Images, Audio and Video
Generating and editing pictures, voices, music and video with AI: how the tools work, how to describe what you want, and the consent, copyright and labelling rules that keep you on the right side of people and the law.
A school PIBG needs a poster for its Merdeka fun run by Friday. Ten years ago that meant a stock photo, a borrowed designer or nothing at all. Today a parent can type a few sentences and get a dozen bright illustrations in a minute, add a narrated video for the WhatsApp group, even a cheerful jingle. The same tools can also put words in the headmaster's mouth that he never said. This page is about using the first power well and never using the second.
How text-to-image works, in one paragraph
Most image generators today are diffusion models. During training, a model sees millions of captioned images, each progressively buried in random noise, and learns to predict and remove that noise step by step. To generate, it starts from pure static and denoises it many times, steered at each step by your text, until a picture consistent with the description emerges. The model is not looking anything up or collaging stored photos; it has learned statistical patterns of what "batik", "golden hour" or "watercolour" tend to look like. That is why it is fluent with common styles, shaky with exact text, hands and counts, and why its output reflects whatever its training data over- or under-represented. The story of how this arrived is told in The Generative Era.
Writing image prompts
An image prompt is a brief to an illustrator who cannot ask questions. The more of the following you specify, the less the model fills in with clichés. Tap each tile.
Notice the request for empty space and "no text". Image models still often garble lettering, so the reliable workflow is to generate the picture and add the words yourself in a design tool such as Canva, Google Slides or PowerPoint. Generate several variations, pick the best, then refine with small changes rather than starting again.
Editing, inpainting and outpainting
Editing is often more useful than generating from scratch. Three operations cover most needs:
- Inpainting: select part of an image and describe what should be there instead — remove a stray bin, change a shirt colour, fix a sixth finger.
- Outpainting (or "generative expand"): extend the canvas so a landscape photo fits a portrait banner.
- Conversational editing: many chat assistants now accept an uploaded picture and an instruction such as "make the background plain white" As of 2026.
The same rule applies as with any photo editing, only more so: editing a picture of a real person to show something that did not happen — a fake crowd, a different location, an altered expression in a news context — is deception, however easy the tool makes it. Use editing on your own images, on generated images, or with permission.
Voices, music and video
Text-to-speech now sounds natural in many languages, including Bahasa Melayu and English with local accents, and is built into phones, office tools and specialist services such as ElevenLabs As of 2026. It is excellent for accessibility, narrating slides or producing a training clip in several languages.
Voice cloning creates a synthetic copy of a specific person's voice from a short recording. Use it only with that person's clear, informed and recorded consent, for a stated purpose. Reputable services require you to confirm consent, and some ask the speaker to read a verification phrase. Cloned voices are a favourite tool of scammers: a "call from your son" asking for urgent money. Agree a family safe word, and see Staying Safe with AI.
Music generators such as Suno and Udio produce complete songs, with vocals, from a text description of genre, mood and subject, or from your own lyrics As of 2026. They are fun for a birthday song or a placeholder jingle. Before using a track publicly or commercially, read the service's licence terms: rights often depend on your subscription tier, and major record labels have taken several of these companies to court over the music used in training — some of those cases have since moved towards settlements and licensing deals.
Video generators — examples include OpenAI's Sora, Google's Veo, Runway and Kling — turn text or a still image into short clips, increasingly with synchronised sound As of 2026. Clips are typically seconds long; keeping a character consistent across shots and getting physics right remain hard. Prompt them like a film director: shot type, camera movement ("slow pan left"), action, setting, lighting and duration. Many people will meet AI video through simpler features first: auto-captions, background removal and translating a talking-head video into another language with lip-sync.
Copyright and consent
Two questions are worth separating: may these tools be trained on other people's work, and who owns what you generate?
Debated Whether training on copyrighted images, music and text without permission is lawful is being fought out in courts and parliaments. Getty Images sued Stability AI in the UK and US; artists brought a class action against several image companies; record labels sued music generators. Outcomes so far are mixed and vary by country, and some disputes have ended in licensing deals rather than judgments As of 2026.
- Training is like learning: many AI developers argue that models learn general patterns, as human artists learn from what they see, and that this is fair use or a permitted text-and-data-mining activity.
- Training is uncompensated copying: many artists, musicians, photographers and publishers argue their work was copied at scale to build products that now compete with them, and want consent, credit and payment.
- Licensing as a middle path: some companies train only on licensed or public-domain material, or pay rights holders; critics reply that this favours the largest firms.
On ownership, the clearest official statement so far comes from the US Copyright Office, whose January 2025 report concluded that material generated purely from a prompt is not protected by copyright, while human selection, arrangement and substantial editing can be. Malaysia's Copyright Act 1987 was written long before generative AI and has no AI-specific rule yet, so how it applies to machine-made work is still unsettled As of 2026. Practically: if owning the result matters (a logo, a book cover), expect to add real human creative work — and check your tool's terms of service.
Consent is simpler. Do not generate realistic images, voices or videos of real, identifiable people without their permission, and never sexualised or humiliating ones. Depending on the content, making or sharing such material can be a criminal offence in many countries — in Malaysia, for example, under the Communications and Multimedia Act 1998, the Penal Code or child-protection laws As of 2026 — as well as a cruelty. (A general pointer, not legal advice.) Be especially careful with children's photos.
Watermarks, content credentials and honest labels
How can anyone tell what was made by AI? Three approaches work together, each imperfect:
- Invisible watermarks are woven into the pixels or audio by the generator. Google DeepMind's SynthID (announced 2023) is a prominent example; a matching detector can recognise it even after mild edits.
- Content credentials follow an open standard from the C2PA, a coalition formed in 2021 by Adobe, Microsoft, the BBC and others. A signed record attached to the file says which tool made or edited it. Cameras, editing apps and AI generators increasingly add them As of 2026.
- Platform labels: major social and video platforms ask uploaders to disclose realistic AI content and add "AI-generated" or "made with AI" labels As of 2026.
Consensus None of these is proof. Metadata can be stripped by a screenshot, watermarks can be weakened, and the absence of a label does not mean a picture is real. Detection tools that claim to spot any AI image are unreliable. So the most important label is the one you add yourself.
If a reasonable viewer might take your image, voice or video for a real photograph, recording or event, say it was made or altered with AI — in the caption, the credits or the file name. Illustrations that are obviously illustrations need less; anything realistic, news-like or showing real people needs more.
Your generated poster has garbled lettering. What is the most reliable fix?
Lettering is a known weak spot; doing text yourself is faster and exact.
A colleague offers to clone the CEO's voice for an internal video "as a surprise". What is missing?
Cloning a real person's voice needs their clear permission for the specific use.
An image has no content credentials. What can you conclude?
Credentials can be stripped by a screenshot; their absence proves nothing.
// Pause and reflect
- Which recurring task — posters, slides, product photos, narration — could media tools save you time on this month?
- Whose face, voice or creative work appears in what you might generate or edit, and have they agreed?
- Where would you put an "AI-generated" label on your next piece, and what wording would you use?
Your notes are saved privately in this browser.
Pick a real need — an event poster, a slide illustration, a social-media header. Write one vague prompt and one using all six tiles, and generate both in a free image tool. Then make one inpainting edit. Add the text yourself, and finish with an honest caption saying how the image was made.
Key terms on this page
Sources & further reading
- Ho, J., Jain, A. and Abbeel, P., “Denoising Diffusion Probabilistic Models” (2020).
- Rombach, R. et al., “High-Resolution Image Synthesis with Latent Diffusion Models” (2021) — the basis of Stable Diffusion.
- US Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (January 2025).
- Coalition for Content Provenance and Authenticity, C2PA technical specification and explainer (accessed 2026).
- Google DeepMind, SynthID (accessed 2026).
- Laws of Malaysia, Copyright Act 1987 (Act 332).