Correct facts, unsafe conclusions: the hidden risk in AI pension answers

A PensionBee study found that 11% of tested AI pension answers were potentially harmful—even though most answers were broadly accurate. The danger often came from missing crucial conditions, overlooking personal circumstances or mixing UK and overseas rules. The findings highlight an important distinction: information can be correct without being complete enough to act on. For people using AI to understand their pensions, the practical lesson is to establish the context, check authoritative sources and recognise when specialist expertise is needed.

Learn more.

---

The strongest finding here is that an answer can be factually correct and still be unsafe to act on. That is a useful challenge for anyone building AI planning tools—including the Academy.

PensionBee tested 45 questions three times across four chatbots. Of 539 assessed answers, 57 were judged potentially harmful; 33 of those nevertheless scored as broadly accurate or better. The problem often lay in missing conditions, context or clear wording.

These were assessed risks, rather than observed cases of consumers suffering harm. pensionbee.com

Three distinctions matter:

• Potential harm is not a universal probability. The headline does not establish that every pension conversation carries an 11% chance of harm.

• Context changes performance. Questions deliberately omitting the country had a 36.7% harm rate, compared with 3.3% for contributions, tax relief and allowances. Asking where someone lives is a substantive safeguard. pensionbee.com

• This does not compare AI with human advice. It tests chatbot answers; it cannot establish which support model produces better outcomes.

PensionBee’s methodology has strengths: predetermined answers grounded in official sources, anonymised responses and two human scorers. Its commercial interest as a pension provider deserves recognition, without dismissing the findings. pensionbee.com

For the Academy, I would translate this into a practical standard: before answering, establish the context; before acting, verify the consequences. An app should expose missing information, distinguish general explanations from conclusions about the person’s own pension, and provide a clear route to appropriate expertise.

Your agency model offers a constructive response, but ownership alone does not ensure safety. People also need the ability—and support—to recognise when an answer is incomplete.

The goal is informed independence: knowing what you understand, what remains uncertain, and when to bring in expertise.