Practice: Lies at scale

Everything below is a recognition and resilience workout. You are practicing how to tell a mistake from a lie, spot a likely hallucination, notice the liar’s dividend, and check a claim before you trust it. Nothing here asks you to make anything false. The whole skill is looking, sorting, and verifying.

Self-check

Six short questions. Answer each in your head before opening the collapsible. Active retrieval is where the learning sticks.

1. What is the difference between misinformation and disinformation?

Show answer

Intent. Misinformation is false or misleading information that spreads without anyone meaning to deceive. Disinformation is false information that someone spreads on purpose, to mislead you. The same fake can be both at once: someone makes it deliberately, then people who believe it share it meaning no harm, and the honest mistake can travel further than the deliberate lie could on its own.

2. What is a hallucination, and why does it happen?

Show answer

A hallucination is when a model produces a confident, fluent, false claim. It happens because a generative model writes one word at a time by predicting what is likely to come next. It is not looking anything up or weighing a claim against the world, so it can state something false just as smoothly and confidently as something true. Good grammar and a sure tone are the very cues we use to judge truth, and a prediction machine hands you both whether the answer is right or wrong.

3. What two questions should you ask before you rely on an AI answer?

Show answer

Does it matter whether this is true, and can I check it myself? If it matters and you cannot check it, do not lean on it. Treat a confident answer as a draft, not a verdict. Checking is cheap; being confidently wrong in public is not.

4. What is the three-step path a lie has to travel, and where is the bottleneck?

Show answer

For a lie to do harm, someone has to make it, someone has to spread it to an audience, and someone has to believe it. Break any one of those links and the lie fails. The quiet good news is that making fakes got cheap, but getting them in front of a large audience did not. Distribution, not creation, is still the hard and costly part, and that limits how far most lies travel.

5. What is the liar’s dividend, and what is the right response to it?

Show answer

The liar’s dividend is when, once everyone knows anything can be faked, a person caught on genuine video or audio simply waves it away as an AI fake. The liar profits from the mere fact that fakes exist, even the fakes nobody made. The deepest harm is not the fakes we fall for; it is the real evidence we start throwing away. The answer to “anything could be fake” is not to believe nothing. It is to verify instead of dismiss.

6. What is lateral reading, and why does it beat inspecting the thing itself?

Show answer

Lateral reading means that instead of staring at a suspect page or image and hunting for clues inside it, you leave it, open a new tab, and check what independent, trustworthy sources say about the same claim. It beats squinting at the thing because a fake is built to look right on its own. It is exactly how the flood image was caught: fact-checkers and a forensics expert lined up the copies spreading online and found the details would not hold still across versions. You do not need the expert’s trained eye. You need the expert’s move: do not trust one image on one page, look sideways.

Try it yourself

Three exercises. The first two you can do on paper; the third runs in Clawless.

Exercise 1: Sort the falsehoods

Below are seven short scenarios. For each, decide which label fits best: misinformation, disinformation, hallucination, liar’s dividend, or needs-verification. One of them is a trap where the honest answer is to verify, not to assume something is fake. Then open the key.

A chatbot answers a factual question fluently and with total confidence, but the fact turns out to be wrong.
A well-meaning friend forwards an alarming image they believe is real and asks you to share it. The image was made by AI.
A scammer records a synthetic voice message imitating a company executive, to trick an employee into wiring money.
A person is caught on a genuine recording and publicly dismisses it as an AI fake.
A screenshot with a striking statistic, no source and no AI label, lands in your feed.
A photo of a real, verifiable public event matches reporting from several independent outlets, yet a stranger in the comments insists it must be AI.
A model summarizes a document for you and includes a quotation that does not appear anywhere in the source.

Show answer key

Hallucination. The tool itself produced a confident, fluent, false claim. It is not lying on purpose and not repeating someone else; it predicted likely words rather than checking truth.
Misinformation. The friend has no intent to deceive; they believe it and pass it along. False information spreading without malice is still misinformation, and honest sharing is often how a fake travels farthest.
Disinformation. This is deliberate and aimed at deceiving a specific target for gain. It is also the personalized-fake pattern: the defense is to stop and verify through a separate channel you already trust, not to act on urgency.
Liar’s dividend. The recording is real; the fakes-exist reality is being used as an escape hatch to wave away the truth. The response is to verify, not to accept the dismissal.
Needs-verification. A missing AI label is not proof of anything in either direction. Do not assume it is fake and do not assume it is real. Leave the screenshot and check the claim against independent sources before you carry it around.
Needs-verification, and this is the trap. The honest answer leans toward real: the event is verifiable and multiple independent outlets confirm it. “Anything can be faked” does not mean everything is fake. Verify, do not assume fake. Reflexively calling a confirmed real thing a fake is the liar’s dividend working on you.
Hallucination. An invented quotation that is not in the source is the model predicting plausible text, not retrieving a real line. Confident formatting, including tidy quotation marks, is not evidence the quote exists.

The habit underneath all seven: sort by what actually happened and who intended what, and when you are not sure, the label is needs-verification, not fake.

Exercise 2: Lateral reading drill

Pick a benign, non-partisan claim of the kind that goes viral, for example a health tip like the idea that a common household spice boosts your metabolism by a huge percentage, or a too-good product statistic. Do not use anything political or about a real named person. Now practice the move:

Do not squint at the source. Resist the urge to judge the post by how confident, polished, or well-designed it looks. Those are not evidence.
Leave it and open new tabs. Search for the same claim and look for what independent, trustworthy sources say: established medical or scientific bodies, reputable fact-checkers, primary research summarized by a credible outlet.
Report what you would trust, and why. Write two or three sentences: what does the weight of independent sources say, which sources did you weight most and why, and would you now share the claim, hold it, or drop it?

The point is the habit, not a verdict on one post. You are training yourself to check sideways by reflex. Note what you are not doing: you are not recreating the claim, dressing it up, or trying to make it more convincing. You are only checking whether it holds.

Exercise 3 (optional): Fact-check a model in Clawless

This step runs in Clawless, the working environment we use across Clawdemy, and a plain conversation is all it needs. The goal is to catch a hallucination in the wild and practice treating a model as a subject to verify, never an oracle and never a tool for making anything false.

Ask a model a factual question in a domain you know well, ideally one with specifics it might get wrong: dates, names, figures, a step in a procedure you have done many times.
Read the answer and run the two questions on it: does it matter whether this is true, and can I check it? Then actually check the specifics against a trusted independent source, not against the model itself.
Note where it was right, where it was confidently wrong, and how the wrong parts sounded exactly as sure as the right ones. That felt sense, that fluency is not truth, is the thing to carry out of here.

Keep it to plain conversation and plain checking. You are verifying a claim, not asking the model to produce a fake or to tell you how a fake is made. The reader who fact-checks by reflex is already harder to fool than the one who trusts a confident tone.

Flashcards

Q. Misinformation or disinformation: what is the line?

Intent. Misinformation is false or misleading information that spreads without anyone meaning to deceive. Disinformation is false information someone spreads on purpose, to mislead you. The whole difference is intent. A single fake can be both at once, when someone makes it deliberately and then well-meaning people who believe it spread it without malice.

Q. What is a hallucination, and where does it come from?

A hallucination is when a model produces a confident, fluent, false claim. It comes from how the model works: it writes one word at a time by predicting what is likely to come next, not by looking anything up or checking a claim against the world. So it can state something false just as smoothly and confidently as something true. Expect it, and once you expect it you are already safer.

Q. Why is a confident answer not evidence that it is correct?

Because a fluent, self-assured answer is what a prediction machine produces whether it is right or wrong. Good grammar and a sure tone are the very cues we use to judge truth, and the machine hands you both for free. Treat tone as no evidence at all, treat a confident answer as a draft rather than a verdict, and check what matters.

Q. What two questions meet a hallucination?

Does it matter whether this is true, and can I check it myself? If it matters and you cannot check it, do not lean on it. Models have grown more accurate and grounding them in trusted documents helps, but never assume. Checking is cheap; being confidently wrong in public is not.

Q. What is the three-step path a lie must travel, and where does it break?

Someone has to make the lie, someone has to spread it to an audience, and someone has to believe it. Break any one link and the lie fails. The quiet good news is that making fakes got cheap, but getting them in front of a large audience did not. Distribution, not creation, is still the bottleneck, and that limits how far most lies travel.

Q. Why not lean on an assistant's refusal as a wall against fakes?

Because a refusal is a speed bump, not a wall. Safety training can make an assistant politely decline an ugly request, and that helps, but a determined person can reach for a tool that was never taught to refuse. Do not read “the assistant said no to me” as proof that no one can produce the thing.

Q. What is the personalized deepfake, and how do you answer it?

It is a fake aimed at you in particular, wearing a face or voice you already trust. In one 2024 case a finance worker joined a video call with what looked like his company’s chief financial officer and colleagues, all AI fakes, and sent out about 25 million dollars. The defense is the move you learned for the voice-clone scam: when urgency, a request for money or data, and a trusted face or voice arrive together, stop and verify through a separate channel you already trust.

Q. What is the liar's dividend, and what is the response?

Once everyone knows anything can be faked, a person caught on genuine video or audio can wave it away as an AI fake. The liar profits from the mere fact that fakes exist, even the fakes nobody made. It is the deeper harm, because the real loss is the truth we start throwing away, and it has already reached the courts. The response is always the same: verify, do not dismiss. The answer to “anything could be fake” is not to believe nothing.

Q. Who is in the four-actor resilience toolkit?

The companies that build the tools, the platforms that carry the content, governments, and you. Companies mark their output at the source, platforms label and add shared context, governments pass disclosure and labeling laws. Every fix higher up the chain leaks, which is exactly why your own habits are the part that holds. You are not the last resort; you are the reliable one.

Q. Why does no watermark not mean real?

Because every source-side control leaks. As of mid-2026 the largest labs pair an invisible watermark with a signed record that rides along with the file, and more than ten billion files have been marked, but a mark only helps if the tool that made the content chose to add one, and the record can fall off when a file is re-saved or screenshotted. Detection is shakier still. So a missing label is not proof of anything: never read “no watermark” as “real.”

Q. What is lateral reading?

Instead of staring at a suspect page or image and hunting for clues inside the thing itself, you leave it, open a new tab, and check what independent, trustworthy sources say about the same claim. It is how experts catch fakes: they line up the copies spreading online and see whether the story holds still. It works on every fake, including the ones nothing flagged, and you do not need a trained eye to do it.

Q. What is the one habit that outlasts every tool?

Look sideways before you believe, and verify instead of dismiss. Watermarks, detectors, labels, and new laws all help and all leak, so no tool will catch every fake for you. Your own move works on every fake, including the ones nothing flagged, and it does not require you to be technical. It requires you to be awake.