Two ways to chat with this model. Try one of the curated examples under the chat box (nothing needed), or unlock freeform chat with a museum access code or your own API key.
AI Time Machine is a working museum of conversational AI. You get one familiar chat interface, with a model picker that goes back to 1966. Ask ELIZA for advice, watch GPT-2 lose the thread mid-sentence, get emoji-bombed by GPT-4o, then put the same question to a current frontier model and compare.
How each answer is produced
Every model in the timeline answers your messages. How the answer is made depends on whether that system still exists. The rule is simple: if a model exists today, we call the real thing. If it doesn't, we recreate it as faithfully as the historical record allows. Each reply carries a badge showing how it was made (live, live recreation, or curated) and its curator note spells out the provenance.
The real algorithms (1966–2008)
ELIZA, PARRY, RACTER, Dr. Sbaitso, A.L.I.C.E., SmarterChild and Cleverbot were never neural networks. They were programs, so they can be genuinely re-implemented. Our ELIZA runs Weizenbaum's keyword decomposition with pronoun reflection: say "can you help me" and the can-you-X rule really fires ("YOU BELIEVE I CAN HELP YOU, DON'T YOU?"). PARRY carries Colby's internal mistrust state, opening guarded and turning hostile as you probe. RACTER fills grammatical templates with random words and weaves your own words back in. These are the same class of machine as the originals.
Toy statistical mimics (GPT-1 and GPT-2)
The early transformers are gone from every API, so we mimic the mechanism. GPT-1 renders as a raw completion engine that continues your text instead of answering it. GPT-2 is a small Markov chain that starts fluent and collapses into repetition, which was its real, documented failure mode.
Era-faithful impersonation (retired LLMs)
GPT-3 through Claude 3.x can no longer be called, because providers retire old models. A current frontier model (Claude Opus 4.8) acts them under a strict per-model brief covering capability ceiling, verbosity, formatting habits, factual reliability, signature tics and knowledge cutoff. This is how GPT-4.5 and GPT-5 come out differently: 4.5's brief encodes its documented warmth, GPT-5's encodes its router-driven brevity. Each impersonated reply carries a curator's annotation identifying which documented quirks appear in that answer. Without an API key you browse curated exchanges for these models instead, recorded through the same pipeline.
The real thing (current models)
Models still served by their vendors answer as themselves, marked with a green ● badge. Current Claude models go through the Anthropic API, and still-served GPTs (GPT-4o, o3, GPT-5.x) through the OpenAI API, directly from your browser with your own key. Nothing is proxied through a server of ours, because there isn't one. The whole museum is a static page.
Sources
The briefs that distinguish, say, GPT-4.5 from GPT-5 are curated summaries of the public record. They are not vendor-endorsed. Where a trait is community consensus rather than documented fact, the curator notes try to say so.
Joseph Weizenbaum, "ELIZA — A Computer Program for the Study of Natural Language Communication Between Man and Machine"Communications of the ACM, 1966. The decomposition rules, reassembly templates and pronoun reflection implemented here come from this paper.
Kenneth Colby et al., PARRY publications and transcriptsArtificial Intelligence, 1971–72, including the test in which psychiatrists could not reliably tell PARRY's transcripts from real patients'. Source of the anger, fear and mistrust affect model.
William Chamberlain & Thomas Etter, The Policeman's Beard is Half Constructed (1984)The book "authored" by RACTER, plus surviving MS-DOS sessions of the Mindscape release, the basis for its template grammar and word-weaving.
Richard Wallace, the AIML specification and A.L.I.C.E. rule sets1995 onward, with Loebner Prize records (2000, 2001, 2004).
Contemporaneous coverage of SmarterChild, Dr. Sbaitso and CleverbotPlus archived session logs. These are the source for SmarterChild's intent routing and Sbaitso's PARITY ERROR meltdown.
OpenAI and Anthropic release announcements, model cards and system cardsGPT-1 paper (2018) through GPT-5.x; Claude 1 (2023) through the current generation. Capability claims, context windows, knowledge cutoffs and release dates come from these.
Well-documented community observationsGPT-4 Turbo's laziness complaints; the GPT-4o sycophancy incident and rollback; GPT-4.5's reception and early retirement; the GPT-5 launch backlash that temporarily restored 4o; Claude 2.1-era over-refusals; Claude 3 Opus's purple prose.
How we test it
An automated harness (eval.js, in the repo) extracts the engines from the shipped page, so the tests run against production code and cannot drift from what visitors see. It has three layers:
Engine correctness. Deterministic checks that documented behaviors fire: ELIZA's "can you" decomposition, PARRY's slide from guarded to hostile, SmarterChild's weather intent, GPT-2's repetition loops.
Historical ground truth. Assertions written from primary sources, independently of the implementation, so they fail when the code faithfully does the wrong thing. This layer exists because early versions shipped confident inaccuracies, like RACTER shouting in caps, that only a source-grounded test could catch.
An LLM judge. A frontier model scores recreations from 0 to 10 against per-model historical briefs, looking for anachronisms: vocabulary that postdates the model, capability beyond the era's ceiling, missing signature tics. The judge has caught real bugs the deterministic suite missed.
The curated keyless exchanges are generated through the same pipeline as live answers and covered by the same suite.
FAQ
How did you model models that no longer exist, like GPT-4.5?
From the public record (see Sources), condensed into a per-model brief that a current frontier model performs under. The brief pins capability ceiling, verbosity, formatting habits, reliability and knowledge cutoff. The curator note on each reply points at which documented quirks appeared in that specific answer, so you can audit the performance.
Why do some models need an API key?
The pre-2019 systems are real algorithms running in your browser. Free, no key. Everything newer needs a large language model, and this site deliberately has no backend to pay for one. Add your own Anthropic or OpenAI key in Settings: calls go directly from your browser to the vendor, and the key never leaves your machine except to that vendor. Without a key you can still browse curated exchanges for every model.
Is my chat stored anywhere?
No. There is no server, no database and no analytics. Conversations live in your browser tab and keys live in your browser's localStorage. See Privacy.
How accurate are the recreations?
As accurate as the public record allows, enforced by the eval suite. They are still recreations, and the curator notes are editorial. Treat everything as an educational exhibit rather than a primary source. Where the museum's research record ends (the late-2025 GPT point releases), the model cards say so plainly.
Why does the timeline compress 1966–2018?
Because the interesting density is recent. More consequential models shipped between 2023 and 2026 than in the previous half-century. The timeline uses a piecewise scale, five decades compressed on the left and the modern era expanded on the right, so both stay readable.
AI Time Machine is a hobbyist, educational project by an independent developer. It is not affiliated with or endorsed by Anthropic, OpenAI, Microsoft, Creative Labs, or any other company or model creator mentioned. All trademarks belong to their owners. The code is open source on GitHub.
AI Time Machine (the "service") is a free, hobbyist, educational project built and maintained by an independent developer (the "developer"). By using the service you agree to these terms. If you do not agree, please do not use the service.
1. What the service is
The service is an interactive exhibit about the history of conversational AI, provided free of charge for education and entertainment. It is a prototype built by a hobbyist — not a commercial product, and not a professional, advisory, or archival service of any kind.
2. Accuracy — none warranted
Historical systems are presented as recreations and approximations. They are made as faithful as the public record reasonably allows, but they are not the original software, and their outputs, the curated example exchanges, the curator notes, the model descriptions, and the timeline are all editorial content that may contain errors, omissions, or simplifications. Outputs from live AI models called through the service may likewise be inaccurate. Nothing produced by the service is advice of any kind, and you must not rely on it for any decision.
3. Your API keys and third-party services
Some features only work if you supply your own API keys (for example from Anthropic or OpenAI). If you do:
your keys are stored only in your own browser and requests are sent directly from your browser to the relevant provider — the developer never receives, stores, or transmits your keys or your conversations;
you are solely responsible for safeguarding your keys, for all use made with them, and for all charges billed to you by your API provider;
your use of those providers is governed by their own terms and policies, which you are responsible for complying with.
4. Acceptable use
You agree not to misuse the service — including attempting to disrupt it, using it to violate any law or third-party right, or using it to violate the usage policies of any API provider whose key you connect.
5. No affiliation
The service is not affiliated with, endorsed by, or connected to Anthropic, OpenAI, Microsoft, Creative Labs, or any other company, model creator, or rights holder mentioned in the exhibit. All product names and trademarks belong to their owners and are used solely for historical identification and commentary.
6. No warranty
The service is provided "as is" and "as available", without warranties of any kind, express or implied, including merchantability, fitness for a particular purpose, accuracy, availability, and non-infringement. The service may change, break, or be withdrawn at any time without notice.
7. Limitation of liability
To the maximum extent permitted by law, the developer's total aggregate liability arising out of or in connection with the service, on any legal basis, is limited to the lower of (a) the amount you have paid the developer for the service and (b) US $1. The service is free, so the amount you have paid the developer is expected to be zero. To the same maximum extent, the developer is not liable for any indirect, incidental, special, consequential, or exemplary loss, loss of profits, data, or goodwill, or for any charges incurred with third-party API providers. Nothing in these terms excludes liability that cannot lawfully be excluded.
8. Open source
The service's source code may be published under an open-source licence. That licence governs the code; these terms govern your use of the hosted service.
9. Changes and general
The developer may update these terms at any time by posting a revised version on this page; continued use after a change is acceptance of it. If any provision of these terms is held unenforceable, the remainder stays in effect. These terms are the entire agreement between you and the developer regarding the service.
The short version: this site collects no personal data. There are no accounts, no database, no analytics and no tracking. The site is a static page plus one small relay function used only for freeform chat on the museum's model budget.
Your conversations
Chats exist only in your browser while the tab is open. If you use an access code for freeform chat, your messages pass through the museum's relay function on their way to Anthropic; the relay stores and logs nothing. The relay sets one cookie: a signed counter of the day's usage, which contains no personal information.
Your API keys
If you add an Anthropic or OpenAI API key in Settings, it is stored in your browser's localStorage on your device and used to make requests directly from your browser to that provider. Your keys and messages travel only between your browser and Anthropic or OpenAI; those requests are governed by the respective provider's privacy policy. Removing your keys in Settings (or clearing site data) deletes them from your device.
Hosting
The static files are served by Vercel. Like any web host, Vercel may keep standard, short-lived infrastructure logs (such as IP addresses of requests) to operate the platform; the developer does not access or use these. See Vercel's privacy policy for details.
Local storage used
The site stores a few non-personal preferences in your browser's localStorage: your API keys and access code (if you add them) and a flag remembering that you've seen the welcome screen. Nothing else.
Questions
The project is open source — you can verify all of the above by reading the code on GitHub.
Sixty years of AI — click a bar or a card to travel
Welcome to the AI Time Machine!
60 years of AI in one chat box
In this interactive digital experience you'll explore AI models from 1966 to the present day.
Exhibit
Settings
Bring your own API keys
The pre-2018 exhibits run their original algorithms in this page, free, no key. Every LLM-era model needs a key. Add either key, or both; each works on its own. Keys are stored only in this browser's localStorage and are sent only to api.anthropic.com / api.openai.com, directly from your browser. This site has no server. You are responsible for your provider's usage charges.
Anthropic key unlocks live answers from current Claude models (marked ●), era-faithful impersonation of every retired model, and AI curator analysis of each exchange. Get one at console.anthropic.com →
OpenAI key lets still-served GPT models (GPT-4o, o3, GPT-5.x) answer as themselves, falling back to recreation if the browser call is blocked. Get one at platform.openai.com →
Museum access code unlocks freeform chat on the museum's own model budget (a small daily allowance), no API key needed. Codes are handed out by the developer while the museum is in preview.
Authenticity: pre-2018 bots run faithful re-implementations of their original algorithms, live in this page — they really were just rules. GPT-1/2 are toy statistical mimics of the real failure modes. Retired models (GPT-3 → Claude 3.x…) no longer exist on any API, so their replies are recreations of each release's documented behavior. Without keys you can browse curated exchanges for every model. How it all works →