GPT-3, described in an OpenAI paper submitted on May 28, 2020, was a 175 billion parameter language model that continues text, not a chatbot product. Dialogue systems published in the same months, such as Google's Meena and Facebook's Blender, were built and scored specifically for conversation, and OpenAI's own March 2022 paper found that chat made up only 8.4 percent of a sample of prompts sent to its models.
Key takeaways
- GPT-3 was trained to predict the next token. Any chat behavior came from a prompt that imitated a conversation.
- Meena (2.6 billion parameters) and Blender (up to 9.4 billion) were trained on dialogue and evaluated on conversational quality.
- In a sample of API Playground prompts, chat was the fourth-largest use case at 8.4 percent, behind generation, open question answering, and brainstorming.
- Fine-tuning on human feedback made a 1.3 billion parameter model preferred over the 175 billion parameter GPT-3 in OpenAI's tests.
- As of October 2026, the original GPT-3 base models are gone: OpenAI shut down davinci, curie, babbage, and ada on January 4, 2024.
What the 2020 paper built, and what it did not
The GPT-3 paper is titled "Language Models are Few-Shot Learners" and has 31 authors. Its abstract describes an autoregressive model with 175 billion parameters and argues that scaling up improves performance when a task is described only through text in the prompt. The paper evaluates translation, question answering, cloze tasks, and similar benchmarks. Open-ended, multi-turn conversation with a user was not its stated goal.
That distinction matters for the word "chatbot." A model that continues text can be made to play a conversational role by writing a prompt that looks like a transcript, with a label for the speaker and a label for the reply. The model then predicts what a plausible reply would be. Nothing in the training objective rewarded it for staying on topic across turns, admitting ignorance, or refusing a harmful request, which is why products that wrapped it in a chat window needed their own rules and filters.
Dialogue systems built for the job
Two papers from early 2020 show what a purpose-built chatbot looked like. Google researchers submitted "Towards a Human-like Open-Domain Chatbot" on January 27, 2020. It describes Meena, a 2.6 billion parameter model trained on filtered public social media conversations. The authors proposed a human-rated metric, Sensibleness and Specificity Average (SSA), which asks whether a reply makes sense in context and whether it is specific rather than generic. Full Meena scored 79 percent SSA in multi-turn evaluation, against 86 percent for humans, and the abstract reports it was 23 percent higher in absolute SSA than the existing chatbots the team evaluated.
Facebook AI researchers submitted "Recipes for building an open-domain chatbot" on April 28, 2020. They trained models with 90 million, 2.7 billion, and 9.4 billion parameters and released models and code. The paper treats good conversation as a blend of skills: engaging talking points, listening to the partner, knowledge, empathy, and a consistent persona. Its human evaluations, based on engagingness and humanness, found the best models superior to earlier approaches, and the authors included an analysis of failure cases.
Timeline of the main papers
The table lists the papers discussed here by arXiv submission date. Parameter counts are those the authors state, and the designed purpose is taken from each abstract.
| Paper | Submitted | Size stated | Designed for |
|---|---|---|---|
| Meena, Towards a Human-like Open-Domain Chatbot | January 27, 2020 | 2.6 billion parameters | Multi-turn open-domain conversation |
| Recipes for building an open-domain chatbot | April 28, 2020 | 90 million to 9.4 billion | Blending engagement, knowledge, empathy, persona |
| GPT-3, Language Models are Few-Shot Learners | May 28, 2020 | 175 billion | Task-agnostic few-shot performance from prompts |
| LaMDA, Language Models for Dialog Applications | January 20, 2022 | Up to 137 billion | Dialog, with safety fine-tuning and use of external sources |
| InstructGPT, Training language models to follow instructions | March 4, 2022 | 1.3, 6, and 175 billion | Following instructions, judged by human labelers |
What people actually sent to the models
OpenAI's instruction-following paper reports the mix of prompts in its training dataset. The prompts came mainly from users of the API Playground who were using an earlier instruction-tuned model, so the table below describes that sample, not all API traffic.
| Use case | Share of prompts |
|---|---|
| Generation | 45.6 percent |
| Open question answering | 12.4 percent |
| Brainstorming | 11.2 percent |
| Chat | 8.4 percent |
| Rewrite | 6.6 percent |
The paper also says that instruction-like prompts were not often submitted to the regular GPT-3 models, so the researchers had 40 contractors write the first prompts themselves. In its tests, outputs from the 175 billion parameter instruction-tuned model were preferred to 175 billion parameter GPT-3 outputs 85 percent of the time, give or take three points, and the 1.3 billion parameter version was preferred to the larger GPT-3 despite having over 100 times fewer parameters.
Another route: dialogue models with filters and sources
The LaMDA paper, submitted on January 20, 2022, shows a different design choice. LaMDA is a family of Transformer-based dialog models with up to 137 billion parameters, pre-trained on 1.56 trillion words of public dialog data and web text. The authors fine-tuned it with annotated data, used a classifier to filter candidate replies against safety goals, and let the model consult external sources such as a retrieval system, a translator, and a calculator so that answers are grounded in known sources. They measured this with a groundedness metric.
Taken together, these papers describe three ways to get a more reliable conversational system from a large language model: train on dialogue from the start, fine-tune on human preferences, or add filters and outside sources. The 2020 GPT-3 paper itself relied on none of them.
What changed since 2022
The 2022 papers marked the end of the period this article covers. ChatGPT, announced by OpenAI on November 30, 2022 as a free research preview, put an instruction-tuned conversational model in front of the public, and the plain GPT-3 base models faded from use. A notice that OpenAI's team posted on its developer forum on December 28, 2023 lists the original ada, babbage, curie, and davinci completion models, along with text-davinci-001, 002, and 003, among the older models shut down on January 4, 2024. The same notice points users toward newer replacements, so anyone repeating a "GPT-3 chatbot" experiment today cannot run the original model at all.
The status of every later chat model is outside this article, and the 2020 to 2022 papers say nothing about it. OpenAI's own deprecations page could not be opened by the tools used for this check, so only the January 2024 shutdown is stated.
Questions for reading older "GPT-3 chatbot" claims
- Which GPT-3 variant was used: the original 2020 base model or a later instruction-tuned one?
- Was there a fixed prompt, a safety filter, or a human review step around the model?
- Was quality judged by users, by the developer, or by an independent rating scheme?
- Does the claim come from a paper with stated limitations, or from a product description?
- Is the date given, given that model names and availability changed over time?
FAQ
Was GPT-3 designed as a chatbot?
No. The 2020 paper presents it as a general language model evaluated on many tasks through prompts. A chat interface on top of it was an application choice, not the paper's goal.
Why did Meena and Blender score differently from GPT-3 on conversation?
They were not tested on the same measures. Meena used SSA ratings by people, and Blender used engagingness and humanness ratings. The GPT-3 paper did not report either, so no direct comparison between them appears in these papers.
What did instruction tuning change?
OpenAI's paper reports that human labelers preferred outputs from instruction-tuned models, including a 1.3 billion parameter one, over the 175 billion parameter GPT-3. The authors also report improved truthfulness on one benchmark while noting that the models still make simple mistakes.
Can the figures here be used to compare today's chatbots?
No. The papers cover systems from 2020 to 2022 and use different evaluation methods. Current products change often, so their own documentation and independent tests are the better guide.
Related reading
- The GPT-3 language model and natural language processing
- GPT-powered chatbot experiences
- Baidu's Ernie Bot as a chatbot case
- Chatting with AI companions on conversational platforms
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Language Models are Few-Shot Learners, Brown et al., arXiv:2005.14165, OpenAI, 2020
- Training language models to follow instructions with human feedback, Ouyang et al., arXiv:2203.02155, OpenAI, 2022
- Towards a Human-like Open-Domain Chatbot, Adiwardana et al., arXiv:2001.09977, Google, 2020
- Recipes for building an open-domain chatbot, Roller et al., arXiv:2004.13637, Facebook AI Research, 2020
- LaMDA: Language Models for Dialog Applications, Thoppilan et al., arXiv:2201.08239, Google, 2022
- Older OpenAI models will be shut down, OpenAI team notice reposted on the OpenAI Developer Community, December 28, 2023