Good Slop, Bad Slop: How to Question AI-Generated Images

A practical guide to checking synthetic images through context, provenance and corroboration rather than one magic visual clue.

The short version

Some images are made by cameras. Some are made by people using drawing tools. Some are made or altered with generative AI. Many are a mixture of all three.

The first question should not be:

“Does the person in the image have six fingers, or is there something else that looks odd?”

It should be:

“What is this image trying to make me think or do, and what can I check?”

Visual clues can be useful. They are not a verdict. Some websites and apps offer an “AI detection” score, but these tools can misclassify difficult examples. For most people, checking the source, context and claim is more useful than looking for a magic percentage.

That is why our Good Slop, Bad Slop test is not a competition to catch a machine. It is a habit of slowing down before an image makes us angry, impressed, frightened or ready to share.

“AI-generated” is not one thing

An image may be:

• entirely generated from a text prompt;
• a real photograph with an AI-assisted edit;
• a human illustration with AI-assisted cleanup;
• a collage combining photographs, stock elements and generated objects;
• a real image reposted with a false caption;
• a synthetic image that has been cropped, filtered or compressed until its original clues disappear.

These cases matter for different reasons. A playful generated picture of a talking dog is not the same as a fabricated photograph presented as evidence of a real event. The tool used to make an image is only one part of the question.

Why visual inspection is not enough

Older examples of generated images often contained obvious problems: strange hands, unreadable signs or objects that melted into one another. Those clues can still be worth noticing, but they are not a reliable universal test.

An image may look strange because it was generated. It may also look strange because it was badly compressed, cropped, edited, photographed in poor light or deliberately manipulated by a human.

Research presented at the 2025 International Conference on Learning Representations tested nine automated image-classification tools against a challenging dataset and reported that almost all of the detectors misclassified AI-generated images as real. The researchers concluded that reliable, general image detection remains unsolved.

That does not make every automated tool useless. It means a result is one signal among several, not a laboratory certificate. A confident percentage on a website should not replace checking where the image came from or whether the claim around it is supported.

The Good Slop, Bad Slop test

Use these five questions together.

1. What is the claim?

Describe the image without repeating its caption.

Is it supposed to show a real event, a person, a product, a place, a scientific result or simply an idea?

The more important the claim, the more evidence it needs.

2. Who first published it?

Look for the earliest credible source you can find. A repost, screenshot or anonymous account may not be the origin.

Check whether the account has a history of satire, advertising, activism, impersonation or recycled material. None of those automatically makes an image false, but they affect how carefully we should interpret it.

3. Is there context outside the image?

Search for related reporting, original photographs, location details, dates, weather, event records or statements from people who were present.

Reverse-image search can help find earlier versions, but a search result is a lead rather than proof. Absence from a search engine does not establish that an image is fake.

4. Is provenance available?

Some files carry metadata or Content Credentials that record aspects of an asset’s origin and editing history. The Coalition for Content Provenance and Authenticity (C2PA) describes Content Credentials as a cryptographically bound record of provenance that can include the origin of an asset, modifications and use of AI.

That information can help, but it is not a truth machine. C2PA explicitly says that provenance does not by itself establish whether an image is accurate or truthful. Credentials may be absent, incomplete or removed during editing and reposting.

5. What would change if it were false?

Before sharing, ask what the image is trying to make you do.

Does it encourage you to donate, buy, panic, accuse someone, vote, join a pile-on or click a link? The emotional or commercial consequence is a reason to slow down, not proof of deception.

A calm visual check

Zoom in if useful, but treat visual inspection as a prompt for questions rather than a final answer.

Look for:

• text that changes shape or cannot be read consistently;
• reflections or shadows that do not match the scene;
• repeated faces, objects or textures;
• impossible joins between objects;
• logos, uniforms or signs that do not match the location;
• a caption that claims more than the image itself shows.

Then ask whether a real photograph could contain the same oddity. A blurry camera image can have defects. A staged photograph can be misleading without any AI. “Looks fake” and “is false” are not interchangeable statements.

What to teach young people

Avoid turning this into a hunt for magical clues. A young person who learns that every fake image has strange fingers may become overconfident just as synthetic images become more convincing.

Instead, practise a three-step conversation:

1. Notice: What caught your attention?
2. Question: What does the image claim, and who benefits if we believe it?
3. Check: What independent evidence could support or challenge the claim?

This keeps curiosity alive without treating every unusual image as suspicious. It also works for ordinary advertising, influencer posts, memes, AI-assisted schoolwork and real photographs used with the wrong caption.

If the image may cause harm

Do not download or redistribute illegal, intimate or restricted material just to investigate it. If someone is being targeted, preserve only the minimum lawful evidence needed, avoid amplifying the image, and use the appropriate platform or official reporting pathway.

BARK can help organise a factual record: where you saw the image, when you saw it, what it claimed, what action you took and which official route may apply. Puppy vs Algorithms does not authenticate the image, investigate the people involved or adjudicate the complaint.

Rosie’s rule of paw

A tool can offer a clue. It cannot do your thinking for you.

The safest response to a dramatic image is not instant belief or instant dismissal. It is a pause, a question and one more piece of evidence.