FT analysis shows AI chatbots err on financial queries at rates above 50 percent
Empirical tests confirm AI models produce unreliable financial answers at rates exceeding 50 percent. The gap stems from absent numeric verification and source grounding rather than training scale. Deployment without mandatory human review creates direct compliance and liability exposure.
The report documented consistent failures on compound interest calculations, regulatory thresholds, and tax bracket mechanics. Error patterns matched known LLM hallucination modes documented in arXiv:2305.18223 where numeric grounding collapsed under multi-step arithmetic. No chatbot surfaced source documents or flagged uncertainty ranges.
Benchmark data from the paper aligns with prior results in "FinBen" (arXiv:2307.16616) showing median F1 scores below 0.45 on financial QA subsets. Operational impact appears in advisory workflows where downstream compliance review must now cover every model output rather than sampled cases.
Regulators already reference similar error distributions in the EU AI Act high-risk classification for credit scoring and investment advice. Firms deploying these systems without per-query verification layers face direct liability under existing fiduciary standards.
Next measurable signal will be SEC or FCA enforcement actions once audit trails capture model provenance at scale.
SEC: By Q4 2025, at least three enforcement actions will cite AI-generated financial advice with documented error rates above 40 percent.
Sources (3)
- [1]Primary Source(https://www.ft.com/content/c0cd359d-df84-4208-a789-ffa864b43666)
- [2]Supporting Source(https://arxiv.org/abs/2305.18223)
- [3]Supporting Source(https://arxiv.org/abs/2307.16616)