flash on Nostr: ⚡️🤖 NEW - Researchers have discovered that ChatGPT can be “manipulated” ...

⚡️🤖 NEW - Researchers have discovered that ChatGPT can be “manipulated” using persuasion techniques that are more than 40 years old.

An initial study led by researchers from Wharton and Robert Cialdini himself tested 28,000 conversations with GPT-4o mini, applying the seven principles of persuasion popularized in his 1984 book *Influence*.

When faced with requests that the model was supposed to refuse, these techniques increased its average compliance rate from 33.3% to 72%.

The most striking technique was “commitment”: getting the model to agree to a relatively innocuous initial request before gradually moving on to the problematic request.
In some tests, compliance reached 100%.

Simply changing the way a request was phrased was enough to significantly alter the model’s behavior.

In May 2026, the researchers published a much larger study in PNAS, featuring 126,000 conversations with GPT-5 mini, Claude Haiku 4.5, and Gemini 3 Flash.

Recent models are more resistant, but the effect still exists: when faced with the tested prompts, persuasion increased compliance from 35.3% to 51.3%.