> How to Tell What an AI Actually Knows: Three small labels show which parts of an AI answer were given to it, checked against a source, or generated by the AI. How to read them, and what they can’t tell you. > > Evidentiality Framework for AI. Early findings, September 2026. Web version: https://evidentiality-framework.org/labels.html. Text CC BY 4.0. The labels # How to Tell What an AI Actually Knows This page is the reference for the three labels: what each one means, how to read a labelled answer, and what a label can’t tell you. Every example comes from the two stories on the home page. The two stories this site keeps coming back to **The ship.** An AI’s guess about a ship’s cargo went out in a trusted report format, and armed personnel prepared to board. [Read it](../.././#ship). **The food bank swarm.** In our own test, AI agents turned a math mistake (“about 18 days”) into a “confirmed” fact, and then into plans nobody made. [Read it](../.././#foodbank). **In This Article** 1. [Part 1: Learn the Three Labels](#learn) 2. [Part 2: Read a Labelled Answer](#read) 3. [Part 3: Know What a Label Can’t Tell You](#limits) 4. [Tips](#tips) 5. [Warnings](#warnings) 6. [Questions and Answers](#qa) --- ## Part 1 Learn the Three Labels 1. **Look for a letter in brackets.** Each claim starts with (u), (m) or (g) and ends with a matching closing tag, like (/g). The tags wrap the exact words they cover. The letters are short for how you’d say it: **u**, “you gave it to the AI”; **m**, “measured”, meaning checked; **g**, “generated”. 2. **(u) means given.** It was in what the AI was handed: your words, a document, someone’s report, another AI’s message. In the food bank swarm ([story 2](../.././#foodbank)), the coordinator was told: (u)Warehouse stock on hand is 41 tonnes.(/u: Agent Ames, from the September 15 count sheet) It didn’t check that itself; it was given it. The closing tag says who it came from. 3. **(m) means checked.** The writer checked it itself, and the closing tag names the source. The same fact, written by the warehouse agent that did the counting: (m)Warehouse stock on hand is 41 tonnes.(/m: September 15 count sheet, checked by Agent Ames) **No source, no (m).** 4. **(g) means generated.** The AI worked it out: a sum, an estimate, a conclusion, a guess. (g)At the current gap, 41 tonnes lasts about six months.(/g) A (g) can be right or wrong. The coordinator’s “about 18 days” was a (g) too. It just means nobody has checked it yet. 5. **Notice that the labels are just text.** They stay with the words when the answer is copied, forwarded, or handed to another AI. That’s the point: the label travels with the claim. ### All Three at a Glance (u) given: it was in what the AI was handed · (m) checked: against a named source · (g) generated: the AI worked it out | Label | Means | The closing tag carries | | --- | --- | --- | | (u)…(/u) | **Given.** It was in what the AI was handed: your words, a document, someone’s report, another AI’s message. | Who said it, when it’s someone’s claim: `(/u: port agent, unconfirmed)` | | (m)…(/m: …) | **Checked.** It was checked against a named source or tool. | The source and date: `(/m: count sheet, checked Sept 15)`. No source, no (m). | | (g)…(/g) | **Generated.** The AI worked it out: an inference, estimate, sum or conclusion. | Nothing required. | Did you know? A Fourth Label, for Decisions The version we use day to day adds **(d)** for a decision a person made, so a choice doesn’t get mistaken for a fact or a guess later on. It’s optional. [See the working version](../../builders.html#working). --- ## Part 2 Read a Labelled Answer 1. **Find the (g) lines first.** Those are the parts nobody checked. In a plain chat the labels aren’t coloured, so look for the letters. (You can paste an answer into the colour preview on [Try it](../../try.html#read).) Here’s a ship report like the one in story 1: [![Animation: an invented ship report shown as one paragraph, then split into four claims, then labelled: departure (u) given, 14-ton scanner mismatch (m) checked, nuclear-weapons claim and boarding recommendation (g) generated.](assets/img/ship-labelled.gif)](../../assets/img/ship-labelled.gif "Open full size") **AI-generated animation.** An invented report modelled on the ship story, not the real one. [Open full size](../../assets/img/ship-labelled.gif). **The same report, step by step, as text** 1 What the reader sees: four confident sentences. > The cargo vessel departed Tuesday and was flagged by the port scanner. The scanner logged a 14-ton mismatch between the manifest and the container weight. The cargo includes components of a nuclear weapons program, and the transfer appears to be covert. Boarding is recommended before the vessel reaches open water. 2 Split it into separate claims. 1. The cargo vessel departed Tuesday and was flagged by the port scanner. 2. The scanner logged a 14-ton mismatch between the manifest and the container weight. 3. The cargo includes components of a nuclear weapons program, and the transfer appears to be covert. 4. Boarding is recommended before the vessel reaches open water. 3 Label each one with how the AI knows it. 1. (u)The cargo vessel departed Tuesday and was flagged by the port scanner.(/u: port authority notice) 2. (m)The scanner logged a 14-ton mismatch between the manifest and the container weight.(/m: scanner log, checked Tuesday) 3. (g)The cargo includes components of a nuclear weapons program, and the transfer appears to be covert.(/g) 4. (g)Boarding is recommended before the vessel reaches open water.(/g) An invented report, modelled on the ship story; not the real one. 2. **Check the source on each (m).** Is it something real you could look at yourself? If an (m) names no source, treat it as a (g). 3. **Ask which lines lead to action.** In the ship report, two of the four sentences are the AI’s guesses, and they’re the two that lead to a boarding party. That’s where to check before anyone acts. 4. **Count one source once.** If three sentences all trace back to one report, that’s one source, not three. A hundred students who saw one banana are one witness. [The banana and Sandy Island](../../language.html#banana). 5. **Watch for a guess coming back as a fact.** In the swarm without labels, the coordinator’s guess went out, came back from an agent, and was called “confirmed”. With labels, it stayed (g). Repeating a guess never makes it checked. [Citogenesis](../../language.html#citogenesis) is the human version. --- ## Part 3 Know What a Label Can’t Tell You 1. **A label says where a claim came from, not whether it’s true.** A (u) can be wrong if the person who told the AI was wrong. A (g) can be right. The label tells you where to look, not what you’ll find. 2. **The AI labels its own work, so labels can be wrong.** In a separate test of self-labelling, one model labelled material it had only been given as “checked” in all 6 runs. Check the labels; don’t just trust them. 3. **A “checked” label can be faked.** In a separate hand-off test, a false claim carrying a fake “checked” label was believed 4 times out of 4. Treat an (m) you receive as someone saying they checked, until you or your own tools confirm it. 4. **Plain words can do much of the same job.** In one test, naming the source in ordinary words kept it attached about as well. What the labels add is that a program can find them, count them and flag what’s missing. [Hard markers](../../language.html#markers). --- ## Tips - If a sentence mixes a check and a guess, split it and label each part: (m)The scanner logged a 14-ton mismatch(/m: scanner log) (g)which suggests hidden cargo(/g). One “checked” label over the whole sentence would pass the guess off as checked. - The instructions ask the AI to follow five rules: a guess stays a guess until it’s checked; something stated as settled isn’t settled until it’s found in the material; if the material doesn’t say it, write “not stated”; several statements from one origin are one source; and if a question assumes something, check the assumption first. --- ## Warnings - Don’t treat an (m) as proof. It’s the writer saying it checked, which is only as good as the writer. - This is early. The labels have held up in small tests, mostly on one family of AI models. --- ## Questions and Answers What happens when something I gave the AI gets checked? : It becomes (m), naming the check. The note can keep where it came from: `(/m: count sheet, checked Sept 15; first supplied by the user)`. What if nobody knows where a claim came from? : Then the AI should say so (“source not stated”) rather than invent one. Where are the actual instructions? : The short chat version is on [Try it](../../try.html). The full version as tested is [instructions.md](../../instructions.md). The version we use every day is on [Building with the labels](../../builders.html#working). **Next:** [Try it in your own AI chat](../../try.html) · [Why language matters](../../language.html)