British fintech company Saturn (Saturn Fintech, London) tested 18 popular AI models, including ChatGPT, Claude, Copilot, Grok, and Gemini, on 121 real financial questions in its September report “Artificial Authority: Should you trust AI to deliver financial advice?” — and found that more than half of the answers were wrong, with errors becoming almost inevitable on complex topics. Financial Times published a piece about the study. Meanwhile, the audience is more trusting than the models’ accuracy allows: according to a PensionBee survey, most adult US users are ready to act on a chatbot’s advice without verifying it.

image
image

What happened

Saturn Fintech, a London-based financial company, published the report “Artificial Authority: Should you trust AI to deliver financial advice?” in September 2026. The test involved 18 popular AI models, including ChatGPT, Claude, Copilot, Grok, and Gemini. The models were asked 121 real financial questions about pensions, taxes, debt, student loans, and savings; each question was repeated five times, and more than 10,000 answers were analyzed in total. On average, models were wrong in 57% of cases. On complex questions, the average error rate rose to 88%, and some models were wrong in 99% of cases. The main failure modes are not knowing current tax rules and inventing “financial rules” that do not exist. Financial Times published a piece about the report.

Context

This failure profile is expected for general-purpose models without grounding in data: LLM knowledge is fixed at the time of training and limited by a knowledge cutoff, while tax rates and credit rules change constantly, so without grounding in official sources the model reproduces outdated or synthesized facts. Invented “financial rules” are a typical hallucination in a high-risk domain, where retrieval only helps partially if the model builds rules on top of or instead of the retrieved text. The “each question five times” scheme additionally measures not only correctness but also answer stability: discrepancies on the same question indicate calibration problems. At the same time, the report’s methodology is not disclosed in open materials, and it was prepared by a fintech vendor rather than independent researchers, so the figures cannot yet be considered a reproduced result. Audience expectations for an “AI advisor” formed before checks on its accuracy appeared: the boundary between “information” and “advice,” whose blurring was warned about by the British regulator FCA in the Mills Review overview, has practically disappeared in user perception.

Why this matters for the industry

The report highlights a regulatory gap: a licensed financial advisor would lose their license for this level of errors, while a language model receives another version update. Positioning a “chatbot as a financial advisor” on a raw model now looks like a defective product pattern, and any offer of “AI gives financial advice” now meets objections with specific numbers. At the same time, a market window opens for a trust layer: grounding in current regulatory data, citing official sources in the interface, notes on the date of currency, guardrails, and a dry-run mode for agentic actions. A practical checklist for ML teams: do not release a raw model for final monetary decisions, build a domain eval on your own questions — the repeated-run methodology is easily reproducible — close time-sensitive facts with retrieval from official sources, and log answers for observability. The stakes are raised by the rollout of agentic payments, which Worldline, ING, and Mastercard launched in June 2026: market participants have to introduce LLM restrictions before hard regulatory rules appear. In the coming months, domain benchmarks will likely become a procurement filter: banks and vendors will demand measurable accuracy on their own questions, not marketing promises.

Why this matters for users

A PensionBee survey of 1,000 adult US users shows the scale of the gap between trust and reliability: 57% are ready to act on a chatbot’s advice without checking, and 23% have already received incorrect financial data from it. The practical takeaway is not to trust chatbots with final monetary decisions: mortgage overpayment amounts, retirement exit strategy, loan parameters. On complex topics like taxes and student loans, almost the entire field of models makes mistakes, and they often do not know current tax rates. Chatbots explain general personal finance principles acceptably, so they can be used as a starting point, but specific figures must be verified against official sources — this is exactly the step that, according to the survey, most people are currently skipping.

What is still unknown / limitations

The report’s methodology is not disclosed in open materials: there is no definition of an error (factual inaccuracy or evasive answer), no description of the judges — humans or LLM-as-judge, no reference answers, and no detailed breakdown by model. Therefore, the figures of 57%, 88%, and 99% can correctly be considered a vendor claim until independent reproduction. The “121 questions, each five times” scheme measures answer self-consistency only when the error definition, judge, and references are fixed; without this, stability and correctness cannot be separated. The report was prepared by a fintech company with its own product interest, not a peer-reviewed research group. It is also unknown which exact model versions and with what settings participated in the test.

Sources

Author

Look at AI, editorial team