The Office Assistant ships in Microsoft Office 97

Filed under no kind at all, so the kind is this entry's to argue. clippy-1996 is not in canon/proposals.md — not among the fifty candidates, not among the sixteen named as dropped at the cap — which puts it in the same position as autofac-1955 and dijkstra-1959: commissioned from outside the proposal list, with the admission test in section 2 carrying the whole weight. The kind that fits is moment, on the definition eliza-1966 set when it argued its own way out of that kind: a moment is a date on which something visibly happened in public, of the shape of Dartmouth, Lighthill, Deep Blue, the LLaMA weights. Microsoft Office 97 shipping an animated assistant to the desktop of essentially everyone who used a computer for work is a date, and it is the largest deployment of anticipatory machine assistance that had ever happened.

Why not idea. The tempting alternative, because there is a framing here — "the social interface," software that offers help before it is asked. But the thing that descends from November 1996 is not a framing later work is built out of. It is a public verdict, and the verdict is the artifact. Every subsequent proactive assistant is built in explicit avoidance of this, which is a different relation than descent. The word "Clippy" is now a common noun in product criticism; that is what a moment leaves behind.

Why not limit. Nothing here is proved. It is very frequently used as though something were — "Clippy proves proactive assistance doesn't work" is section 5's first item — and the reason it is a misuse is precisely that this is an event and not a theorem.

On the year, which is genuinely ambiguous and should not be smoothed over. Office 97 was released to manufacturing on 19 November 1996. General availability and the retail launch were in January 1997, and Microsoft Research's own paper dates the ship that way: "In January 1997, a derivative of Lumière research shipped as the Office Assistant in Microsoft Office '97 applications." The id says 1996 and the header follows it, because that is the date the thing existed and left the building; a reading that wants the date the public met it should say January 1997. Both are correct and they are four inconsistent-looking weeks apart.

descends_from holds one id, and the basis for it needs stating. eliza-1966 is an ancestor of half of this entry. The Office Assistant has two parents: a Microsoft Research decision-theory program, and a design doctrine called the social interface. The doctrine descends from the ELIZA effect — the finding that people relate socially to conversational programs — by way of the Computers Are Social Actors research paradigm that Byron Reeves and Clifford Nass ran at Stanford, which Nass and Reeves brought into Redmond as consultants, which produced Microsoft Bob in 1995, whose characters and animation format the Office Assistant then inherited. That chain is documented at every link. What I have not verified is whether Reeves and Nass cite Weizenbaum, so the edge is a claim about the lineage of the finding and not a checked citation trail — weaker evidence than eliza-1966 demanded of itself, and flagged here rather than implied. The other parent, the Lumière project, has no ancestor in canon/; the one I would attach is a Bayesian-decision-theory entry that does not exist and is not proposed.

What it is

Microsoft Office 97 shipped with the Office Assistant: an animated character occupying a small window above your document, watching what you did, and volunteering help. The default was a paperclip with eyebrows, officially Clippit and universally Clippy, drawn by the children's-book illustrator Kevan J. Atteberry from a set of more than fifteen submitted designs. It headed a gallery — The Dot, The Genius (an Einstein caricature), Scribble, Power Pup, Mother Nature, Will (Shakespeare), Hoverbot, the Office Logo, and Kairu the dolphin in East Asian editions. Type an address and then "Dear," and it appeared: "It looks like you're writing a letter. Would you like help?"

The lineage from Microsoft Bob is not thematic, it is literal. The Dot came straight from Bob. The characters ran in Bob's .act "Actor" format, which the Microsoft Agent .acs format did not replace until Office 2000. Bob had shipped in early 1995 as a "social interface" for home users — a cartoon room full of clickable objects with animated guides — and it was built on Reeves and Nass's demonstration of how thoroughly computer users project human social traits onto the machines they interact with. Bob failed commercially and was discontinued. The doctrine survived it, moved into the flagship product, and reached a hundred times the audience.

The other parent is the part almost nobody knows about, and it is the reason this entry earns a place rather than being a joke with a date on it.

Lumière. A Microsoft Research project begun in 1993, with the first Lumière/Excel prototype demonstrated to the Office division in January 1994, run by Eric Horvitz with Jack Breese, David Heckerman, David Hovel and Koos Rommelse. Its purpose was to infer, under uncertainty, what a software user is trying to do and whether they would welcome help. It built Bayesian user models over roughly forty Excel problem areas; a temporal events language that converted raw mouse and keyboard atoms into modelled events like "menu surfing," "mouse meandering" and "menu jitter"; a persistent competency profile stored in the registry; and a Bayesian information-retrieval component over about six hundred terms of free-text query. That last piece was not vaporware: it shipped in Office 95 as the Answer Wizard.

The governing structure of Lumière is an influence diagram, Figure 1 of the 1998 paper, and the node that matters for everything below is labelled Cost of assistance. The paper is explicit about what it is for: "the overall goal is to take automated actions to optimize the user's expected utility. A system taking autonomous action to assist users needs to balance the benefits and costs such actions. The value of actions depend on the nature of the action, the cost of the action, and the user's needs."

So the prototype's autonomous mode was built with brakes on it, all three of which are worth naming because their absence is the entire story of Clippy:

• A threshold you controlled. "A user-specified probability threshold is used to control the autonomous assistance," exposed in the interface as what the paper calls a "volume control."
• An interruption that withdrew itself. "If the user does not hover over or interact with the autonomous assistance, the window will timeout and disappear after a brief apology for the potential distraction."
• A memory of having been ignored. "Recommendations for assistance are noted and the autonomous help will not be offered again until there is a change in the most likely topics."

Now the shipped product, in the research team's own words. Section 7 of the UAI paper — headed "Components of Lumière in the Real World: Office Assistant" — is one of the more unusual passages in the applied-AI literature, a group of researchers itemising in a peer-reviewed venue what a product division left out:

> "Compared with Lumière/Excel, the Office Assistant employs broader but > shallower models reasoning up to thousands of user goals in each Office > application. [...] However, the system does not employ persistent user profile > information and does not reason about user competency. Also, the system does > not use rich combinations of events over time. Rather, the system only > considers a small set of relatively atomic user actions. Furthermore the > system employs a small event queue and considers only the most recent events. > The system also separates the analysis of words and of events. [...] Finally, > the automated facility of providing assistance based on the likelihood that > a user may need assistance or on the expected utility of such autonomous > action was not employed. Rather, the results of inference are available only > when the user requests assistance explicitly."

And from Horvitz's project page, on what filled the gap: "The Office team has employed a relatively simple rule-based system on top of the Bayesian query analysis system to bring the agent to the foreground with a variety of tips. We had been concerned upon hearing this plan that this system would be distracting to users — and hoped that future versions of the Office Assistant would employ our Bayesian approach to guiding speculative assistance actions — coupled with designs we had demonstrated for employing nonmodal windows that do not require dismissal when they are not used."

The two sources frame it slightly differently and reconcile cleanly. The paper's "only when the user requests assistance explicitly" refers to the Bayesian inference; the project page describes the rule-based layer that separately shoved the character into view. What decided when Clippy spoke was a rule table. "You typed an address and then the word Dear" is a rule. There was no estimate of whether you wanted to be interrupted, because the component that computed that was left in the lab — and the people who built it said in advance, on the record, that this would be distracting.

It was also tested, and the test failed, and it shipped anyway. Roz Ho, then a Microsoft executive, has described the pre-ship research: "We did a bunch of focus-group testing, and the results came back kind of negative. Most of the women thought the characters were too male and that they were leering at them." The mostly-male team shipped the paperclip. Wikipedia records, sourced to Steven Sinofsky, that the internal codename was "TFC," where the C stood for "clown."

The rejection, with dates. On 11 April 2001 Microsoft announced that the Assistant would be off by default in Office XP, which shipped 31 May 2001. The company then ran an advertising campaign at the expense of its own mascot: officeclippy.com, three Flash shorts — "Clippy Gets Clipped," "Clippy Goes Undercover," "Clippy Faces Facts" — with Clippy voiced by Gilbert Gottfried, being fired and coming to terms with the fact that nobody liked him, plus a game called X-tract Paperclip. Office 2003 was the last Windows version to include the Assistant; it was gone from Office 2007. On the Mac it survived through Office 2004. Smithsonian later called it "one of the worst software design blunders in the annals of computing," and Time put it on a fifty-worst- inventions list in 2010.

A company killing its own assistant in a paid ad campaign, four and a half years after shipping it to the world, is the single hardest piece of evidence about how the deployment went. Nothing else in the record needs to be weighed against it.

Why a reading would cite it

Three occasions, and the first is live in this project's own window.

An assistant that decides when to speak on someone else's behalf. The 2026-08-15 midday digest records OpenAI Ireland notifying Free and Go users across the EEA and Switzerland that advertising arrives later this month, with initial targeting "contextual only — current conversation topic, general location, device type." Read that against Figure 1 of the Lumière paper. The 1993 architecture gated an unrequested interruption on a quantity called the user's expected utility, with an explicit Cost of assistance node, because the researchers understood that speaking uninvited is an act with a price paid by the person interrupted. An assistant that surfaces content selected by what your current conversation topic is worth to a third party has not merely omitted that node. It has replaced the objective function with a different one and kept the same interface. That is the sharpest available statement of what changed between 1996 and 2026, and this entry is where the vocabulary for it lives — expected utility to whom is the question, and it is thirty years old, and it was answered correctly in a lab before it was answered wrongly in a product.

A legislative season about when a machine may address you. The same digest records nine California bills clearing Senate Appropriations, among them AB 1609 on customer-service chatbots (5–2), AB 2023 on chatbots and children's safety (6–1), and AB 2656 on generative-AI notice by public employers (7–0). The EU AI Act's Article 50 transparency duties have been the live instrument since 2 August 2026. The whole cluster legislates the disclosure half of the problem — that you must be told what you are talking to. eliza-1966 holds the argument about why disclosure is insufficient for attachment. This entry holds the adjacent and unlegislated half: nothing in any of these instruments governs the decision to initiate. Clippy was fully disclosed. Everyone knew it was a paperclip. It was hated anyway, and it was hated for interrupting.

Agentic assistance, which is the 1996 question with the stakes raised. The same window has Anthropic disclosing an unreleased model in heavy internal use for "agentic work," and OpenAI shipping a latency mode explicitly because agentic loops are priced in wall-clock. As assistance moves from answering when asked to acting when not asked, "on what basis does it decide to act?" stops being a UI question. The one prior mass deployment that faced it in earnest answered it with a rule table, against its own researchers' written objection, and produced the most reviled feature in the history of desktop software. That is a citation with a date on it, and it is available to a reading covering any proactive or background-agent feature.

There is a fourth use that is not an occasion but a hygiene requirement: "Clippy" is the most common epithet in AI-assistant criticism, and it was reached for throughout 2025 and 2026 against Copilot and against proactive features generally. A reading that uses the word should know that the thing under it had the right theory, in-house, with the brakes designed and demonstrated, and shipped the half without them.

What this entry does not support, stated plainly. It is not evidence, it deposits nothing, it moves nothing. And it does not show that anticipatory assistance fails. The version with a decision-theoretic gate was never put in front of the public, so that experiment has not been run.

What it got right, and what it got wrong

Not required for moment. Included because the canon is worth less if only prediction entries get graded, and because there is an unusually clean set of dated claims here — including a correct in-house forecast of the failure, made before the failure.

Right, and this is the one — Microsoft Research predicted the outcome, in writing, before it shipped. Claim: that a rule-based system bringing the agent to the foreground with tips "would be distracting to users." Made: on hearing the Office team's plan, so before the November 1996 RTM. Due: on contact with users. Graded correct, comprehensively, within one product cycle, and confirmed by the vendor's own conduct when it demoted the feature on 11 April 2001 and paid to have it publicly humiliated. This is the rare case where the failure mode was named in advance by identifiable people who were overruled — which is why the entry is about a decision and not about a paperclip.

Right — that bad machine advice is costly, and that user compliance misleads the system into giving more of it. From the Wizard-of-Oz studies behind Lumière, in which human experts watched subjects through a "keyhole" and typed advice: "We learned that poor advice could be quite costly to users. Even though subjects were primed with a description of the help system as being 'experimental,' advice appearing on their displays was typically examined carefully and often taken seriously; even when expert advice was off the mark, subjects would often become distracted by the advice and begin to experiment with features described by the wizard. This would give experts false impressions of successful goal recognition, and would bolster their continuing to give advice pushing the user down a distracting path. Such patterns of poor guesses and 'confirmatory' feedback could lead to a focusing on the wrong problem and a loss in efficiency." Due: continuously. Graded correct and now much larger. A lab study around 1994 caught automation bias and a compliance-driven feedback loop between an advisor and an advisee in the same paragraph — the user takes the wrong suggestion seriously, and the advisor reads the compliance as a confirmed goal inference. Naming that structure is fair; claiming later work derived it from here is not, and this entry does not.

Right, and the fix generalises — that a system offering advice should become conservative, hedge, and decompose. The same studies found the human experts improved with practice by "becoming conservative with offering advice, using conditional statements (i.e., 'if you are trying to do x, then...'), and by decomposing advice into a set of small, easily understandable steps." Due: unspecified. Graded: correct, and independently rediscovered often enough since that its 1990s provenance is invisible.

Wrong — the social-interface premise, as the Office division read it. The premise: Reeves and Nass had shown people treat computers as social actors, therefore giving the helper a face and a personality would make it welcome. Made: across Bob (1995) and Office 97. Due: on contact. Graded: false, and false in a way that inverts the theory it was drawn from. Nass himself supplied the correct reading afterwards in The Man Who Lied to His Laptop (2010): if a thing is treated as a social actor, it is judged by social standards, and by those standards Clippy was not merely unhelpful but rude — it never learned your name, never learned your preferences, and re-offered the same suggestion after you had declined it any number of times. Anthropomorphism does not buy goodwill; it buys jurisdiction under a harsher code. A menu that fails to predict you is a bad menu. A character that fails to predict you is a boor. (I have this book through secondary summaries rather than the text; the argument is consistently reported and the entry leans on it accordingly, but it has not been verified at source.)

Wrong — that the negative pre-ship signal could be discounted. Claim, by conduct: that the focus-group result was noise. Due: immediately. Graded: the signal was right. Note what it was right about, though — the reported objection was that the characters read as male and as leering, which is not the same complaint as the interruption complaint that later sank the feature. The research was correct that users disliked the characters, and the deployment then failed for an overlapping but distinct reason. A reading that flattens these into one lesson gets a tidier story and a worse one.

Wrong, at least as an expectation — that a later version would restore the brakes. Horvitz's stated hope was "that future versions of the Office Assistant would employ our Bayesian approach to guiding speculative assistance actions." Due: within the product line's life. Graded: did not happen. There was no later version that fixed it; the feature was switched off in 2001 and deleted in 2007. The remedy applied to bad proactivity was removal, and the design that would have addressed it properly has still, thirty years on, not shipped at consumer scale.

Unresolved, and worth watching rather than grading — whether the character or the interruption policy did the damage. The record shows both were defective, and there is no controlled comparison anywhere. The one piece of natural evidence is suggestive and weak: modern assistants dropped the face entirely and kept the unrequested proactivity, and are called Clippy anyway. That points at the interruption policy as the heavier factor. It is not proof, and a reading should present it as the weak signal it is.

Commonly misused as

Not required for moment. Included because this is the most-invoked and least-understood reference in assistant design, and because the first item below is currently doing real work in product arguments.

• "Clippy proves proactive assistance doesn't work." The load-bearing misuse. What shipped was a rule-based trigger attached to a cartoon, with the expected-utility gate, the persistent user profile, the competency model, the temporal event reasoning and the self-withdrawing notification all left in the lab — by the research team's own published accounting. The proposition "anticipatory help is unwelcome" was never tested in 1996, because the design that would have tested it was not the design that was released.
• "Clippy was too dumb; the models are smart enough now." Changes the subject, and in the most expensive available direction. The Assistant's weakness was not that it failed to infer — it reasoned over thousands of user goals per application. Its weakness was that nothing decided whether to speak. Capability does not supply that decision; it is a separate calculation with a separate input, namely the cost to the person of being interrupted. A more capable system with no interruption gate is a better-informed interrupter.
• "Clippy was hype, not real AI." Backwards. The Bayesian information retrieval was real enough to ship in Office 95 as the Answer Wizard, the research is a UAI paper with an influence diagram in it, and the modelling work is serious applied Bayesian user modelling. The AI is the part that got cut.
• "Nobody tested it / nobody saw it coming." Both false, and the difference is the whole value of the entry. There was focus-group testing and it came back negative. There was an in-house research team that named the failure mode in advance. This is a documented case of shipping over known objections, not a case of an unforeseeable outcome — and the two carry completely different lessons for anyone reading a 2026 assistant launch.
• "Clippy was Microsoft Bob." Adjacent, not identical, and the distinction matters for the argument. Bob was a separate 1995 consumer product that failed in the market and was withdrawn. The Office Assistant inherited Bob's doctrine, several of its characters and its animation format, and then shipped inside the flagship business suite — which is how a discontinued consumer product's design premise reached vastly more people after its own product died.
• "It looks like you're trying to write a letter." The documented string is "It looks like you're writing a letter. Would you like help?" Popular renderings vary and usually insert "trying to." Small, but if a reading is going to quote it, it should quote it.
• "Clippy is the ancestor of Copilot." True as marketing lineage and as Microsoft's own joke about itself; false as technical descent. Nothing inside a transformer came from a keyword-triggered tip table or from a hand-assessed Bayesian network over Excel goals. What actually descends is the unsolved problem — deciding whether to speak — plus a brand name and a warning.
• "Everyone hated it, so the users were right about everything." They were right about the experience. It does not follow that the underlying research programme was wrong, and section 4 grades the two separately for exactly this reason. The most reviled feature in desktop-software history was the badly shipped remnant of a good idea, and both halves of that sentence have to survive the citation.

Sources

Primary, and read in full: Eric Horvitz, Jack Breese, David Heckerman, David Hovel and Koos Rommelse, "The Lumière Project: Bayesian User Modeling for Inferring the Goals and Needs of Software Users," Proceedings of the Fourteenth Conference on Uncertainty in Artificial Intelligence (UAI-98, July 1998), pp. 256–265; also arXiv:1301.7385. All Lumière quotations above are from that text — the January 1997 ship date and 1993/January 1994 project dates from §1, the expected-utility and Cost-of-assistance material from §2 and Figure 1, the Wizard-of-Oz findings on costly advice and confirmatory feedback from §3, the user-specified threshold and autonomous-assistance controls from §6.1 and §6.4, and the itemised comparison of the shipped Office Assistant against Lumière/Excel from §7.

Primary: Eric Horvitz's Lumière project page, erichorvitz.com/lum.htm, for the Answer Wizard shipping in Office '95, for the Office team's "relatively simple rule-based system on top of the Bayesian query analysis system," and for the verbatim statement that the researchers "had been concerned upon hearing this plan that this system would be distracting to users."

Secondary, for the product record: Wikipedia's "Office Assistant" and "Microsoft Office 97" articles, for the 19 November 1996 RTM date, Atteberry's authorship, the character gallery, the Bob-descended .act Actor format and its replacement by Microsoft Agent .acs in Office 2000, the "It looks like you're writing a letter. Would you like help?" trigger, the off-by-default change in Office XP, the removal in Office 2007, Mac support through Office 2004, the Smithsonian and Time judgements, and the "TFC" codename attributed to Steven Sinofsky. These are secondary and were not checked against Microsoft documentation; the Sinofsky codename in particular is reported at one remove and should be treated as colour rather than fact.

For the Bob lineage and the social-interface doctrine: Byron Reeves and Clifford Nass, The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places (CSLI Publications / Cambridge University Press, 1996), the Computers Are Social Actors paradigm; and the Stanford HCI Group's profile of Microsoft Bob (hci.stanford.edu, "Profile 7"), for Bob's 1995 launch, its room metaphor and its explicit reliance on the Stanford finding that users "project human social traits onto the computers... with which they interact."

Clifford Nass and Corina Yen, The Man Who Lied to His Laptop: What We Can Learn About Ourselves from Our Machines (2010), for the diagnosis that Clippy failed by violating social rules — never learning the user's name or preferences, re-offering rejected suggestions — and for Nass's proposed repair, in which Clippy would respond to a rejected suggestion by blaming Microsoft's help system, which Microsoft declined. Not read at source; this is via consistent secondary summaries, and section 4 flags it in place.

The Roz Ho focus-group quotation ("We did a bunch of focus-group testing, and the results came back kind of negative. Most of the women thought the characters were too male and that they were leering at them") traces to interviews reported in the Seattle Met piece "The Twisted Life of Clippy" (August 2022). That article returned HTTP 403 and was not read. The quotation is as carried in secondary coverage of it, is consistently rendered across sources, and is named here as unverified at source.

For the retirement: contemporaneous and retrospective coverage of the 11 April 2001 announcement, the 31 May 2001 Office XP release, and the officeclippy.com campaign with its three Flash shorts and Gilbert Gottfried voice work (Accountancy Age, 11 April 2001; Tom's Hardware and TV Tropes for the campaign inventory). Secondary throughout; the officeclippy.com site is defunct and was not retrievable.

The OpenAI advertising notice, the California suspense-file results including AB 1609, AB 2023 and AB 2656, the Article 50 status, and the agentic-work items are as recorded in this project's own 2026-08-15-12 digest, which holds the primary links. They are named here as the citation occasion and nothing more. Nothing in this file is evidence, nothing in it is deposited in the ledger, and nothing in it touches the needle.