BBC and the European Broadcasting Union have found that 45% of news queries submitted to four leading generative AI assistants returned erroneous answers, according to research published this month on the BBC media center site.

The joint study examined responses from OpenAI's ChatGPT, Microsoft Copilot, Google Gemini, and Perplexity. Researchers characterized the systems as "dangerously self-confident," a label that should trouble HR leaders deploying these same tools for policy research, compliance lookups, and employee-facing chat.

What did the study actually measure?

The research, summarized this month by industry analyst Josh Bersin, tested how the four assistants handled factual news content. Across the platforms, roughly 45 of every 100 queries produced information the researchers judged incorrect. Publicly available summaries do not detail sample size, query categories, or the evaluation rubric.

The figure lands squarely inside HR's technology stack. Recruiters drafting job descriptions, benefits teams verifying statutory citations, and learning managers sourcing regulatory guidance lean on these assistants daily. A 45% error rate on news questions is a useful proxy for risk on policy questions, though not a perfect one.

Takeaway: A public benchmark exists now, and HR leaders can use it when negotiating with vendors.

Why should HR treat 45% as a workplace problem?

Most enterprise AI rollouts in HR begin with assistants used for drafting and research. A compliance officer asking a model for an FMLA deadline, an EEO filing window, or a state-specific leave rule risks receiving a confidently stated wrong number. The risk compounds when the assistant paraphrases statutes it has not read accurately.

HR teams also field employee questions through chatbots built on the same engines. A benefits inquiry that returns an inaccurate eligibility rule is not a technical glitch. It is a liability event that can drag into ERISA, ACA, or wage-and-hour exposure.

Takeaway: Vendor accuracy claims need HR-specific testing, not general benchmarks.

What does the study leave out?

The published summary does not name the model versions tested. It does not disclose whether researchers separated fact-retrieval failures from reasoning errors, a distinction that matters for any audit. It also does not separate factual mistakes from confabulations, the category of invented detail that concerns employment lawyers most.

The 45% figure describes news queries, not workplace policy. HR teams should hesitate before transposing the number onto their own use cases without internal testing. The absence of a vendor rebuttal to date is itself a data point.

Takeaway: No vendor has yet published a counter-benchmark for HR-domain queries.

What can HR do this quarter?

Procurement teams that cite vendor accuracy in business cases now have an external reference. HR leaders evaluating assistants should press vendors for published error rates on their own domain, not on news samples. Three practical steps follow:

• Require vendors to disclose model versions, training cutoffs, and evaluation rubrics
• Build human verification into any workflow that cites statutes, regulations, or case law
• Pilot internal accuracy tests before extending assistants to employee-facing channels

Takeaway: Treat the 45% figure as a floor for risk planning, not a ceiling.

What comes next?

The BBC and EBU have signaled further research is in development. HR teams should follow that work, particularly any segment that examines workplace policy content. Vendors are likely to contest the methodology, and a credible rebuttal would shift the calculation.

Until such evidence arrives, HR's prudent course is to assume the same error range applies to internal use and to design guardrails accordingly. The cost of those guardrails is small. The cost of treating an assistant as a verified source is not.