompt did not significantly reduce the effect. A separate screening model reduced it in most tested settings, but results varied by scenario and reasoning level.
The authors said bias cannot be fully removed and recommended comparison testing and human review across the workflow. The controlled experiment used OpenAI models only, so it does not establish how the results would apply to other providers or live investment operations.
India's rules assign responsibility to financial intermediaries using these systems. A February 2025 amendment by the Securities and Exchange Board of India (SEBI) covers investor data, legal compliance and results from AI or machine-learning tools that regulated intermediaries act on or implement, including third-party products.
SEBI's June 2025 consultation paper proposed fairness safeguards, continuous testing and monitoring, and independent audits. It linked growing AI use to greater access to data and computing and advances in software and hardware. In an Aug. 19, 2026, speech, SEBI Chairman Tuhin Kanta Pandey said the regulator would issue guidelines for responsible AI and machine-learning use in markets “shortly.”
