Direct answer
AI Town is an open-source, deployable virtual town populated by AI characters that live, chat, socialize, and remember their experiences, built by a16z as a starter kit on top of Convex and a game engine [1]. It is not a commercial game with goals or levels; it is a working sandbox for generative-agent research and for developers who want to build their own agent simulations [2]. You can run it locally or deploy it to the cloud, then watch characters form relationships and spread information on their own [1][2].
What Is AI Town and What Can You Actually Do With It?
AI Town is best understood as a reference implementation rather than a product. The repository describes it as a "deployable starter kit for building and customizing your own virtual town," where AI characters live, chat, and socialize [1]. Because the whole thing is open source and written in TypeScript, you can read every part of the simulation, from the agent loop to the rendering layer, instead of treating it as a black box [1]. That transparency is the point: the project exists so other developers can fork it and change it.
What you can do with it falls into three practical buckets. First, you can observe: deploy the town, let the agents run, and watch conversations, friendships, and gossip emerge without scripting any of it [2]. Second, you can customize: swap the character sprites, rewrite the personalities, change the map, or point the agents at a different language model to see how behavior shifts [1]. Third, you can build on it: the same architecture — a reactive backend plus a vector store for memory — can be repurposed for customer-support simulations, NPCs in a real game, or internal agent testing [2].
It is worth being clear about what AI Town is not. It is not a consumer app you sign into, it does not have win conditions, and it is not a benchmark that produces a score you can cite [2]. Treat it as a laboratory. If you are a student writing about generative agents, it gives you a concrete, inspectable example you can describe accurately rather than vaguely [1][2].
How Does AI Town Work Under the Hood — Memory, Planning, and Agent Architecture?
The design descends directly from the "Generative Agents" research, which proposed giving each character a memory stream: a running natural-language log of everything the agent has done, said, and observed [3]. When an agent needs to act, it retrieves from that stream using three signals — recency, importance, and relevance to the current situation — so a trivial remark from a minute ago can outrank an important event from yesterday, or vice versa [3]. On top of retrieval, agents reflect: they periodically synthesize raw observations into higher-level inferences about themselves and other characters, and those inferences are written back into memory to be retrieved later [3]. A planning layer then turns those inferences into a coarse daily plan that gets broken down into hour-long and then minute-level actions [3].
In the original Smallville experiment, twenty-five agents with this architecture were placed in a small town and simply left alone [3]. Emergent behaviors appeared without being programmed: agents passed along information about a candidate's party, one agent asked another to the event, and the group coordinated arrival times [3]. That result is the reason AI Town exists as a demo — it makes the architecture tangible and reproducible [2][3].
AI Town adapts that design to a modern web stack. Convex handles the reactive database and serverless functions so that every client sees the same world state in real time, while a vector database stores embeddings of agent memories for similarity search [1][2]. The rendering layer is a tile-based game engine, so the town looks like a 2D game even though the interesting work is happening in prompts and database transactions [1]. If you want to modify behavior, the highest-leverage place to start is the agent's prompt and memory schema, not the graphics [1][3].
Does Turnitin Detect AI-Generated Writing, and How Is AI Text Flagged?
Yes — Turnitin produces an AI writing indicator alongside its similarity report, and it works by segmenting a submission and estimating what share of the prose was likely generated by an AI model [4]. The indicator only evaluates long-form prose sentences; it does not score reference lists, bibliographies, or short-form answers, which is why a paper can show a low AI percentage while still containing AI-assisted passages [4]. Importantly, when the detected share falls below Turnitin's confidence threshold, the report displays *% instead of a precise number, signalling a low-confidence result rather than a clean bill of health [4].
The report is explicitly framed as an indicator, not a verdict. Turnitin advises that the score be interpreted alongside an instructor's knowledge of the student's writing, the assignment context, and any drafting history [4]. That matters for anyone reading about AI Town and generative agents: writing about AI is not the same as writing with AI, but prose that leans heavily on generic, model-flavored phrasing can still be flagged even when a human typed every word [4].
Practically, the safest workflow is to check before you submit rather than after. If you have drafted a paper with ChatGPT, Claude, or Gemini — or if you are worried that your own phrasing reads as machine-like — an independent pre-submission check gives you the same kind of AI and similarity report your institution would generate, so you can revise the flagged passages while there is still time [4]. Understanding how segmentation and confidence thresholds work also helps you read a report correctly instead of panicking over a single highlighted paragraph [4].
If the architecture behind AI Town makes you curious about how AI text is actually identified in practice, the more pressing question is usually about your own draft — and that is exactly what turnitin0 was built for. Turnitin0 lets you upload a .docx, .pdf, or .txt file and receive the same two PDFs your professor sees: a Turnitin AI detection report and a similarity report, delivered in under 15 minutes in 98% of cases. Files are checked without being added to Turnitin's student paper database, so your draft stays yours.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
Get [Real Turnitin](https://www.turnitin0.com/) AI & Similarity Report
FAQ
Is AI Town a game?
No. It renders like a 2D game, but it has no objectives, scoring, or progression — it is an open-source sandbox for observing and building generative-agent simulations [1][2].
Do I need to know machine learning to run AI Town?
Not to run it. The repository is a TypeScript project you deploy with standard web tooling; understanding the memory and planning layers helps mainly when you want to change agent behavior [1][3].
Which language models can AI Town use?
The starter kit is model-agnostic in practice — you supply API credentials for the model provider you want, so you can compare how different models behave in the same town [1].
Why does a Turnitin AI report sometimes show *% instead of a number?
That display appears when the detected AI share falls below Turnitin's confidence threshold, meaning the signal is weak and should not be read as either a pass or a fail [4].
Can writing about AI trigger a false positive?
It can. Detection works on prose patterns, not on topic, so heavily templated or generic phrasing may be flagged even when you wrote it yourself — which is why reviewing the segmented report before submitting is useful [4].