Images, Audio, and Multimodal AI

Generative AI is no longer only about text. Beginning around 2022, tools that turn a written prompt into an original image reached the public, and systems that clone or synthesize a voice followed close behind. The same models increasingly handle several kinds of input and output at once, changing what you can make and what you can trust. A photograph or a voice recording used to be fair evidence that something happened; once a convincing image or voice can be generated on demand, that assumption weakens for everyone. Grasping roughly how these systems work is now part of basic media literacy, not a specialist skill.

From noise to picture: diffusion, sketched plainly

Most modern image generators use an idea called diffusion. In training, the system sees countless images progressively corrupted with random visual noise and learns to reverse the process — to predict what a slightly cleaner version would look like. To generate, it starts from pure noise and refines it step by step, while a text prompt steers each refinement toward whatever the words describe. No mathematics is needed to grasp the shape of it: begin with static, and repeatedly ask "what would make this look a little more like the prompt?" until a picture emerges.

The same idea, in sound and across modes

Voice and audio synthesis follow the same logic in another medium. A voice model learns the statistical texture of speech — timbre, rhythm, how sounds flow — and generates new audio that is plausible given a script or a short voice sample. A multimodal model goes further, handling more than one kind of data together: because these modes share one internal representation, it might read an image and answer a question in words, or turn text into a picture, moving fluidly between them. But the thread never changes — predict a plausible next piece, whether a word, a region of pixels, or a slice of sound. Fluency across modes is not comprehension.

Deepfakes and provenance

Because these systems produce convincing media, they enable deepfakes — synthetic images, video, or audio built to look like a real person or event. The defense is not a sharper eye; realistic fakes are designed to pass inspection. The defense is provenance: where the file came from, who published it first, whether a trustworthy source corroborates it. The useful question shifts from "does it look real?" to "can I trace where it came from?"

Type "a red fox reading a newspaper in a library, soft light." The generator fetches no photo. It starts from random noise and refines it over many prompt-guided steps until pixels resembling a fox, a newspaper, and warm light emerge. The result can look photographic while depicting something that never existed. That is the point to hold onto: a generated image answers "what would a plausible picture of these words look like?" — not "what really happened?"

Not every AI-made image is a deepfake. A designer mocking up a fantasy book cover, or a studio using voice synthesis to fix one line of an audiobook without re-recording, is producing synthetic media openly and honestly. The harm in a deepfake comes from passing synthetic media off as an authentic record of a real person or event. Synthetic is not the same as deceptive; the deception lies in the false claim of authenticity, not in the tool.

It is well documented that around 2022 several text-to-image systems became widely available to non-experts, growing quickly from a research novelty into everyday consumer tools. Without over-claiming any single product's specifics — model names and quality change constantly — the reported pattern is clear: prompt-driven image generation moved from labs into public hands within a very short window, and debate about consent, artists' work, and misinformation followed almost immediately. The same period saw voice-cloning tools grow good enough that short samples could yield convincing imitations. The durable lesson is the shape of the shift, not any benchmark: convincing synthetic media became cheap and accessible faster than shared norms for handling it.

• "The generator searches for and pastes existing images together." It refines noise toward a plausible result using learned patterns; it is not a collage search.
• "If it looks real, it is real." Realism is exactly what these systems optimize; appearance is no longer evidence of authenticity.
• "Deepfakes only threaten celebrities and politicians." Voice and face synthesis can target ordinary people too, in scams and harassment.
• "Multimodal AI understands what it sees and hears." It relates and generates across modes fluently, but fluency is not comprehension.

You are shown a striking image and an audio clip attributed to a public figure. Decide what you would check, and in what order, to establish provenance before believing or sharing either.

Think Like a Maester: When media can be generated on demand, "does it look real?" stops being a useful test. Ask where it came from and who can vouch for it — provenance, not appearance.

1. In plain terms, how does a diffusion-based image generator turn noise into a picture?
2. What does it mean to call a model "multimodal"?
3. Why does convincing synthetic media weaken photos and recordings as evidence?
4. What distinguishes honestly-made synthetic media from a harmful deepfake?
5. Why is checking provenance a better defense than trying to spot fakes by eye?

Generative AI reaches well beyond text. Diffusion-based image generators begin with random noise and refine it toward a picture that fits a prompt; voice systems do the equivalent for sound; and multimodal models move across text, images, and audio through one shared representation. The unifying idea is the same as in text generation — predict a plausible output, with no grasp of truth. The public arrival of these tools around 2022 made convincing synthetic media cheap and widespread, which is why deepfakes matter and why the trustworthy response is to verify provenance rather than to trust appearances.

• Spotting Misinformation and Disinformation
• Using Generative AI Responsibly