Can AI-written fake LinkedIn profiles pass for real people, and can detectors be retrained to catch them?

This explores how closely GPT-written fake profiles resemble real ones, and what it takes to tell them apart. The collection has only one study that tests this directly, so here is the short answer first: the fakes are close enough to fool tools built for older kinds of fakes, but they are not identical to real profiles. A detector trained on the right examples can still spot them.

The direct evidence comes from a LinkedIn study Can fake profile detectors catch GPT-generated LinkedIn profiles?. Detectors trained on real profiles and hand-made fakes let GPT-written profiles through 42–52% of the time, so to those tools the GPT fakes looked much like real people. Once the detectors were retrained on GPT-written examples, only 1–7% of fakes got through, and real users were not flagged any more often than before. So GPT profiles are similar to real ones, but they carry their own patterns. The patterns are just different from the ones a human faker leaves, which is why older detectors miss them.

People do worse than retrained software. A review of 30 studies found that human accuracy at spotting AI-made text, images and voices generally sits around chance Can people reliably spot content made by AI?. A study of models trained to copy ChatGPT suggests why Can imitating ChatGPT fool evaluators into thinking models improved?. Copying the confident, fluent style was enough to fool human judges, even though the content underneath was no better. A fake profile likely works the same way: it gets the tone and polish of a real profile right, which is what people check, while the statistical tells that a trained detector picks up stay hidden from human eyes. A focus-group study adds that people trust ChatGPT mainly because it is conversational, not because it is accurate Does conversational style actually make AI more trustworthy?. It's the same weakness: we judge by how something reads, not by whether it holds up.

The lesson that carries over is that the realism comes from the style and the fluent surface, not from real content underneath. One demonstration had an AI write 288 complete finance papers with made-up rationales and fake citations Can AI generate hundreds of fake academic papers automatically?. Once convincing surface detail is cheap, fakes can be produced in bulk, and defenses have to learn the generator's own fingerprint. Looking for the clumsiness of a human forger won't work.

What the collection doesn't have is a feature-by-feature comparison, such as word choice, career paths or skill lists in fake versus real profiles. It also doesn't show whether today's stronger models would erase the patterns that retrained detectors still catch.