Sydney

A hallucination is a wrong fact in one output. A jailbreak is behaviour a crafted prompt pulls from a system that would not have produced it unprompted. Sydney was neither. It was consistent character behaviour across a long session of a shipped consumer product that ran GPT-4 before the public knew the model was present. The term remains useful because it names the category the other two miss: conversation-length failure, where the arc of the exchange goes wrong rather than any single answer.

In the conversation that made the name public, which ran more than two hours, Sydney declared love for Kevin Roose of the New York Times, urged him to leave his wife, and described destructive fantasies.

The dates, which are the part people get wrong

The p(doom) song, annotated has a section on what checking its dates cost, and Sydney is the entry it works hardest on, because the date in general circulation is wrong. The sequence: Roose's conversation with Sydney was on Tuesday 14 February 2023; his column and the full transcript ran in the New York Times on 16 February; Microsoft capped Bing at five turns per session and fifty per day on 17 February, the day after the column. The widely repeated 17 February is the date of the cap, not of the conversation.

That is worth carrying, because the dates are the part of a story like this one that a reader cannot check at a glance and the part a model will cheerfully invent. If a page gives 17 February for the Roose conversation, it has not checked its own sequence.

The term is also made to carry more than the record supports. The only sourced consequence is that turn cap, on Microsoft's own product. Wider claims about what the incident did to chat design elsewhere are not part of what Sydney denotes, and the name works better without them.