Charta is on Google Play: practise on the go. Get the app
Charta

Explain where AI and large language models fall short or pose risks in investment work

CFA® Level I · Quantitative Methods · Introduction to Financial Data Science · QM.20.d

The idea

AI tools bring risks that the user remains responsible for managing. Hallucination is fluent output that is false or unsupported, such as invented figures or citations. Bias arises when training data reflect historical prejudice, which the model can reproduce. Pasting confidential client or firm data into an external tool risks privacy breaches and leakage. Deep models are hard to explain, which complicates justifying decisions to clients and regulators. Outputs can change with small changes in prompt wording. In backtests, a model trained on text written after the test period introduces look-ahead bias. Together these form model risk, and the standard response is validation, documented controls and human oversight of every consequential output.

A practice question

From the question bank, with the reason each wrong option is wrong.

In the context of large language models, a hallucination is:

  1. a refusal to answer a sensitive questionWhy not: Declining to answer is a refusal, not the generation of false content.
  2. a slow response caused by a long promptWhy not: Response time is a performance issue, unrelated to the accuracy of the content.
  3. fluent output that is false or unsupported — correct

LLMs generate likely-sounding text rather than retrieving verified facts, so they can state wrong figures or cite sources that do not exist with apparent confidence.

A harder one

A credit model trained on historical lending decisions consistently scores one group of applicants lower for reasons unrelated to credit risk. This illustrates:

  1. a hallucinated credit scoreWhy not: The model is not inventing facts; it is reproducing patterns present in past decisions.
  2. sampling error in a small data setWhy not: The effect is consistent and systematic, which points to bias rather than random error.
  3. bias inherited from training data — correct

Models learn from the data they are given. If past decisions were biased, the model can reproduce or amplify that bias, which is why training data and outputs need testing for fairness.

The hardest tier

Two analysts get different ratings on the same bond from the same LLM after slightly rewording an otherwise identical request. This illustrates:

  1. prompt sensitivity — correct
  2. look-ahead biasWhy not: Look-ahead bias is using information unavailable at the time; nothing here depends on dates.
  3. data leakageWhy not: No confidential or future information has escaped; the change comes from wording alone.

LLM outputs can shift with small changes in wording, so firms should test prompts for robustness and not treat a single response as a stable model output.

In the app: 6 practice questions on this outcome, a narrated walkthrough with diagrams, spaced review that brings it back before you forget it.

Take the free diagnostic →

More in Introduction to Financial Data Science