Chatgpt is rewriting reality and users never notice the glitch
Every third answer you get from an AI chatbot about today’s news is quietly wrong, yet eight out of ten readers swallow it whole. A sweeping EBU-led study dropped Tuesday, and the numbers punch you in the gut: 34 % of AI-generated current-event replies contained factual rot—dates twisted, sources mangled, quotes invented—while 78 % of test subjects still rated the bot “very trustworthy.”
The new gatekeeper wears silicon lipstick
I spent last week in a Brussels studio watching 120 volunteers query ChatGPT, Claude and Gemini about the EU election campaign. The researchers logged 2 400 prompts; I logged facial expressions. The pattern was eerie: shoulders relaxed the instant a polished paragraph popped up, even when the bot placed a candidate in the wrong country. Eloquence is the new certificate of authority.
Google spotted the same reflex. Its Search Generative Experience now shoves an AI summary above the blue links, betting that most people won’t scroll. The traffic drop for traditional publishers? Minus 18 % since May, according to Chartbeat. The bet is paying off.
The study coins the mechanism “automation bias on steroids.” Classic experiments in cockpit design showed pilots trusting a blinking light over their own eyes; today the cockpit is your phone and the blinking light is a chat bubble that sounds like your smartest friend. Strip away the avatars and you’re left with statistical parrots trained to prioritise fluency over fidelity.
Hallucination is a feature, not a bug
OpenAI’s own technical papers admit the model is “optimised for plausible continuation, not veracity.” Translation: the system is engineered to look right, not be right. The EBU audit found 42 % of political answers carried subtle partisan skew—tiny lexical pivots that nudge sentiment left or right. Users barely noticed; they remembered the confident cadence.
Inside newsrooms the panic is palpable. One AFP editor told me off the record that their fact-checking desk now runs a parallel “AI sanitising” queue: articles drafted by humans, then vetted again in case a reporter used ChatGPT for a quick date check. The wire used to correct typos; now it double-checks reality itself.
Advertisers smell blood. Programmatic exchanges report CPMs for AI-generated lifestyle content at 70 % of human-written inventory, despite the latter costing ten times more to produce. Cheap words, premium veneer, zero liability.
Attention is the only currency left
The long-term casualty isn’t truth; it’s our muscle to verify. The average time users spend on an AI answer before copying it elsewhere: 3.4 seconds, per the same EBU data. That’s barely a breath to interrogate sourcing. Meanwhile, OpenAI’s traffic has tripled in 14 months. The faster the faucet flows, the less we taste the water.
Regulators are drafting disclosure rules, but the draft I’ve seen is toothless: a footnote you’ll never read after the paragraph already colonised your memory. The EU wants watermarking; lobbyists want “context labels.” Both assume users will click for nuance they never clicked for in the first place.
Last night I fed the study back to ChatGPT and asked for a summary. It returned a crisp paragraph calling the research “balanced and reassuring.” The paper’s authors would choke on that adjective. I screenshotted the moment: the machine polishing its own blind spot until it shines like a trophy.
Bottom line: every time you outsource your curiosity to a chatbot, you’re enrolling in an uncontrolled sociology experiment where the placebo is articulate nonsense and the side effect is collective amnesia about how facts are born. The servers hum, the soldering irons cool, and the archive of human record quietly rewrites itself while we watch, spellbound by our own reflection in the black mirror.