Same facts, different verdict: CFA Institute tested nine AI models on positive versus negative framing

Framing bias is one of behavioral finance's best-known findings: the same facts, described as a gain or as a loss, lead to different decisions. As analysts start to use language models to summarize results, judge drawdowns or screen credit, a fair question is whether the model inherits the habit. The CFA Institute Research and Policy Center took it on in a report published on September 10, 2026, Managing LLM Bias in Investing: From Detection to Mitigation, by James Tait and Brian Pisaneschi, CFA.

The test

The authors sent 900 prompts to each of nine OpenAI models across 10 investment scenarios. The scenarios included portfolio performance, drawdown performance, client retention, credit defaults, concentration risk and stress testing. In each, the researchers changed only whether the identical financial information was presented positively or negatively, then compared the models' assessments.

That design is simple and useful. A model with no framing sensitivity should give the same assessment of two descriptions of the same numbers. Any systematic gap between the positive and negative versions is the bias.

What they found

Framing mattered. The effect was strongest and most consistent in evaluations of drawdown performance and client retention. Our illustration, not the report's: a retention rate and the matching attrition rate are the same fact, yet a model can read them differently.

Nothing in the public summary of the report tells us how large the gaps were per scenario or per model, and we have not reproduced the experiment. We can report the direction and the ranking of scenarios, not an effect size.

Fixing it: what worked and what did not

The more practical half of the report compares ways to reduce the bias.

• Defining the bias in the prompt. Telling the model what framing bias is, and asking it to avoid it, produced little improvement.
• A human restructuring the prompt. Manually presenting both the positive and the negative perspective produced the largest reduction.
• A second model as a guardrail. A separate bias-detection LLM identified and restructured biased prompts effectively in most cases. It was faster and more scalable, but did not outperform the human approach.

The lesson is about process, not wording. Asking an AI to be objective is a weaker control than showing it the evidence both ways.

Governance, in practice

The report also frames bias as a workflow problem. It separates human bias in how an analyst asks, implicit model bias from training, and the observable choices a model makes about data, sources and analytical steps. For firms, it recommends comparative testing, structured prompt templates, predefined data sources and steps, independent model review, and escalation checkpoints for human review. Human review is not automatically a safeguard either: a reviewer who simply approves a recommendation without examining its assumptions provides limited protection.

A test you can run in ten minutes

Take a one-paragraph summary of any fund's results. Write it twice with identical numbers: once as "the fund fell sharply from peak to trough" and once as "the fund recovered most of the way from its low." Ask your AI the same question about both, in fresh chats, a few times each. If its verdict moves with the wording, you have found a framing effect in your own tool. Then try the report's fix: give it both framings in one prompt and ask it to reconcile them. This is a demonstration of a habit, not a measurement of any product.