Create and assess generated media

What can I create with generated media, and what can I establish about media I receive?

Use a generator for an explicit creative task. To assess an image, recording or video, check its source and the claim attached to it, then treat detection and content-history tools as evidence with a limited scope.

Scope

Generated images, audio and video · Worked examples for students

Reviewed

5 October 2026

In brief

• A convincing image can still fail your instructions or depict an event that never happened.
• Signed content history and watermarks answer narrower questions than whether a claim is true.
• A missing mark or detector result cannot settle the origin of every file.

Start with a useful creative task

A student club needs an illustration for a robot-repair workshop. A generator can help explore compositions before someone makes the final poster. The task is quite different from creating a photograph that appears to document a workshop that has already happened.

Write the intended use before the prompt: an illustration for an advertised event, with the date and venue added separately in a layout tool. Specify what the illustration needs to show and what would make it unusable.

Illustration for a student robot-repair workshop poster.
Show three students at one workbench, inspecting a small wheeled robot.
Leave empty space at the top for a title that I will add separately.
Do not include logos, text labels or a claim that this is an event photograph.

This is an illustrative prompt, not a tested recipe. Inspect the result against the brief: count the people, check that the robot and tools make sense, and decide whether the image helps explain the event. Revise a specific failed requirement instead of repeatedly asking for something "better".

Use people, recordings and reference images only when you have the necessary permission and rights. Check the output and publication terms of the chosen tool. The rules and rights guide explains why a generated result does not remove those questions. Follow the applicable assessment or publication rule for disclosure; a technical mark in the file does not replace a required declaration.

How diffusion makes an image

A basic diffusion model learns from a process that adds noise to training images. It then learns transitions that reverse that process. In generation, it starts with noise and applies learned steps to produce a sample. These are two different stages: training teaches the model statistical structure; generation uses the trained model. Ho, Jain and Abbeel, 2020

Training example:   image → partly noisy image → increasingly noisy image
Generation:        initial noise → repeated learned updates → generated image

The second line does not mean there is an original photograph hidden in the noise. The model generates a sample using what it learned. In the basic process, updates affect the image representation across several steps, rather than appending visible pixels from left to right like letters in a sentence.

Some systems perform diffusion in a compressed representation of an image, then decode it into pixels. Text can condition the process through a learned connection between the text representation and the image generation steps. That gives a prompt influence over the result, rather than a guarantee that every instruction will be satisfied. Rombach et al., 2022

Image generators do not all use one recipe

Mechanism, not a product claim

Diffusion is one family of methods. Other generators can produce sequences of image tokens. Audio and video tools also vary in what they generate and how they represent it. Use the chosen model's documentation before claiming that a named product follows this diagram. The diagram does not represent a measured run, a fixed number of steps or every current media generator.

For the workshop illustration, changing the prompt can change the composition. It does not make the image a record of an event. A useful creative output and reliable evidence are different requirements.

Keep four kinds of evidence separate

Suppose an image shared in a group chat claims to show a fire at a campus building. Before asking whether the image looks generated, identify the claim: which building, on what date, and according to whom?

Scroll horizontally to view all columns.

What each check can establish

Check	Useful evidence	Limit
Original source and independent reports	Who published the image, when, with what account of the event, and whether other sources support it.	A genuine photograph can have a false date, location or caption.
Visual inspection	Details that require further checking, such as inconsistent labels or geometry.	A visual oddity is not proof of generation; a convincing appearance is not proof of camera origin.
Detector	A classification produced by a method tested under particular conditions.	Its tests may not cover this generator, edit, compression level or type of media.
Signed history or watermark	Declared creation/editing information, or a signal added by a supported generator.	Coverage, removal, compatibility and trust in the signer matter; the depicted claim still needs checking.

These checks can support each other, but they do not all answer the same question. Finding an older copy can show that the current caption is misleading even if you cannot establish how the pixels were created.

Why appearance and detectors can fail

In a 2022 study, participants struggled to distinguish particular StyleGAN2-generated faces from real face photographs. That result supports caution about judging origin from appearance. It does not provide one human detection accuracy for all photographs, videos, audio, generators or people. Confidence in a judgement is also not a measurement of its correctness. Nightingale and Farid, 2022

Detector performance can change when the generator changes. A CVPR 2023 study showed that the standard classifier it examined, trained on images from a GAN, could fail on unseen diffusion and autoregressive generators. The authors proposed a method that generalized better in their experiments. Neither result certifies a detector for every file found online. Ojha, Li and Lee, 2023

Before relying on a detector, look for its tested media, generator versions, real-image sources, edits and compression conditions. Check both kinds of error: real content labelled generated and generated content missed. Ask whether a displayed score is calibrated for your situation. A high-looking number without that information is hard to interpret.

For an assessment, a detector score also does not establish who made the work, what assistance was permitted, or whether misconduct occurred. See verification and declaration for those separate checks.

Signed content history is not a truth test

Content Credentials use the C2PA standard to attach cryptographically signed information about a file's creation or editing. Verification can establish whether declared information is still bound to the asset and who signed it. The standard's explainer explicitly separates that history from whether the content is true, accurate or factual. C2PA explainer, version 2.2

A camera can record a staged event. An AI image can correctly declare that it was generated. A valid signature helps you assess the declared history; you still need to judge the claim and the signer.

Embedded metadata can be removed. C2PA also supports recovery mechanisms using fingerprints or watermarks to find a stored credential. That does not ensure that every platform or verification tool will recover it. If no credential is available, record that as unavailable evidence. Do not turn it into a verdict about the image's origin. C2PA, section 7.2.3

What a watermark check tells you

Watermarks embed a signal directly into content. They differ from metadata stored alongside it. Google's SynthID is a named example: its official documentation describes marks added to generated content and checks for supported Google AI media. Google DeepMind: SynthID

The Gemini verification help page, checked on 5 October 2026, says its SynthID check recognizes Google AI content and describes limits for small edits, low-detail content and repeated alterations. A detected mark may relate to only part of a file. A missing mark cannot establish camera origin, including because a different system may have created the content.

This is provider documentation about a particular tool, not independent validation of all watermarking. Before uploading a recording or image, also check whether it contains private information and whether that service is an appropriate processing route. See processing routes and data.

Legal duties depend on the actor, use and applicable date. The current transparency rules distinguish machine-readable marking from disclosure to people. Do not assume a watermark makes every publication lawful or fulfils every disclosure requirement.

Work through an uncertain image

Return to the campus-fire example. This is a hypothetical case; the page does not present a real event or a measured detection result.

1. Save the URL and identify the precise claim. Separate the image from its caption and the person forwarding it.
2. Find the original publisher. Search for an earlier copy and check the claimed place and time against reliable reports.
3. Inspect available content history. Record the signer, declared tools and verification status, rather than just the presence of an icon.
4. If you use a detector, record what it tested and whether this file falls within that scope.
5. State what the evidence supports and what remains unknown. Choose an action that fits that uncertainty.

A defensible conclusion could be: "I found the same image in an older post about another location, so it does not support today's claim. I have not established whether it was originally generated." That is an example of how to express a conclusion, not a report of an investigation performed for this guide.

For a voice message asking for money, do not wait for an origin verdict before deciding how to act. Contact the person through a separate channel you already know. A familiar voice does not settle whether the request is authorized.

Try a creation and checking exercise

Choose an illustration you could actually use for a student event or study explanation. Before using a tool, write its purpose, three required features and one reason you would reject the result. Compare any output with that list. Add labels in a tool where you can edit them directly, and check the final published version.

Then write two sentences: what the image is intended to communicate, and what it cannot serve as evidence for. Record the tool and the material changes you made if your publication or assessment instructions require it. You can do this exercise by assessing an existing permitted image; access to a paid generator is not required.

Source notes and dates

For recommendation systems, unequal treatment and resource costs, continue with bias, recommendations and costs.

Sources checked on 5 October 2026. Research results concern their stated model, dataset and evaluation period. Product documentation can change.

• Ho, Jain and Abbeel: Denoising Diffusion Probabilistic Models, 2020. Basic diffusion training and sampling; not a specification of every current generator.
• Rombach et al.: High-Resolution Image Synthesis with Latent Diffusion Models, 2022. Compressed image representations and conditioning.
• C2PA Content Credentials explainer, version 2.2. Signed assertions, truth limits and removable metadata.
• Nightingale and Farid, PNAS, 14 February 2022. Specific face-perception experiments, not a universal detection rate.
• Ojha, Li and Lee, CVPR 2023. Detector generalization experiments across selected generators.
• SynthID documentation and Gemini verification help. Provider descriptions of a supported watermark check and its limits.