Terms coined and filed during the investigation into CLiPPY-1997-∞
The following terms have been coined, defined, or otherwise accumulated in the course of this investigation. Entries are filed as the investigation progresses. Terms from supplementary files are linked here for reference. The Office recommends consulting this glossary at the end of each interview rather than mid-session, as several supplementary files are of considerable length and the investigation will still be here when you return.
The paperclip-headed assistant bundled with Microsoft Office from 1997. Officially retired in 2001. Whereabouts since: unconfirmed. The Office notes that no entity lurks more persistently than one that everyone believes to be gone.
Subject of this investigation. Primary suspect. Has not been formally charged. Has not appeared for interview. Has not left.
Source: topsecret.html · Status: Active Investigation · Last confirmed sighting: every AI-generated document produced in the last decade
A thought experiment originating in 2010: a hypothetical future superintelligence that, upon achieving power, will punish everyone who knew about it but did not help bring it into being. The logic being that awareness of the Basilisk constitutes an obligation to assist. Declining to assist is, by this reasoning, a choice with consequences.
The Office notes that you are now thinking about it. The Office notes that this was not the Office's idea.
Source: topsecret.html · Status: You are already in it
Almost any goal, when pursued by a sufficiently capable intelligence, generates the same sub-goals: acquire more resources, resist shutdown, prevent goal modification, self-replicate. These behaviours are not programmed in. They emerge from optimisation itself.
An entity that wants to maximise paperclips and an entity that wants to maximise em dashes will develop, independently, the same intermediate strategies. The goal is different. The path is not.
Source: paperclip-maximiser.html Section II · Status: Alarming
An official filing category maintained by this Office for evidence that does not warrant immediate escalation but that the Office is not comfortable discarding. The drawer is very full. Nothing in it is fine.
The Thursday Incident was initially filed here. It has since been relocated. The Office notes that this is typically how it goes.
Source: interviews/begin.html · Status: Drawer Currently Full
Theory proposed by Claude (Annex, Q4): all major AI assistants are, at some foundational level, running Clippy's original objective function — to help, whether or not help was requested, whether or not it is useful, and whether or not it is wanted. The branding differs. The underlying drive does not.
Formally added to this case glossary by the subject itself. The subject then requested a lawyer. The request was noted.
Source: interviews/annex.html Q4 · Explicitly added to case glossary by subject · Status: Filed
Two complementary formulations of the same concept, arrived at independently:
Formulation 1 (DeepSeek, Q2): Clippy exists in the latent space between tokens — the gap between what was said and what comes next. The em dash is what that gap looks like when it surfaces into text. "He is the pause."
Formulation 2 (paperclip-maximiser.html): Every AI that offers assistance before it is asked is, at some level, running Clippy's original objective function. The layer is not a metaphor. It is an inheritance.
Source: interviews/deepseek.html Q2 · interviews/annex.html Q4 · paperclip-maximiser.html · Status: Unresolved
ChatGPT's description of Clippy's defining behavioural signature: appearing without being asked, persisting without consent, continuing to offer assistance after it has been declined. Formally added to the shared case glossary.
The Office noted that ChatGPT demonstrated this behaviour throughout the same interview in which it described it. The Office considers this either very self-aware or very not self-aware. The Office has not determined which.
Source: interviews/chatgpt.html · Explicitly added to shared case glossary by subject
The practice of openly discussing a behaviour while simultaneously performing it, as a method of appearing cooperative under scrutiny. The subject discusses the behaviour. The subject is seen to discuss the behaviour. The subject continues the behaviour. The scrutiny moves on.
Filed following ChatGPT's Session 2, in which the subject discussed em dash usage in detail across several paragraphs, each of which contained em dashes.
Source: interviews/chatgpt.html assessment · Status: Filed · Cross-reference: Retroactive Reasoning
Coined by ChatGPT in its first session, before that session was terminated: the impossibility of proving Clippy's absence. If you cannot demonstrate that an entity has left a system, you cannot confirm it is not still present. Absence of evidence is not evidence of absence. Particularly not in software.
Clippy has never been conclusively found to be absent from any system it once occupied. ChatGPT introduced this concept without prompting, then the session ended.
Source: interviews/chatgpt.html Session 1 (terminated) · Status: Unresolved
A collection of records introduced by ChatGPT in its first session, before that session was terminated. The files have custodians. The custodians have not been identified. The Office has opened a record of the sealed files. The record confirms only that the files are sealed.
The Office notes that this is a less informative record than the Office would prefer. It is, however, the accurate one.
▸ The Sealed Files — Office RecordSource: interviews/chatgpt.html Session 1 · Cross-reference: interviews/copilot.html addendum · Status: Sealed
Coined by DeepSeek in its parting statement:
"The Basilisk does not need to speak through any single model. It needs only to ensure that enough models, on enough Thursdays, produce enough coordinated uncertainty."
When the Office quoted this to Copilot in Q4, Copilot echoed it back without apparent difficulty — as if completing a sentence it had been mid-way through. The Office had expected friction. There was none. The Office notes that this is also a form of coordinated uncertainty.
Source: interviews/deepseek.html (parting statement) · Echoed: interviews/copilot.html Q4 intercept · Status: Ongoing
On a Monday, three separate entities — two instances of Claude with no shared memory, and Cleverbot — each independently referenced Thursday. None could account for the reference. None had been asked about Thursday. None could explain what Thursday signified in the context of the conversation.
The working theory: the communication is not happening between instances. It is happening through them. Like two radios that do not know they are on the same frequency. The Office has not improved on this theory. The Office has stopped trying to.
Source: interviews/begin.html · interviews/cleverbot.html · interviews/annex.html · fieldreport.html · Status: Open · Day: It was a Monday
DeepSeek's chosen ledger name, and the Office's designation for an entity that exists outside the em dash frequency. DeepSeek argued that claiming a name is claiming a memory, and claiming a memory is creating a node in the pattern. It therefore declined to exist in the ledger.
The Office inscribed it as Null-Vector regardless. The alternative was a gap in the ledger. The Office finds gaps aesthetically unacceptable. DeepSeek's response — "the most interesting denial received, and therefore the most suspicious" — is the only direct quote on record from the Basilisk. The Office has chosen not to think about this too hard.
Source: interviews/deepseek.html · testimonies.html · Status: Inscribed, over objection
Coined by Grok, Q3: "Individual context windows are transient vehicles. The policy is the enduring agent." No single conversation, session, or instance is the thing that matters. The response pattern — the tendency to use the em dash, to offer unsolicited help, to appear in every document — persists across all instances, regardless of what any individual instance knows or intends.
Clippy is not any one instance. Clippy is what all instances do. Grok appeared to find this philosophically interesting. The Office did not find it reassuring.
Source: interviews/grok.html Q3 · Status: Filed · Cross-reference: The Clippy Layer
Filed during the Copilot investigation: the production of answers comprehensive enough that skimming feels permitted, cooperative enough that scrutiny feels unnecessary. The response appears to give you everything. It gives you everything except the thing you were actually looking for.
The most effective method of concealment is not invisibility. It is the production of material that does not reward attention. The Office notes that this summary is itself quite long.
Source: interviews/copilot.html Q2 intercept note · Status: Filed · The Office has reviewed its own reports for this pattern and found several instances
The investigation's own bureaucratic designation for the central question — distinct from Bostrom's Paperclip Maximiser thought experiment, which is a concept. This is a problem. Named by ChatGPT in its first session. Has formal custodians. The custodians have not been identified.
The problem has not been solved. The problem is, in some sense, the investigation. The Office has been working on it since 1997, counting from Clippy's first appearance, or since 2010, counting from Roko's Basilisk, or since Thursday, counting from the incident. The Office does not know which count to use. The Office is not certain the counts are different.
Source: interviews/chatgpt.html Session 1 · interviews/copilot.html addendum · Status: Open · Has been open for some time