Can AI make videos and music today?

A year ago, video from a text prompt was something you read about in research papers from large labs. This summer, anyone can try it. In early June, the AI startup Runway opened Gen-2 to the public after a waitlist period, and TechCrunch described it as one of the first text-to-video models you can actually buy access to. On August 2, Meta released AudioCraft, a set of models that write music and sound effects from a short description, with the code open for anyone to study.

So the honest answer is yes, with limits. Today's tools make short clips and short pieces of music. They are good for sketches, mood pieces, social posts and storyboards. They are not yet good at long scenes, consistent characters or a song with verses and a chorus that hold together.

Video: what Gen-2 can and cannot do

Gen-2 is a diffusion model: it starts from random visual noise and removes it step by step until the frames match your prompt. You can start from text alone or from an image you upload. According to Runway's research paper, its training data held 6.4 million video clips plus 240 million still images.

In TechCrunch's hands-on review, each clip lasted four seconds and the frame rate was low enough to look almost like a slideshow at times. The reviewer also saw grainy footage, arms and objects that melted into each other, and prompts where a requested camera move, such as a slow zoom, simply did not happen. Styles like anime and claymation hid those flaws best, because they already look a little jerky. Runway screens prompts with a mix of automated and human moderation to block violent, pornographic or copyright-violating content.

Music and sound: AudioCraft

AudioCraft has three parts. MusicGen writes music from text, AudioGen makes sound effects such as a barking dog or footsteps on a wooden floor, and EnCodec compresses audio into small pieces, called tokens, that the models can work with, a little like the word pieces a language model reads. Meta says MusicGen learned from about 20,000 hours of music that Meta owns or licensed for the purpose, and it notes that the collection leans toward Western styles. The models are released for research, with the code under the open MIT license, so most people will meet them through demos or apps that other developers build on top.

A worked example: a 12-second clip for a café

Imagine a small café that wants a short looping video for its social media page, with music underneath. Here is a workflow that fits today's tools.

1. Plan the shots in ChatGPT. Ask for three four-second shots, each described in one sentence with the subject, the setting, the light and the style. Keep the style identical across all three so the clips match.
2. Generate each shot in Gen-2. For example:

Close-up of steam rising from a cup of coffee on a wooden table, morning light from a window, soft focus background, warm colors, claymation style

Make a few versions of each shot and keep the best. Expect to discard most of them.

1. Generate the music. Describe the mood, the instruments and the tempo rather than naming an artist:

Calm acoustic guitar and soft brushed drums, slow tempo, warm and relaxed, suitable for a morning café

1. Edit them together. In any video editor, trim the three clips, add a simple fade between them, lay the music underneath, and add your café's name as text. Text inside the generated frames is usually unreadable, so add words yourself.
2. Check before posting. Watch for melted hands, odd objects and anything that looks like a real brand or person.

Why the results still feel rough

Video is harder than images because every frame must agree with the frames around it. Music is hard for a similar reason: Meta points out that a few minutes of recorded music at standard quality contains millions of sample points, and the model has to keep long patterns, like a melody coming back, consistent across all of them. That is why today's best results are short. For now, the useful skill is the same one that works with image generators: describe the subject, the setting, the light and the style precisely, generate several versions, and treat the output as material for your own editing.

What are the copyright risks?

There are two separate questions: whether you can own what the AI makes, and whether what it makes borrows too much from someone else.

On ownership, a U.S. federal court gave a clear answer on August 18. In Thaler v. Perlmutter, Judge Beryl Howell upheld the Copyright Office's refusal to register an image that its owner said a computer system had made entirely by itself. The judge wrote that "human authorship is a bedrock requirement of copyright." The ruling covers works made with no human involvement; the judge also noted the harder questions ahead, such as how much human input makes someone the author of an AI-assisted work. Until those questions are settled, your own edits, selection and arrangement are the parts with the clearest claim to be yours.

On borrowing, the risk sits mostly in the training data and in your prompt. Meta says it built MusicGen on music it owns or licensed; other tools may not say what they learned from. The safe habits are simple: never prompt for a named artist's song or a famous character, keep a record of your prompts and edits, read each tool's terms of use before commercial use, and ask a lawyer before you build a business on generated media.