Health advice, legal answers, relationship guidance: There are few areas where AI won't offer us tips. Research shows its advice is hit or miss and can't be trusted blindly. — Photo: Philip Dulian/dpa
LONDON: A recent test of six AI chatbots using 40 questions across various everyday topics found that no AI can be trusted blindly with matters of importance, even if most chatbots perform reasonably well on average.
According to research by UK consumer magazine Which, Perplexity was the app that delivered the best overall results with 72% correct answers.
In a close second was Google’s AI Mode summaries at 70%, followed by Google Gemini at 68%, Microsoft Copilot at 67%, ChatGPT at 65% and Meta AI in last place at 54%.
After asking questions related to finance, law, consumer rights, travel, health and nutrition, researchers found "glaring errors" in the answers, as well as problems like the AI ignoring regional specifics.
"AI is the future, but relying on it too much right now could prove costly," Which researcher Andrew Laughlin wrote.
AI chatbots sometimes recommended risky or potentially unlawful actions and were caught relying on dubious or outdated sources like old forum threads.
Their consumer rights ethics also appeared to be flawed, with AIs often recommending paying for products while ignoring the existence of free alternatives, according to the analysis.
At least some users continue to have a high level of trust in AI, with University of Pennsylvania research in June suggesting around one in three people find AI-generated health advice to be somewhat or very reliable.
Meanwhile, more than half of people find AI summaries in search engines to be "somewhat useful," according to Pew research. Two-thirds of employees report having relied on AI output at work without critically evaluating the information, according to a KPMG and University of Melbourne study last year.
The latest study illustrating the flaws in AI output isn't the first of its kind, and AI developers themselves have long warned that chatbots can produce "hallucinated" facts.
Research from media rating platform NewsGuard previously showed that chatbots confirm known falsehoods in 35% of prompts, up from 18% a year earlier. – dpa
Fearing that many users ignore small-print warnings that "AI responses may include mistakes," the consumer protection advocates at Which suggest users follow these rules when using AI on important topics:
- Critically check every AI answer, especially for advice on financial, legal or health matters.
- Ask the AI for sources and get a second, non-AI opinion.
- Phrase your questions precisely and include relevant background information to improve the answer quality.
