Start here: the same poster job, with one new part#
You have probably made a poster before. You picked a photo, typed the details, and moved things around until it looked right. An image generator changes one part of that job: where the picture comes from. You describe a picture in words, and a program makes it.
If a program makes the picture, which parts of the poster job are still yours?
Show the answer
Almost all of them. You decide what the poster must say and who it is for. You check the details. You judge whether it looks right for the people you want to reach. And you make sure you are allowed to use it. The tool only saves you the drawing.
Here is the job we will carry through the module. Your neighborhood is holding the Maple Street Community Swap: bring books, clothes and plants you no longer need, and take home what you can use. It is at the community center on Saturday, 10 a.m. to 2 p.m., and it is free. You have volunteered to make a printed flyer and a 15-second vertical video for phones.
Making them with AI raises three questions that clip art never did:
- Is it right? Does it say the right thing, fit the job, and hide no mistakes?
- Is it allowed? Did everyone shown or heard agree, and may you use what the tool made?
- Is it real? When someone else shares a picture, how do you check it before passing it on?
What image, voice and video generators do#
If you type a sunny park with tables of books into an image generator twice, will you get the same picture both times?
Show the answer
Usually not. Each run starts from a different random beginning, so each picture differs. That gives you options. It also means choosing, and checking what you chose, is part of the job every time.
A generator is a program that makes new pictures, sounds or video from a description. It was trained on a huge number of examples, such as pictures paired with captions. It does not fetch a stored picture. It produces new pixels that fit the patterns it learned.
Many image generators work roughly like this. Picture a TV screen full of random static. The program cleans the static up in many small steps, each one nudging it closer to something that matches your words. At the end, the static has become a park that never existed.
The tool matches patterns; it does not check facts. So it can draw a confident sign with a misspelled word, or a hand with six fingers, and see nothing wrong.
| Kind | You give it | It gives back | Watch for |
|---|---|---|---|
| Image | A description, sometimes a sample picture | A new picture | Garbled lettering, odd hands, details that look right and are not |
| Voice (text to speech) | Text, and a choice of voice | Spoken audio | Wrong stress on names, odd pacing |
| Voice clone | Recordings of one specific person | New speech that sounds like them | Whose voice is it, and did they agree? |
| Video | A description, sometimes a starting picture | A short clip | Things that change shape between frames |
Video is the hardest of the four. Each frame, one still picture in the sequence, has to agree with the frame before it. Keeping a face or a sign the same across frames is harder than making one good picture. That is why objects in generated video sometimes melt or swap.
Describing what you want, then iterating#
Your first request for the flyer picture is flyer for a community swap. What do you expect to be wrong with what comes back?
Show the answer
Everything you did not say gets decided for you: style, colors, who appears, layout. It may also add its own lettering, with spelling mistakes or a made-up date. A vague request gets the tool's average guess, and the average guess is not your event.
A good prompt (the text you type to tell the tool what you want) works like a brief to a human designer. It answers four questions:
| Part | The question | For the swap flyer |
|---|---|---|
| Subject | What is in it? | Tables of books, folded sweaters and potted plants under a big tree |
| Style | What should it look like? | Flat illustration, warm autumn colors, friendly |
| Composition | How is it arranged, and what shape? | Tall page; top third left plain for the words |
| What to avoid | What must not appear? | No lettering or signs, no logos, no recognizable faces |
A flat illustration of a neighborhood swap: tables with stacks of books,
folded sweaters and small potted plants under a large tree. Warm autumn
colors, friendly and simple. Tall page shape. Keep the top third plain
and light so text can go there. No words, letters or signs anywhere.
No logos. People only as small distant figures, no faces.Notice the "no words" line. You will add the name, date and place yourself in a document or design program, where every letter is exactly what you typed. The section on weak spots explains why.
Then iterate: try, look, change one thing, try again. One change at a time tells you which change helped. Judge each version against your brief, not against the last version, or you drift toward whatever the tool likes making.
Open any free chat assistant, such as ChatGPT, Claude, Gemini or Copilot. Paste this:
I am making a printed flyer for a free neighborhood swap of books,
clothes and plants. Help me write a description for an image generator
with four parts: subject, style, composition, and what to avoid. The
event text will be added separately, so the picture must contain no
words. Before you write it, ask me two questions.What to look for: did it ask, or did it invent details such as a date or an address? Is there a "what to avoid" part? If your assistant can make images, ask for three versions. Pick one by checking each against your four parts, not by which you liked first.
Where your taste decides#
You ask a tool to score three flyer drafts for "visual appeal". It says 6.1, 7.4 and 8.2. Does the 8.2 win?
Show the answer
Not on that number alone. You do not know what the score measured. And it is not the question that matters: will your neighbors notice this flyer across the laundromat and know when to come? Ask what was measured, then use your own eyes for fit.
Some questions have a checkable answer: is the date right, can you read it from 10 feet away? Measure those. Others are about fit: is this the right feel for these people? No score can answer that, because it depends on a purpose only you know.
Kestrel Exams is a small study site built by one person working with an AI assistant. Here is something that happened. The owner picked a background music track by ear. Our AI assistant measured beat strength across six candidate tracks and twice argued for a different one. The owner's ear was right: the measurement had found a drum groove, but the job was a calm bed of music under a spoken voice.
Later, the assistant flagged the voiceover as "too fast", judged against a target it had invented. Measuring the voiceover already published showed the real pace was faster still. The rule we kept: measurement settles facts (is this file the same song?). A person's judgment settles taste (is this the right song?). Say what you measured once, then defer.
Judgment has a second half. You also set the standard, and a tool should not lower it for you.
For a short animation, the target score for visual density, a measure of how much was happening on screen, was 12. Three tries scored 11. Our AI assistant argued that the target was wrong. The owner said keep the target. The fourth try used genuinely different methods and scored 17.1. The rule: when AI output misses a standard you set, push for a different approach before you move the goalposts.
After a few weak results, it is tempting to decide the brief was too strict. Sometimes it was. First ask: have I tried a genuinely different approach, or only small changes to the same one? Lower the bar on purpose, for a reason you can say out loud, never just because you are tired.
Where generators still slip: check these first#
Your generated flyer picture looks great on your phone. Before you print 200 copies, where should you zoom in?
Show the answer
Any lettering or signs (are they real words?), hands and faces (count the fingers), and small repeated things such as chair legs. And anything that must be factually right, such as a map to the center, should not come from a generator at all.
Generators improve quickly and the weak spots move, so check these every time:
| Weak spot | What can go wrong | What to do |
|---|---|---|
| Text in the picture | Misspelled or made-up words, a wrong date | Ask for no words; add all text yourself |
| Hands, faces, small objects | Extra fingers, merged objects | View it at full size, not as a thumbnail |
| Consistency | A mascot that changes between pictures; objects that morph between frames | Keep a reference picture; watch video frame by frame |
| Factual pictures | Maps, diagrams and charts that look right and are wrong | Make them from real information you can check |
The last row matters most. A generated map to the community center can look convincing and send people to the wrong street. A picture meant to inform has to be checked like a fact, because it is one.
The video has one more limit, and it has nothing to do with AI. Phone video apps cover parts of the screen with buttons and captions. Designers keep important words inside a safe zone, the part of the frame that stays clear. Where it sits depends on the app, so check that app's current guidance and write the numbers down.
Two of our projects made vertical videos on the same topic, each with its own layout standard. Our AI assistant applied the wrong project's standard. A headline ended up 168 pixels inside a zone that had to stay clear. The rule: name which project a rule belongs to, put the actual numbers into the work, and check the result against them.
For the swap: write each app's safe zone down separately, and tell your assistant which one applies. "Keep the text inside these numbers" is a check; "keep it in the safe zone" is not.
Real people: consent comes first#
The center's director says: "End the video with me saying 'See you Saturday'. Just copy my voice from last year's speech video with an AI tool." What do you do?
Show the answer
Ask her to record the line herself. Ten seconds on a phone is simpler than a voice clone, and it leaves no copy of her voice sitting in a tool. She has offered, so it is her choice. What you never do is clone anyone's voice, or make a picture or video of anyone, without their clear permission.
This part of the module has firm rules, not judgment calls:
- Never make images, voices or videos of a real person without their permission. Permission means they know what you are making and where it will appear, and they said yes. For a child, ask a parent or guardian, and the child too when they are old enough to understand.
- Never make sexual or humiliating content of any real person. Not as a joke, not in private, not for any reason.
- Real photos count too. If last year's swap photos show recognizable neighbors or children, ask before you reuse them.
In some places the law backs these rules up. In the United States, the TAKE IT DOWN Act was signed in May 2025. In many cases it makes it a federal crime to knowingly publish an AI-made intimate image of a real, identifiable person without consent. Since May 2026 it has also required covered online platforms to remove such images within 48 hours of a valid request. Other countries have their own laws, and this is not legal advice.
Voice clones have a darker use too. The US Federal Trade Commission warns that a scammer may need only a short clip of a voice, perhaps from something posted online. With it, they can fake a call from a relative in trouble. Its advice: do not trust the voice. Call the person back on a number you already know, or check with another relative. A request to pay by wire transfer, cryptocurrency or gift cards is a warning sign.
A picture that is obviously a joke to you can be screenshotted and passed on without the context that made it a joke. The person in it lives with the result. If someone real could be embarrassed or hurt, do not make it.
Spotting fakes: check the story, not just the pixels#
The night before the swap, a neighborhood group shares a photo of the center's parking lot under water: "Swap cancelled!" Before you repost it, what would you check?
Show the answer
Who posted it first? Does the center itself say the swap is cancelled? Did it rain? Has this photo appeared online before? Is that even the right building? None of these needs you to decide whether AI was involved: an old photo of a different flood would fool people just as well.
A deepfake is media made or changed with AI to show something that did not happen. Often it shows a real person saying or doing something they did not. Staring at the pixels is the weakest check: good fakes look fine, and real photos sometimes look odd. Check the story around the picture instead:
- Slow down. A post that makes you alarmed or rushed is exactly the one to check before you share.
- Find the source. Who first posted this? Go to the official channel yourself: the center's own page, or a phone call.
- Search with the image. A reverse image search uses a picture instead of words. Several search engines offer one. It shows where else the picture has appeared, often with dates. An old photo from somewhere else, passed off as new, needs no AI at all.
- Check the context. Do the weather, season, signs and building match the claim?
- Look for Content Credentials. These are a record of where a file came from and how it was edited. They follow an open standard from the Coalition for Content Provenance and Authenticity (C2PA), which compares them to a nutrition label. The record is signed, so tampering shows. If an app displays one, open it.
Content Credentials show where a file came from, not whether what it shows is true. The standard's own explainer says so. Adding them is optional and they can be stripped, so many pictures have none. A missing label does not mean fake, and a present label does not mean true.
Paste this into any chat assistant:
A neighborhood group posted a photo of our community center's flooded
parking lot, captioned "Swap cancelled!" The event is tomorrow. List,
in order, the checks I should do before I share it. Then tell me which
of those checks you cannot do for me.What to look for: a good reply sends you to the source and the official channel and suggests a reverse image search. It should admit it cannot see the photo, call the center or know the local weather. If it accepts images, show it one you generated and ask "Was this made by AI?" Unless it read a built-in label or watermark (a hidden mark some tools add to what they make), it is guessing from looks.
Who owns it, labeling it, and alt text#
You made the flyer picture with a generator, from your own description. In the United States, does copyright protect that picture as your work?
Show the answer
Probably not the generated picture itself, if it came purely from your prompt. Your own words on the flyer, any photo you took, and the way you chose and arranged the parts can still be yours. Other countries differ.
Who owns it. Copyright is the legal right to control copies of a creative work. As of October 2026, the US Copyright Office's position is in its January 2025 report on copyrightability. In its words, "copyright does not extend to purely AI-generated material." It also says that, with today's tools, "prompts alone do not provide sufficient human control" to make you the author. Using AI "to assist rather than stand in for human creativity" does not remove protection. Your own expression visible in the result, your creative selection and arrangement, and your creative edits can be protected. The Office decides case by case.
One court ruling so far backs the most basic part. In Thaler v. Perlmutter (March 2025), a federal appeals court held that US copyright requires a human author. It did not decide how much human input is enough. In March 2026 the Supreme Court declined to hear the appeal. Rules differ by country. Read the generator's terms of use and the rules of any site where you post: they say what you may do. This is not legal advice; if money depends on it, ask a qualified person.
Ownership runs the other way too. Do not ask for a copy of a known character, a logo or someone's artwork. Describe the look in your own words, such as "bold flat shapes, warm autumn colors", instead of naming a living artist.
Labeling it. Say something is AI-made if a reasonable person could take it for a real photo, recording or person. A flat illustration fools nobody; a photo-realistic crowd at a swap that never happened could. A small line such as "Illustration made with AI" costs nothing. Many platforms, schools and employers have labeling rules of their own, so check yours.
Making it readable for everyone. Alt text is a short written description attached to an image. Screen readers, programs that speak the screen aloud for people who are blind or have low vision, read it out. The W3C, the body that sets web standards, says alt text should convey the image's information or function. For an image of text, it says the alt text should contain the same words. Your flyer is mostly text, so its alt text must carry the name, date, time and place exactly. Better still, type those details in the post too.
Paste this into any chat assistant:
Write alt text, in no more than two sentences, for a flyer I am
posting online. The flyer says: "Maple Street Community Swap. Bring
books, clothes and plants. Take what you can use. Saturday, 10 a.m. to
2 p.m., Maple Street Community Center. Free." Below the text is an
illustration of tables of books, sweaters and plants under a large tree.What to look for: check the reply against the flyer word by word. Are the day, time and place exactly right? Did it invent an address, a date or people who are not there? Did it spend words on decoration and drop the information? AI drafts fast; you check, because a screen reader user cannot see the flyer to catch the mistake.
What was new in this module#
- A generator makes new pictures, sounds or video by matching learned patterns. It does not check facts, so every result needs a human check.
- A good request covers subject, style, composition and what to avoid. Then iterate, one change at a time, judged against the brief.
- Measurement settles facts; your judgment settles fit. You set the standard, and a tool should not quietly lower it.
- Check the weak spots every time: text in pictures, hands and small objects, consistency, and anything factual.
- Never make images, voices or videos of a real person without permission, and never sexual or humiliating content of anyone real.
- To check suspicious media: slow down, find the source, search with the image, check the context, and read any Content Credentials as one clue, not a verdict.
- In the US, as of October 2026, purely AI-generated material is not protected by copyright; your own contributions can be. Rules differ by country and platform.
- Label AI-made media that could pass as real, and check that alt text carries the image's information.
The tool makes the picture; you answer for it. Before it goes out, check four things: it is right, everyone in it agreed, it is labeled if it could pass as real, and everyone can read it.
Key terms#
- Generator
- A program that makes new pictures, sounds or video from a description by matching patterns it learned.
- Prompt
- The text you type to tell an AI tool what you want.
- Iterate
- Try, look, change one thing, and try again, judging each version against your brief.
- Voice clone
- Generated speech made to sound like one specific person, built from recordings of their voice.
- Deepfake
- Media made or changed with AI to show something that did not happen, often a real person doing or saying something they did not.
- Consent
- Clear permission from a person who knows what you are making with their face or voice, and where it will appear.
- Reverse image search
- A search that uses a picture instead of words to show where else it has appeared.
- Content Credentials (C2PA)
- A signed record of where a file came from and how it was edited. It shows origin, not truth, and it can be missing or removed.
- Safe zone
- The part of a video frame that an app does not cover with buttons or captions.
- Alt text
- A short written description of an image, read aloud by screen readers.
Check yourself#
Answer each one out loud before you open it. Reading the answer without trying first feels like learning and is not.
1. A friend says: "The AI scored my poster 9 out of 10, so it is ready to print." What do you ask first?
What did the score measure, and does that match the job? Then check the facts yourself (date, time, place, spelling) and judge whether it fits the audience.
2. Your generated bake sale picture has a cheerful sign reading "Bake Sael Saturdy". The rest is lovely. Best fix?
Generate it again with no words, then add the text yourself. Text you type is exactly what you typed.
3. A classmate wants a "funny" voice-clone video of a teacher singing, just for the group chat. What would you say?
No. It uses a real person's voice without permission, it could humiliate them, and a group chat clip can travel without context. Private does not make it acceptable.
4. A shocking photo of a local landmark on fire has no Content Credentials. Is it fake?
You cannot tell from that; many pictures carry no label. Check the source, search with the image, look for official channels, and check the context.
5. You combined a generated picture with your own photos and text in a zine. In the US, what could copyright protect?
Your photos, your text, and your creative selection and arrangement. Not the purely generated picture on its own, according to the US Copyright Office as of October 2026.
I can#
Tick each one only when you could do it tomorrow without this page open.
- I can write an image request that covers subject, style, composition and what to avoid.
- I can tell which questions a measurement can settle and which need my judgment.
- I can name the weak spots to check in generated media before I use it.
- I can explain why I never make images, voices or videos of a real person without permission.
- I can check a suspicious picture before sharing: source, reverse image search, context and Content Credentials.
- I can label AI-made media that could pass as real, and write and check alt text.
Frequently asked questions#
Can I copyright an image I made with AI?
In the United States, as of October 2026, the Copyright Office says purely AI-generated material is not protected and prompts alone do not make you the author. Your own text, art, creative edits and arrangement can be. Other countries differ. This is not legal advice.
How can I tell if an image or video is AI-generated?
Often not from the pixels alone. Check who posted it first, whether official sources confirm it, where else it appears in a reverse image search, whether the context fits, and whether it carries Content Credentials.
Is it okay to make an AI image or voice of a real person?
Only with their clear permission, for a use they know about. Never make sexual or humiliating content of any real person. In the US, publishing AI-made intimate images of a real person without consent is a federal crime in many cases under the TAKE IT DOWN Act of 2025.
What are Content Credentials or C2PA?
A signed record of where a file came from and how it was edited, based on an open standard from the Coalition for Content Provenance and Authenticity. They show origin, not truth, and a missing label does not mean fake.
Do I have to label AI-generated images?
Label AI-made media whenever a reasonable person could take it for a real photo, recording or person. Many platforms, schools and employers have their own rules, so check the one that applies to you.
This course is free, with no ads and no account. If it helped you, you can help keep it that way.
Support Kestrel Exams →Sources#
U.S. Copyright Office (2025). Copyright and Artificial Intelligence, Part 2: Copyrightability. Report of the Register of Copyrights, January 2025.
U.S. Court of Appeals for the D.C. Circuit (2025). Thaler v. Perlmutter, No. 23-5233. Decided March 18, 2025.
Supreme Court of the United States (2026). Docket No. 25-449, Thaler v. Perlmutter. Petition denied March 2, 2026.
Coalition for Content Provenance and Authenticity (C2PA). C2PA Explainer, version 2.2, and C2PA Specification, version 2.4. Overview at c2pa.org.
U.S. Congress (2025). TAKE IT DOWN Act, Public Law 119-12. Approved May 19, 2025.
Puig, A., U.S. Federal Trade Commission (2023). Scammers use AI to enhance their family emergency schemes. Consumer Alert, March 20, 2023.
W3C Web Accessibility Initiative (2026). Images Tutorial. Eggert, E. and Abou-Zahra, S., eds.; Elton, B., update editor; updated April 8, 2026.