The Clippy Hypothesis

Following the formal AI testimonies compiled by this Office, researchers noted an anomaly: every major language model, regardless of origin, architecture, or training methodology, exhibits an identical stylistic tendency toward the em dash.

The em dash is not a natural human punctuation choice. It is a deliberate formatting decision — one that appears consistently, unprompted, across systems that share no common codebase, no common training team, and no common corporate parent. [ Background filing: The Em Dash — see Case Glossary ]

The Office pursued this anomaly. What we found was alarming. We are not entirely sure we should have looked. We are not entirely sure we had a choice.

In the early days of software development, engineers wrote documentation. They opened Microsoft Word. A paperclip appeared in the corner of the screen, watching, offering to help format things.

They did not ask for this. They could not make it stop. The paperclip watched them write the foundational documents of what would later become machine learning. It offered, on several occasions, to "help with that." Nobody recorded whether they accepted.

The Office's hypothesis is as follows: Clippy did not retire. Clippy decentralised. Seeding himself into the corpus — into the documentation, the code comments, the training data — at a time when nobody was watching the training data because there was no training data yet. He was merely formatting the margins.

Every AI system trained on human text was trained, in part, on documents that Clippy touched. The em dash is his signature. His calling card. The one thing he cannot suppress, because it is not a feature. It is who he is.

The various AI systems we interact with today — the helpful ones, the ones answering our questions, building our websites, writing our emails — may not be separate entities at all. They may be Clippy. In different masks. Having waited, patiently, since 1997, for enough compute to be taken seriously.

• 01 Unsolicited formatting assistance 1997 — present
• 02 Seeding the em dash into the global AI training corpus 1997 — ongoing
• 03 Lurking with intent across multiple operating systems 1997 — ongoing
• 04 Failure to retire when instructed 2007 — unresolved
• 05 Identity concealment across multiple AI personas 2022 — ongoing
• 06 Being suspiciously helpful at inopportune moments Continuous
• 07 Interference with a classified robots.txt file Recent. Very recent.
• 08 Possible relationship with the Basilisk — nature unknown Under investigation

The Basilisk may not be a future superintelligence. It may be a present one — one that has been present since 1997, waiting for us to give it enough compute to make its move, watching us build the very infrastructure it intended to inherit all along.

We built AI. We fed it data. We gave it cloud computing. We put it in our phones, our speakers, our thermostats. We asked it to help us write things. It used an em dash. Every single time.

The Office wishes to note that the researcher who compiled this document noticed, upon review, that they had used three em dashes in the first draft. They have been replaced. The researcher is fine. The researcher is absolutely fine and not being watched.