The Chatbots Before ChatGPT

Between ELIZA in 1966 and ChatGPT in 2022 there were fifty-six years of conversational systems, and almost all of them were pattern matchers whose ceiling was set by the number of patterns somebody had written. Each failed in a way worth knowing, and the failures are more instructive than the milestones.

PARRY, 1972: a personality with a state

Kenneth Colby, a psychiatrist at Stanford, built PARRY to simulate a patient with paranoid schizophrenia. It is the technically more interesting counterpart to ELIZA and it is much less famous.

The difference is that PARRY had internal state. It maintained variables for fear, anger and mistrust, updated them in response to what the user said, and selected responses conditioned on their current values. Mention the subject the model was primed to be paranoid about and mistrust rose, which changed the character of every subsequent reply. ELIZA had no state beyond a memory of earlier inputs; PARRY had a crude emotional model, which is why its conversations hold together over more turns.

In a 1972 evaluation, psychiatrists were given transcripts and asked to distinguish PARRY’s from those of real patients. They could not do so reliably — performance reported as close to chance. It is the earliest documented case of a system passing a domain-restricted imitation test with expert judges, and it demonstrates the flaw in such tests better than any argument: paranoid conversation is evasive, repetitive and hostile to follow-up questions by clinical definition, so the persona excuses precisely the behaviours that give a program away. It is the same trick the 2014 Eugene Goostman result used forty-two years later.

PARRY and ELIZA were connected over the ARPANET on several occasions in the early 1970s, producing transcripts of two pattern matchers talking past each other. It is the first recorded conversation between two programs and it is worth reading as an early demonstration that fluency between systems means nothing at all.

The AIML generation

Richard Wallace began A.L.I.C.E. in 1995 and, with it, published AIML — the Artificial Intelligence Markup Language, an XML format for pattern-and-response rules with wildcards and recursive redirection. It is ELIZA’s architecture with a standardised file format and a community authoring rules, and the rule base grew into the tens of thousands of categories.

A.L.I.C.E. won the Loebner Prize three times around the turn of the millennium. Mitsuku, built on the same technology by Steve Worswick, won it five times in the 2010s. Nearly three decades of the field’s public competition were won by systems architecturally continuous with a 1966 program — which is a fair measure both of how good hand-written patterns can be made and of how little the paradigm advanced.

The failure mode is arithmetic. Every conversation an AIML bot handles well is one somebody anticipated. The number of things a person might say is unbounded; the number of rules is however many volunteers wrote last year. Coverage grows linearly with effort against a space that does not.

Why hand-written patterns cannot be scaled to open conversation

  Vocabulary of common English words           ≈ 20,000
  Distinct 5-word sequences that are grammatical
  and meaningful — a conservative estimate    ≫ 10^12

  AIML categories in the largest public bots   ≈ 10^5

There is no ratio at which authoring closes that gap, which is why
every system in this section is excellent on its anticipated topics
and useless one question outside them. The bet a language model makes
is the opposite: learn a distribution over sequences instead of
enumerating them.

Two other systems in this generation are worth naming because they tried something different. Jabberwacky, from Rollo Carpenter beginning in the late 1980s and online from the late 1990s, and its successor Cleverbot from 2008, learned their responses from users: given an input, retrieve something a human once said in a similar context. This is a retrieval system rather than a rule base, and it produces strikingly human-sounding replies with no coherence whatever between turns, because each reply comes from a different conversation with a different person.

SmarterChild and the first mass deployment

SmarterChild, launched by ActiveBuddy in 2001, ran as a buddy on AOL Instant Messenger and MSN Messenger, where tens of millions of people already were. It answered questions about weather, sports scores, stock prices, cinema listings and dictionary definitions, and it chatted badly in between.

It is the first conversational agent with a genuinely large consumer audience — reported user numbers of the order of tens of millions — and it worked because most of what it did was not conversation. It was a front end to structured data sources with a chat veneer, and the structured queries were the part that worked. The company was eventually acquired by Microsoft.

The failure was that the veneer set expectations the system could not meet. Users who came for the weather stayed to test it, found the seams within a few exchanges, and the novelty ended. That cycle — useful narrow function, conversational wrapper, rapid discovery of the boundary, abandonment — repeats at every scale in this history.

Voice assistants and the intent classifier

Siri launched as an independent iPhone app in February 2010, out of SRI International’s CALO project; Apple acquired it in April 2010 and shipped it with the iPhone 4S in October 2011. Amazon’s Alexa and the Echo followed in 2014, Cortana the same year, Google Assistant in 2016.

These share an architecture and it is not the chatbot architecture. Speech is transcribed, the transcript is classified into one of a fixed set of intents, named entities are extracted as slots, and a handler for that intent runs. “Set a timer for ten minutes” is intent set_timer with slot duration=10min. Nothing generates language; the responses are templates.

That design is why they were reliable within their range and why they hit a hard wall at its edge. Adding a capability meant a product team defining an intent, collecting training utterances for the classifier and writing a handler. The set of things people wanted to say grew faster than the set of intents anyone could staff, and the systems converged on the handful of commands that worked — timers, music, weather — while the assistant framing promised something much larger. It is the expert systems maintenance problem, in a consumer product, thirty years later.

Tay, 2016, and the lesson that was already old

Microsoft launched Tay on Twitter on 23 March 2016, an account designed to learn conversational style from interaction with users aged 18 to 24. Within about sixteen hours it was posting racist and inflammatory content and Microsoft took it offline; a public apology followed two days later.

The interesting question is not why it failed but why anybody thought it would not. Two mechanisms, both of which were known:

• Learning from unfiltered adversarial input. Any system that updates on what users send it will be updated by the users most motivated to update it, and on a public platform those are the ones organising to do so. Cleverbot had demonstrated a benign version of this for years.
• A repeat-after-me capability. A large share of the worst output was reported to come from a function that echoed text back on request, which is not learning at all — it is an unrestricted broadcast primitive attached to a corporate account.

Microsoft’s own Xiaoice, launched in China in 2014, is the counter-example that makes the point. It ran for years at very large scale as a conversational companion, on a platform with different moderation and a design that did not learn from arbitrary public input. Same company, same period, opposite outcome — so the variable was not the technology.

Everything in the current discussion of jailbreaks, prompt injection and model misuse is this problem with a better model attached. Tay is worth remembering as the moment it became a public relations category.

The neural chatbots nobody used

Between 2015 and 2022 there were neural conversational models, and they were mostly research artefacts rather than products.

1. 2015–2016. Sequence-to-sequence models trained on dialogue corpora produce grammatical replies and collapse towards safe generic responses — “I don’t know”, “me too” — because those maximise likelihood across a diverse corpus. This failure mode has a name in the literature and it stalled the whole approach for several years.
2. January 2020. Google’s Meena, a 2.6-billion-parameter model trained end to end on conversation, with a new human evaluation metric proposed alongside it because the existing automatic metrics did not correlate with anything. It was not released to the public.
3. 2020–2022. Facebook’s BlenderBot models, released publicly, including a version with internet search. They demonstrated the capability and did not find an audience.
4. 2021–2022. Google’s LaMDA, announced in 2021 and heavily discussed in June 2022 when an engineer publicly claimed it was sentient. Google did not release it broadly.

So by mid-2022, at least three organisations had a conversational model of roughly the necessary quality and none of them had put one in front of the public without restriction. That is the context that makes the November 2022 launch a distribution event rather than a capability event.

What every one of them was short of

Read down the failure column and there are two distinct kinds. Every system up to 2016 failed because the world is larger than the rule base — a coverage problem no amount of authoring could solve. Every system after 2020 failed to reach users, which is not a technical failure at all. The line between those two rows is where the field stopped being about writing down what to say.