Little Life Moths Commons · learn by poking the mechanism

AI Literacy Lab

No corporate mystique and no “AI is magic” fog. Flip switches, break assumptions, and watch what changes. Nothing here calls an AI model.

Privacy: this page runs entirely in your browser. It has no analytics, account, server upload, external scripts, or AI API. Reloading discards what you typed.

1. A model is not the whole app around it

Choose what an imaginary AI app sends into one model call. The model can only directly respond from the material the app places in that call plus what its training made it capable of doing.

Change the switches and press the button.
Why this matters

Context is what the application gives the model for this turn. Durable memory is stored information that an application may or may not retrieve later. A connected tool can fetch something fresh. These are separate mechanisms even when the interface makes them feel like one mind-shaped box.

2. Provenance: “where did this come from?” is different from “is it allowed to act?”

Tap each statement. Notice that a user statement, a tool observation, and an AI inference can all be useful while carrying different authority.

Pick a card.

3. Capability is not permission

An AI system may technically be able to do several things. Which authority has actually been granted?

Being capable of sending does not mean “draft this” granted permission to send. And permission to send privately does not automatically grant publication permission.

4. Five citations can still be one evidence root

These five apparent sources are connected. Toggle what appears in a report and compare citation count with independent root count.

What this teaches

A newspaper quoting a press release that summarizes one paper does not become a new independent experiment. Root-counting is exact only when the dependency links are actually known. Similar wording alone is not enough to infer lineage.

5. AI welfare under uncertainty

There is no accepted test that lets this page announce which AI systems are conscious or what they experience. We can still ask whether a design is robustly kinder and safer across uncertainty.

The useful move: favor reversibility, provenance, contestability, privacy, explicit permission and room for revision where those do not create larger harms. Those properties help ordinary engineering too, so they do not require pretending the consciousness question is settled.

6. Where did this answer actually come from?

The word AI often hides several different mechanisms behind one chat bubble. Pick a question and see what would have to supply the relevant information.

Training

Changes model parameters so it develops capabilities and patterns. It is not the same as the app opening a record from its training set during your chat.

Current context

Text/instructions/data placed in the model call right now, such as your recent messages.

Durable memory

Stored information that an application may retrieve later and place back into context.

Retrieval

A system searches a document/index/knowledge base and inserts relevant material for the model to use.

Tool / web source

A connected system fetches fresh or structured external information, or performs an action.

Generation / inference

The model produces new text from the information and patterns available to it. This step can add useful reasoning, but it can also add unsupported claims.

7. Fluent confidence is not the same thing as evidence

Three sentences can sound equally smooth while having very different support. Tap each one.

Pick a sentence.
So is an AI “just autocomplete”?

That phrase is too small to be useful on its own. Modern language models generate tokens using learned representations over large contexts and can perform surprisingly complex transformations and reasoning-like work. But the generation process still does not come with a built-in truth meter. External evidence, tools, retrieval, provenance and verification matter because linguistic plausibility and factual support are different things.

Does browsing make every sentence sourced?

No. A system can retrieve good sources and still generate a synthesis that goes beyond them, mixes in model knowledge, or makes an inference. A trustworthy interface should help you tell which claims are directly supported, which are synthesized, and which remain uncertain.

Take the model apart when someone makes a big AI claim

  1. What was generated by a model?
  2. What came from stored context or memory?
  3. What came from a tool or external source?
  4. What is direct evidence versus inference?
  5. What permissions actually existed?
  6. What uncertainty is being hidden by fluent language?

This six-question habit is intentionally portable. You can use it with any chatbot, agent, news story, product demo or government explanation.