Lies at scale

You are scrolling, half paying attention, when a photo stops you cold. A little girl sits alone in a small boat on brown floodwater. She is crying. She is holding a soaked puppy. The caption says she is a child from the storm that just tore through whole towns. Your chest tightens. Your thumb moves toward the share button.

Do not. There is no girl. There is no puppy. No boat, no rescue, no child in danger at all. A machine drew the picture, and it fooled a great many kind people who only wanted to help.

Lesson 8 was about what AI does to your work. This last lesson is about what AI does to something even more personal: what you are able to trust. The same tool that can draft your emails can also pour convincing false words, images, and voices into the world faster than anyone can check them. That sounds like a reason to be afraid. By the end of this lesson it will be a reason to feel steadier, because the point of this track was never fear. It was standing on solid ground.

Two words, one difference

Not all false information is the same, and the difference has a name worth learning. The Harvard Kennedy School course this track adapts, created by Sharad Goel, Dan Levy, and Teddy Svoronos, starts with two plain terms. 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 whole difference is intent.

The flood picture was both at once. Someone made it. Then thousands of people who believed it shared it, meaning no harm, and the honest mistake traveled further than the deliberate lie could have on its own.

The machine’s own false claims

Some false information starts with the tool itself. To see why, recall the one idea this whole track rests on, from Lesson 1: a generative model writes one word at a time by predicting what is likely to come next. It is not looking anything up. It is not weighing a claim against the world. So it can state something false just as smoothly, and just as confidently, as something true.

This behavior finally gets its name. A hallucination is when a model produces a confident, fluent, false claim. Lesson 1 promised you would learn why answers sound confident even when wrong. This is the answer, and hallucination is its name.

It is a slippery word, and the course’s own teacher is not fond of it. It makes the machine sound like it sees or believes things. It does neither. Some researchers just say false claims. Either way the behavior is real, and once you expect it you are already safer.

One well-known case shows the shape. In an early public demo in 2023, a major chatbot was asked about the James Webb Space Telescope and answered, cleanly and confidently, that the telescope had taken the very first picture of a planet outside our solar system. It had not. That first image came nearly twenty years earlier, from a different telescope entirely. Nothing in the answer looked wrong, and that is the trouble. Good grammar and a sure tone are the very cues we use to judge truth, and a prediction machine hands you both for free.

Treat a confident answer as a draft, not a verdict. Before you rely on something that would cost you if it were wrong, ask the two questions the course borrows from one researcher: 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.

When the lie is on purpose

Disinformation is not new. People have spread lies to sway others for as long as there have been others to sway. What generative AI changed is the price and the polish. Making a convincing fake used to take real skill and time. Now it is cheap, fast, and good.

The course offers a clear way to see the threat, and where it can be stopped. For a lie to do harm, three things have to happen. 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.

There is quiet good news in that chain. 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.

You might hope another limit protects you: the safety training that makes an assistant politely refuse an ugly request, which you met in Lesson 6. It helps, but do not lean on it as a wall. A refusal is a speed bump, and 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.

The sharper risk is a fake aimed at you in particular, wearing a face or a voice you already trust. In early 2024, a finance worker at a global engineering firm joined a video call with what looked and sounded like the company’s chief financial officer, along with several familiar colleagues. Every person on that call was an AI fake. Convinced by faces he recognized, he sent out about 25 million dollars before anyone caught it. The case was reported widely months later.

The defense is the same steady move you learned for the voice-cloning scam in Lesson 6. When urgency, a request for money or data, and a trusted face or voice all arrive together, stop. Verify through a separate channel you already trust. Hang up and call back on a number you had before. Ask something only the real person would know. The fake is built to rush you, so slowing down is most of the defense.

The liar’s dividend

Now the deepest harm, and it is not the one you would guess. It is not the fakes we fall for. It is the truth we start throwing away.

Once everyone knows that anything can be faked, a new escape hatch opens. A person caught saying or doing something real, on genuine video or audio, can simply wave it away. That is AI, they say. That is fake. Researchers call this the liar’s dividend: the liar profits from the mere fact that fakes exist, even the fakes nobody made. As one digital-forensics expert has put it, in the age of deepfakes anyone can deny reality, and it takes only a few convincing fakes to poison the well until everything looks suspect. After one major news event in 2023, there were very few actual deepfakes in circulation, yet many people refused to believe real footage, simply because now they could.

This has already reached the courts. Lawyers now raise what some call the deepfake defense, claiming genuine recordings are AI fabrications. In one case a company argued that recorded statements by its own chief executive might be fakes; the judge was not persuaded. Judges now say openly that they worry about deciding a real person’s fate on evidence they cannot be sure is real.

The liar’s dividend works by making you cynical about everything, and cynicism can feel like safety. It is not. The answer to “anything could be fake” is not to believe nothing. It is to verify instead of dismiss.

The resilience toolkit

So what pushes back? The course sorts the responders into four groups: the companies that build the tools, the platforms that carry the content, governments, and you. Every fix higher up the chain leaks, which is exactly why your own habits are the part that holds.

Start with the companies. Their main move is to mark their own output at the source. Two techniques, both defined in Lesson 6, do the work: watermarking, a hidden mark buried inside AI content, and detection, software that tries to guess whether something was machine made. As of 2026 the largest AI labs have mostly agreed to pair an invisible watermark with a signed record, riding along with the file, that says what made it; more than ten billion files have been marked this way in the last few years. Real progress, and still leaky. A mark only helps if the tool that made the content chose to add one, and the signed record can fall off when a file is re-saved or screenshotted. Detection is shakier: the company behind the best-known tool for spotting AI writing shut it down for being wrong too often, including flagging real human writing, which hit people writing in a second language hardest. Never read “no watermark” as “real.”

Platforms come next. They label what they can, sometimes detecting AI automatically from that signed record, and some let ordinary readers attach a shared note to a viral post, adding the context a fake leaves out. Useful, and still partial: labels catch the honest and the automatic, not the determined.

Governments hold the heaviest levers, and Lesson 6 walked through them: new laws are starting to require that AI content be disclosed and labeled, with major rules landing in 2026. But the course flags an honest tension between free expression and safety. A government that gets to decide which sources are legitimate can slide, in the name of fighting lies, toward silencing reporting that is true but inconvenient. One government pulled a major broadcaster’s press credentials in 2023 on that pretext. Watch this lever from both sides.

Which leaves you, and you are not the last resort. You are the reliable one. The single most useful habit has a name: lateral reading. Instead of staring at a suspect page or image and hunting for clues inside the thing itself, you leave it. You open a new tab and check what independent, trustworthy sources say about the same claim. That is exactly how the flood picture was caught. Fact-checkers and a forensics expert did not just squint at it. They lined up the copies spreading online and found the story would not hold still. The child’s shirt changed color from one version to the next. The puppy’s markings changed too. The boat did not stay the same. Those are the fingerprints of a machine that cannot keep its own details straight across copies. 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.

Two older defenses still matter. When spam email flooded in, we learned its telltale signs and mostly stopped clicking, and we can learn the signs of an AI hoax the same way. And supporting real, local journalism rebuilds the very thing disinformation eats away: sources you have earned reasons to trust.

One last comfort, and a practical one: bad actors still answer to cost. Make lies costlier to spread and easier to catch, and we take back ground. This is a fight we are genuinely in, not one we have lost.

Why this matters when you use AI

Confident is not correct. A fluent, self-assured answer is what a prediction machine produces whether it is right or wrong. Treat tone as no evidence, and check what matters.
The real target is your trust. The worst harm is not one fake you believe. It is slowly losing the ability to believe anything. Guard against that by verifying, not by going numb.
The habit beats the tool. Watermarks, detectors, and labels all leak. Your own move, look sideways before you believe, works on every fake, including the ones nothing flagged.

Common pitfalls

Reading confidence as truth. Hallucinations fool us because they sound exactly like correct answers. If you are convinced mostly because a reply was fluent and sure, that is the moment to check.

Waiting for a label to warn you. A missing AI label does not mean content is real. Plenty of fakes carry no mark at all. Absence of a warning is not a green light.

Curdling into cynicism. Deciding everything online is fake hands the liar exactly what they want. The goal is not to believe nothing. It is to verify before you believe, and before you dismiss.

What you should remember

Misinformation is false information spread without intent to deceive; disinformation is spread on purpose. The line is intent.
A hallucination is a confident, fluent, false claim from a model, produced because it predicts likely words rather than checking truth. Expect it; check anything that matters.
The liar’s dividend is the deeper danger: once anything can be faked, real evidence gets waved away as fake. Verify, do not dismiss.
Watermarking, detection, labels, and new laws all help, and all leak. None is a complete shield.
Your strongest habit is lateral reading: leave the suspicious thing and check what independent, trusted sources say. It is how experts catch fakes, and you can too.

Where the track leaves you

There is no next lesson. This is the end of a road we started nine lessons ago, so let us look back.

You began with a machine that writes one word at a time by predicting likely text, powerful but not a mind. From there you learned to ask it well, to work beyond the chat window, to judge which tasks it should touch, and to run a whole real project with it. Then the track turned to the harder question, what all of this means for your world: an honest map of the risks, the fight over who owns what AI makes, a clear-eyed look at your own job, and now this, keeping your footing when the same machine can flood the world with lies.

Here is what you carry out. Not a shield that blocks every fake, because no one has that. Something better: your own judgment, sharpened. You can tell a mistake from a lie. You can expect the confident wrong answer and check it before it costs you. You can feel the tug of a fake built to move you, and slow down instead of sharing. You can look sideways before you believe. None of that requires you to be technical. It requires you to be awake, and you are.

Lesson 1 said that once you understand the trick, the fear starts turning into fluency. Nine lessons on, that is what happened. You are not more replaceable than the day you started. You are harder to fool, steadier on your feet, and more in charge of what you let into your head. That is real power, and it is yours to keep.

If you remember one thing

Misinformation is a mistake, disinformation is a lie, and a hallucination is the machine’s own confident wrong answer, and you can now name all three.
No tool will catch every fake for you, so the durable defense is a habit: verify instead of dismiss, and look sideways before you believe.
This track never promised to make you fearless. It promised to make you harder to fool, and that is where fluency was leading all along.