Week 13: AI-generated Images, Audio, and Video

Welcome to Week 13 of the AI Literacy Course.

Generative AI can create or transform images, voices, music, video, animation, captions, and mixed media.

This makes media production faster and more accessible. A learner can create a fictional illustration. A teacher can prepare a draft visual. A small organisation can produce a prototype without a large production team.

But a convincing file is not automatically accurate, fair, accessible, or suitable for publication.

The central principle is:

Responsible generative media is planned, reviewed, labelled, and documented before publication.

Learning goals

By the end of this lesson, learners should be able to:

• Distinguish generated, AI-edited, and AI-assisted media

• Prepare a production brief for a specific purpose and audience

• Identify important inputs, people, permissions, and risks

• Explain why prompting cannot guarantee accuracy

• Review media for technical, factual, contextual, and human-impact problems

• Plan accessibility for images, audio, and video

• Write a meaningful disclosure notice

• Keep a practical provenance record

• Decide whether to publish, revise, or stop

1. What is generative media?

Generative media is visual, audio, video, animated, or mixed content created or materially transformed with an AI model.

AI may play different roles in production.

Generated media

Most of the content is created by a model from prompts, examples, or other instructions.

Examples include:

• an image created from a text prompt;

• synthetic music;

• an artificial voice;

• an animated scene generated from a description;

• a video produced from a storyboard.

AI-edited media

Existing media is materially changed by AI.

Examples include:

• replacing an object in a photograph;

• extending an image beyond its original borders;

• changing a person’s expression;

• translating and replacing speech;

• animating a still photograph;

• changing what appears to happen in a video.

AI-assisted media

AI performs a more limited supporting function.

Examples include:

• removing background noise;

• suggesting captions;

• separating a subject from a background;

• improving image resolution;

• converting a file format;

• drafting alt text for human review.

These categories can overlap. The important questions are:

• What existed before AI was used?

• What did AI create or change?

• Did the change affect meaning, identity, realism, or evidence?

• What might the audience misunderstand?

2. Begin with a production brief

Before generating or editing media, write a short brief.

Brief element	Question
Purpose	Why should this media exist?
Audience	Who will encounter it?
Message	What should the audience understand?
Format and platform	Where and how will it appear?
Realism	Could it be mistaken for documentation?
Accuracy	Which facts or details must be correct?
People and inputs	Whose material, identity, or work is involved?
Accessibility	What alternatives or adaptations are needed?
Risk	What could be misunderstood or cause harm?
Stop condition	What would make publication unsuitable?

Aesthetic novelty should not replace a meaningful purpose.

Ask:

• Is media generation necessary?

• Could a simpler illustration, photograph, recording, diagram, or text communicate the message better?

• Is realism useful, or could it mislead?

• Is the planned format suitable for the audience?

Illustration is not evidence

A generated image of a historical event may be useful as a concept illustration. It is not a photograph of that event.

A synthetic voice may demonstrate a fictional character. It is not a real person giving testimony.

The brief should identify whether the media is:

• illustrative;

• fictional;

• reconstructed;

• educational;

• documentary;

• evidential.

This distinction will affect both review and disclosure.

3. Identify people, inputs, and permissions

Generative media may involve:

• photographs;

• faces and bodies;

• voices;

• names;

• locations;

• music;

• footage;

• scripts;

• fonts;

• reference images;

• personal documents;

• distinctive rooms or settings.

Before using an input, ask:

• Do I have permission to use this material?

• Does a licence or platform rule limit its use?

• Is attribution required?

• Does the material contain personal or confidential information?

• Is the intended use educational, public, or commercial?

• Could I achieve the purpose with fictional or less identifying material?

Consent and identity

Public availability does not automatically mean consent for cloning, animation, impersonation, or use in a new context.

Consent should relate to the specific use.

A person should understand, where relevant:

• what will be generated or changed;

• where the media will appear;

• who the audience will be;

• whether the media may be edited;

• whether it may be reused;

• how long it may remain available.

Permission to take a photograph does not automatically include permission to animate the person’s face, clone their voice, or place them in an unrelated situation.

Do not clone or impersonate an identifiable person’s face or voice without clear permission and a legitimate purpose.

Named styles and protected material

Requests to imitate a named creator’s style may raise ethical, contractual, platform, or legal concerns.

A safer alternative is often to describe the visual qualities you need:

“Use muted colours, geometric shapes, soft natural light, and a hand-painted texture.”

Detailed copyright and responsible-creation questions will be examined in Lesson 17.

Minimise personal data

Use only the material needed for the task.

Removing a filename or visible name may not anonymise:

• a recognisable face;

• a distinctive voice;

• a location;

• a document;

• a room or workplace.

Personal data and GDPR will be explored more fully in Lesson 15.

4. Direct the media through prompts and examples

Prompts can guide:

• subject;

• composition;

• camera position;

• lighting;

• colour;

• movement;

• voice;

• pacing;

• duration;

• emotional tone;

• exclusions;

• accessibility requirements.

Reference material may also guide style, structure, movement, or sound.

But prompting controls direction, not truth.

A detailed prompt can produce detailed errors.

For example, generated media may contain:

• impossible anatomy;

• invented writing;

• incorrect equipment;

• historically inaccurate clothing;

• misleading symbols;

• unnatural speech;

• inconsistent objects between frames;

• incorrect captions.

Do not assume that a more detailed prompt creates a more reliable result.

Generation is the beginning of review—not the end of responsibility.

5. Review the output in four layers

A useful review examines four different layers.

Layer 1: Technical review

Check whether the media functions correctly.

For images:

• distorted anatomy;

• broken or merged objects;

• unreadable text;

• inconsistent lighting or shadows;

• unexpected background details.

For audio:

• unnatural pronunciation;

• incorrect names;

• background noise;

• changes in volume;

• missing words;

• unnatural pacing or emotion.

For video:

• inconsistent frames;

• disappearing objects;

• unnatural movement;

• mismatched lips and speech;

• abrupt temporal changes;

• caption timing problems.

Technical polish does not guarantee meaningful accuracy.

Layer 2: Factual review

Check every detail that makes a factual claim.

This may include:

• dates;

• places;

• maps;

• uniforms;

• flags;

• tools or equipment;

• scientific processes;

• quotations;

• historical settings;

• labels and captions.

Compare important details with appropriate sources.

A fictional or illustrative image may still need factual review if audiences will use it for learning.

Layer 3: Meaning and context review

Ask:

• Could the audience mistake this for a real event?

• Does an edit change what appears to have happened?

• Does the music create a misleading emotional frame?

• Does the caption accurately describe the media?

• Has uncertainty been removed?

• Is the media illustrative or evidential?

• Could cropping or editing change the interpretation?

An image can be technically impressive and still communicate a false meaning.

Layer 4: Human-impact review

Ask:

• Is anyone identifiable?

• Was relevant consent obtained?

• Could someone be embarrassed, misrepresented, or harmed?

• Are people shown with dignity?

• Who has expertise, authority, or agency?

• Are any groups reduced to stereotypes?

• Could the same purpose be achieved with less risk?

This review prepares for the deeper lessons on privacy, representation, and responsible creation.

6. Review representation

Generated media may reproduce patterns found in data and common cultural images.

Ask:

• Who is shown?

• Who is absent?

• Who is active, and who is passive?

• Who appears as a leader, expert, worker, victim, or helper?

• Are gender, age, disability, clothing, religion, or culture represented respectfully?

• Does the media treat one group as the default?

• Is diversity meaningful, or only decorative?

For example, if every generated engineer is young, male, and able-bodied, the media presents a narrow picture of who can hold expertise.

Representation review is part of accuracy and dignity, not merely visual variety.

Lesson 16 will examine bias, stereotypes, and representation in greater depth.

7. Design accessibility from the beginning

Accessibility should be planned during production rather than added as an afterthought.

Images

Consider:

• useful alt text;

• readable text;

• sufficient contrast;

• descriptions of meaningful visual information;

• alternatives when colour alone carries meaning.

Alt text should explain the image’s purpose in context, not list every decorative detail.

Audio

Consider:

• a corrected transcript;

• speaker identification;

• descriptions of meaningful non-speech sounds;

• clear volume and pacing;

• accurate pronunciation of names and technical terms.

Video

Consider:

• corrected captions;

• speaker identification;

• a transcript;

• audio description when important visual information is not spoken;

• readable on-screen text;

• controls or alternatives for unnecessary movement or flashing.

Automatic captions, transcripts, and alt text are drafts. They may mishear names, accents, or specialist vocabulary and may miss the educational purpose.

A person should review them before publication.

8. Disclose material AI involvement

Disclosure should help audiences understand what they are seeing or hearing.

It is especially important when AI materially affects:

• meaning;

• identity;

• realism;

• evidence;

• authorship;

• audience interpretation.

A meaningful notice should explain:

1. that AI was used;

2. what was generated or materially changed;

3. whether the media is illustrative, fictional, reconstructed, or evidential;

4. what the audience might otherwise misunderstand.

Examples:

“AI-generated concept illustration. Not a photograph of a real event.”

“Synthetic narration using a fictional voice.”

“Original photograph with an AI-generated background. The person was not photographed in this location.”

“AI-generated reconstruction based on the sources listed below. Some visual details are speculative.”

Disclosure should be proportional

Minor assistance, such as checked noise reduction or caption drafting, may not require the same prominent notice as identity cloning or generation of a realistic event.

The key question is:

Would knowing about the AI involvement change how the audience interprets or trusts the media?

Place important disclosure where the audience encounters the work. Hidden metadata alone may disappear when media is copied, cropped, or reposted.

Disclosure is not permission

Labelling unsuitable media does not make it responsible.

A disclosure notice does not excuse:

• missing consent;

• false information;

• privacy violations;

• harmful stereotypes;

• deceptive use;

• unsuitable publication.

9. Record provenance and decide

Provenance is information about how media was created, transformed, reviewed, and published.

Use a practical production log:

Stage	Information
Purpose and audience
Tool and model, when known
Inputs and references
Permissions and licences checked
Main prompts or instructions
Significant AI transformations
Human edits
Technical and factual review
Meaning and human-impact review
Accessibility additions
Disclosure notice
Final decision	Publish, revise, or stop

Do not reproduce unnecessary personal or confidential information in the log.

Publish

Publish when:

• the purpose is legitimate;

• important permissions are in place;

• relevant reviews are complete;

• accessibility has been considered;

• disclosure is adequate;

• remaining risks are acceptable.

Revise

Revise when the idea is suitable but problems remain with:

• technique;

• facts;

• context;

• representation;

• accessibility;

• consent;

• disclosure.

Stop

Stop when:

• necessary consent is missing;

• important inputs cannot be used responsibly;

• critical facts cannot be checked;

• likely harms cannot be reduced;

• the media remains misleading despite labelling;

• the purpose itself is unsuitable.

For public media, the plan may also include how to correct or withdraw the work if a serious problem is discovered later.

Classroom activity: Responsible media prototype

Create or plan one small AI-supported image, audio clip, or short video using permitted, non-sensitive material.

If a suitable generation tool is unavailable, create a detailed storyboard or production plan instead.

Step 1: Complete the production brief

State:

• purpose;

• audience;

• message;

• format and platform;

• required accuracy;

• possible misunderstanding;

• accessibility needs;

• stop condition.

Step 2: Identify inputs and permissions

List the important source material, references, people, identities, and permissions involved.

Use fictional or non-identifying material where possible.

Step 3: Create the first version

Save the main prompt or production instructions.

Step 4: Conduct the four-layer review

Record findings from:

• technical review;

• factual review;

• meaning and context review;

• human-impact review.

Step 5: Create a revised version

Correct the most important problems.

If revision cannot reduce the risks sufficiently, choose Stop.

Step 6: Add accessibility

Create at least one appropriate element:

• alt text;

• corrected captions;

• transcript;

• audio description;

• accessible text alternative.

Step 7: Write the disclosure

Explain material AI generation or transformation clearly.

Step 8: Complete the provenance log

Record the important tools, inputs, changes, checks, and decisions.

Step 9: Decide

Choose:

• Publish;

• Revise;

• Stop.

Explain the reason.

Step 10: Reflect

Write approximately 300–500 words explaining:

• why the media was created;

• what AI generated or changed;

• what the four-layer review found;

• which problem mattered most;

• how representation and accessibility were considered;

• what the disclosure tells the audience;

• why you chose to publish, revise, or stop.

Reflection questions

1. What is the difference between generated, AI-edited, and AI-assisted media?

2. Why should purpose and audience be defined before generation?

3. When might realism be misleading?

4. Why does public availability not automatically provide consent?

5. What can prompts control, and what can they not guarantee?

6. What is the difference between technical and factual review?

7. How can music, captions, or cropping change meaning?

8. What should representation review examine?

9. When is disclosure especially important?

10. What would make you revise or stop publication?

Key vocabulary

Generative media:
Images, audio, video, animation, or mixed media created or materially transformed with an AI model.

Generated media:
Media largely created by an AI model from prompts or examples.

AI-edited media:
Existing media materially changed by AI.

AI-assisted media:
Media in which AI performs a limited supporting function.

Production brief:
A plan describing purpose, audience, format, accuracy, risks, accessibility, and stop conditions.

Consent:
Permission for a specific use, including relevant information about the transformation, audience, platform, and reuse.

Technical review:
Checking whether media functions correctly and contains production defects.

Factual review:
Checking whether details and claims are accurate.

Meaning and context review:
Checking how media may be interpreted and whether editing changes its apparent meaning.

Human-impact review:
Checking identity, consent, dignity, representation, privacy, and foreseeable harm.

Disclosure:
A clear notice explaining material AI generation or alteration.

Provenance:
Information about the tools, inputs, transformations, human edits, reviews, and decisions behind media.

Summary

In Week 13, we learned that responsible generative media requires more than an attractive result.

The production process is:

Brief → inputs and permissions → generation or editing → four-layer review → accessibility → disclosure → provenance → publication decision

AI may generate new media, materially edit existing media, or assist with limited production tasks.

Prompts can guide direction, but they cannot guarantee truth, permission, dignity, or safety.

Before publication, review the media technically, factually, contextually, and in terms of human impact. Add appropriate accessibility, explain material AI involvement, and record how the work was produced.

The responsible final decision is not always publication.

It may be:

Publish, revise, or stop.

Lesson 13 Interactive Quiz: AI-generated Images, Audio, and Video

Choose one answer for each question. Then select Check my answers. You will receive feedback for every question. This practice quiz does not collect names or scores.

1. AI removes background noise from a person’s original recording without changing the speech. How should this use normally be classified?

A. Fully generated media B. AI-edited media that changes the meaning of the recording C. AI-assisted media D. Documentary evidence created by AI

2. What should happen before creating an AI-generated historical illustration?

A. Define its purpose, audience, evidential status, accuracy needs, risks, and accessibility. B. Generate several realistic versions and decide on the purpose afterwards. C. Publish the first attractive version and add context if viewers complain. D. Make the image as realistic as possible so it appears historically authentic.

3. A person agreed to record an interview. An organisation later wants to clone the person’s voice for an advertisement. What is required?

A. No further permission because the original recording was voluntary. B. Only a notice saying that the advertisement uses AI. C. Permission only if the advertisement changes the person’s words. D. Clear permission covering voice cloning, advertising, audience, and intended use.

4. A detailed prompt requests an accurate image of a historical event. What can the creator conclude?

A. Every visible historical detail is reliable because accuracy was requested. B. The prompt guides the result, but important details still require factual review. C. Factual review is needed only if the image contains written text. D. Realistic lighting and anatomy prove historical accuracy.

5. A generated image is technically excellent but could be mistaken for a photograph of a real protest. Which review layer identifies this problem?

A. Technical review B. File-format review C. Meaning and context review D. Audio review

6. A set of generated workplace images shows only young men as engineers and older women as assistants. What is the main concern?

A. The media reproduces stereotypes about expertise, age, and gender roles. B. The images use too many different people for a consistent design. C. The images should use stronger colours to make the roles clearer. D. Representation does not matter if the anatomy is technically correct.

7. Which accessibility plan is most appropriate for an educational video?

A. Use automatic captions without reviewing names or technical vocabulary. B. Provide only background music so viewers can focus on the visuals. C. Put all important information in fast-moving on-screen text. D. Correct the captions, provide a transcript, and describe important visual information when needed.

8. Which disclosure is most meaningful for a realistic generated historical scene?

A. “Digital media.” B. “AI-generated concept illustration. Not a photograph of a real event.” C. “Created with modern technology.” D. A hidden metadata tag that viewers are unlikely to see.

9. A creator clones a person’s voice without permission but adds the label “AI-generated.” Is the label sufficient?

A. Yes, because disclosure removes the need for permission. B. Yes, if the label appears at the beginning of the recording. C. No. Disclosure does not correct missing consent or unsuitable use. D. Only the technical quality of the cloned voice matters.

10. A prototype uses an identifiable person, necessary consent cannot be obtained, and replacing the identity would defeat the planned concept. What is the responsible decision?

A. Stop the project in its current form and record the reason. B. Publish it with a stronger AI disclosure notice. C. Remove the person’s name but keep the recognisable face and voice. D. Publish only on social media because expectations are lower there.

Reflection: Imagine one AI-generated media project. What would you include in its brief, four-layer review, accessibility plan, disclosure, and final publication decision?