Study reveals users blindly trust chatgpt even when it hallucinates
ChatGPT lies, users nod. A sweeping pan-European study dropped last week showing that millions treat OpenAI’s chatbot like a digital oracle, swallowing fabricated election dates, phantom health alerts and nonexistent laws with the same calm they once reserved for Google’s top blue link. The kicker: the fancier the prose, the deeper the credulity.
The hallucination numbers that should freeze every newsroom
Researchers at the European Broadcasting Union fed 3,700 breaking-news prompts to ChatGPT, Google’s Bard and Microsoft’s Bing in 24 languages. Roughly four in ten answers carried factual errors; 14 % invented entire events. Yet 72 % of the 12,800 survey participants rated the bot’s replies “very reliable.” When debriefed, most shrugged: “It sounded right.”
The finding is a cold shower for publishers who already see referral traffic from Google search eroding. If audiences now outsource even headline scanning to a bot that invents sources on the fly, the supply chain of reality itself wobbles. Loic Damour, lead author at France Télévisions, put it bluntly: “We are watching the birth of a self-soothing machine that edits the world to fit the user’s comfort zone.”
Why eloquence beats accuracy inside the human brain
Psychologists call it automation bias: the same glitch that once made airline pilots trust a faulty altimeter now makes TikTok teens repeat ChatGPT’s claim that the Eiffel Tower moved to Berlin last summer. The bot’s tone—measured, confident, sprinkled with fake citations—triggers a shortcut in our cognition. We surrender scrutiny to save calories.
Tech firms profit from that surrender. Every confident answer keeps users inside the chat window longer, feeding ad profiles and subscription tiers. Google, once obsessed with “ten blue links,” now touts Bard atop search results even while its own disclaimers warn the bot “may give inaccurate information.” The contradiction is not lost on media lawyers: if platforms present machine prose as an answer rather than a guess, liability questions pile up.
The quiet redistribution of civic memory
Inside newsrooms, editors see the shift in real time. During Spain’s snap election in July, regional outlet La Voz de Galicia caught ChatGPT telling voters that ballots could be cast online—a right that does not exist. The fabrication spread faster than the paper’s correction. By the time the story was debunked, WhatsApp groups had already archived the lie as fact.
Across the Atlantic, a similar pattern played out around Kansas’s abortion referendum. Volunteers who relied on ChatGPT to craft canvassing scripts unknowingly repeated polling-place rules that expired in 2020. The error rate mattered less than the saved minutes. Campaigns optimized for speed, not truth.
Regulators already late to the spectacle
Brussels’ AI Act, set to take effect in 2026, demands “risk management” for generative systems but leaves consumer expectations untouched. The loophole is glaring: no clause forces a chatbot to reveal when it is guessing. Until then, literacy campaigns fall to librarians and aging fact-checking nonprofits armed with browser plug-ins no teenager installs.
Meanwhile, OpenAI quietly updated its fine-print last month, removing the promise to “reduce hallucinations by 90 % within two years.” The commitment vanished without press release—an edit few users noticed, because users rarely scroll past the first paragraph. The company now markets “custom instructions” that let paying subscribers order the bot to “always sound certain.” Demand is soaring.
The EBU team refuses to moralize; they simply ran the math. If current growth curves hold, by 2025 more Europeans will read AI-generated news summaries than front pages. Each point of market share shifts ad euros away from newsrooms that still pay stringers to verify courthouse spelling. The casualty is not just jobs; it is the communal record of what actually happened yesterday.
There is no killer app for skepticism. The study closes with a single, stark histogram: the steepest line tracks not error frequency but user delight. The smoother the lie, the louder the applause. Until that curve bends, the bot will keep talking—and the world will keep nodding along.