IntegrAuth Academy
Identity, security and practical AI — free, from zero. Byte-sized lessons, real diagrams, hands-on labs, zero fluff. Two programs, each with its own exam and certificate.
English is hard? Your browser can translate this page into Tamil, Telugu, Kannada, Malayalam or Hindi. Tech words stay in English so you learn the real names.
How?
- Chrome on phone: ⋮ menu → Translate
- Chrome on computer: right-click → Translate to …
- Edge: the translate icon in the address bar
Images, voice and video
Priya needs a banner, a voice-over and 5 interview transcripts. AI can do a first try. Your goal: see how these tools work, and where they fail.
🧪 Interactive lab — enable JavaScript to play with this one.
- Many image tools start from random dots, like TV static. Your prompt steers each clean-up step.
- A good image prompt names the subject, style, framing, light and frame shape.
- Check text, counts and hands in AI images. A made image is never proof.
- A transcript can include words nobody said. Check quotes against the recording.
- Copy a real person's face or voice only with their Consent.
- Urgent call from family asking for money? Hang up. Call back on a number you already have.
How AI makes a picture
First, the model learns. It sees millions of images, each with a caption. Random noise is added to each image, step by step, until only static is left. At each step, the model learns to guess the noise. Then it can take the noise out.
To make a new picture, it runs backwards. It starts from pure random static. It removes a little guessed noise at a time. Your prompt steers every step. This kind of tool is a Diffusion model.
Voice tools
Some tools turn speech into text. This is Speech-to-text. It can mishear noise, accents and names. It can even add whole phrases nobody said.
A tool can also copy a real person's voice from recordings. This is a Voice clone. Scammers use cloned voices in phone calls (deepfakes and AI scams).
Now take a pretend urgent call.
🧪 Interactive lab — enable JavaScript to play with this one.
Go deeper: seed, steps and guidance
- Seed: the number that picks the starting static. Same model, prompt, settings and seed usually give the same image. A new seed gives a new image.
- Steps: how many times it removes noise. More steps usually means more detail and a longer wait.
- Guidance: how strongly each step follows the prompt. Too low, and the image drifts from your words. Too high, and colors get harsh, with less variety.
Not every tool uses diffusion. Some build an image piece by piece, the way a language model writes text. This is called Autoregressive generation (what an AI model is). In March 2025, OpenAI said the image tool in ChatGPT worked this way. The same prompt tips work either way.
A model learns patterns, not stored pictures. But it can sometimes copy a training image closely.
Go deeper: prompts, edits and weak spots
"A lighthouse" leaves every choice to the model. Say more:
- Subject: a red-and-white lighthouse on a cliff
- Style: flat vector poster
- Framing: lighthouse on the left, empty sky on the right
- Light and color: at dusk, warm orange sky
- Frame shape: 16:9 for a banner, 1:1 for a profile image
Change one thing at a time.
To fix one part, mark that area and describe what should go there. This is Inpainting. Some chat edits redraw the whole picture, so compare before and after.
Weak spots, as of September 2026: text in images, counting, exact layouts and hands. Anything you don't say gets filled with what was common in training. So "a doctor" may mostly come out as a man (bias and fairness).
Video tools, such as the one in Google's Gemini app, Runway and Kling, make short clips. Movement can look wrong, and faces can change between shots. Music tools such as Suno and Udio make whole songs, singing included. Who owns the output is a hard question (who owns AI output?).
Go deeper: more on voice
A 2024 study tested a popular speech-to-text model. About 1% of its transcripts had whole phrases that were not in the audio. Will you quote it, or put it in a medical or legal record? Check those lines against the recording.
The reverse tool reads text aloud. This is Text-to-speech.
A voice clone can need very little audio. In March 2024, OpenAI said its tool could copy a voice from a 15-second sample. Some services allow their best clones only of your own voice, after a check. In the US, robocalls with AI voices generally need your consent first.
Go deeper: labels and credentials
Could people think it is a real recording? Then say it is AI-made. Sites like YouTube require this. Since August 2026, the EU AI Act does too, for deepfakes. India's IT Rules, in force since February 2026, ask big platforms to label AI-made media.
A file can carry a signed record of how it was made and edited. These are Content Credentials (C2PA). A screenshot or a website can remove them. So a missing credential proves nothing. And a credential shows where a file came from, not whether it is true (deepfakes and AI scams).
Want to use it for business? Check the tool's license first.
Everything you upload is data: a photo, a voice, your face. Check what the tool may keep and learn from (safe AI at work).
Open Diffusion Explainer. Watch noise turn into a picture as you change the seed.