Recognising AI slop: how to spot AI images and text

Humans produce rubbish – including on the internet. That has always been the case. But with the rise of powerful AI tools, this tendency has taken on a new dimension: images and text can now be generated in seconds, at virtually no cost and in enormous quantities. Fortunately, the most important AI detector is still the one between your ears.

From junk mail to AI slop

Low-quality content has always existed – from trashy literature and junk mail to AI-generated material. Since the emergence of powerful AI tools, however, the possibilities for mass-producing images and text in particular have increased dramatically. Absurd content repeatedly circulates online – such as the countless variations of “Shrimp Jesus” that appeared in 2024.

A memorable term has become established for this kind of digital rubbish: “slop” – a word associated with mush, animal feed or simply waste. Humans produce it online both deliberately and unintentionally, particularly when tools are easy to use and generate amusing results.

AI slop arises above all when content is released into the world without care, contextual verification or subject-matter expertise. The tricky part is that AI-generated images and text often look deceptively convincing. Inexperienced users may then neglect the duty of care that should always accompany the use of AI – or even begin to question their own prior knowledge. One clarification is important: not every piece of AI-generated content is automatically slop. Used properly, AI provides valuable support in analysis, drafting text or simply structuring ideas. The question is therefore not whether you use AI, but how you use it.

Three characteristics of AI content

What typically distinguishes AI-generated content? Cognitive scientist Cody Kommers and colleagues have identified three fundamental characteristics:

Asymmetric effort: Creating the content costs hardly any time or money. Understanding, contextualising and disproving generated material, by contrast, can require considerably more effort.

Mass producibility: AI content is part of an ecosystem of continuous mass generation and distribution that never truly comes to rest. Sometimes sources or watermarks are provided. Often, however, provenance is missing altogether.

Surface competence: At first glance, everything looks convincing – an image appears photorealistic, a text reads fluently. Yet behind it there may be neither craftsmanship nor any genuine communicative intention.

Researchers including teams from Imperial College London and Stanford University estimated that by mid-2025 more than one third of all newly published web pages already contained AI-generated or AI-assisted content – and the trend was rising. What figure will we see by the end of 2026? It is therefore time to train our ability to recognise AI content more effectively.

Spotting AI-generated images: four indicators

1. Backgrounds and text: When generating images, AI tends to prioritise the centre of the composition. What happens in the background often receives less attention – resulting in buildings with incorrect windows, signs containing unreadable characters and invented lettering.

2. Physics, light and reflections: Shadows may not match the light source, people and objects may be illuminated inconsistently, and mirrors may remain strangely empty. Reflections continue to be among the persistent weaknesses of current AI image models.

3. The devil is in the detail: Limbs, jewellery and teeth frequently give AI-generated images away – and the list of possible errors is considerably longer. A careful second look is always worthwhile.

4. The uncanny valley: If an image triggers an instinctive sense of discomfort, that can be a useful signal. The human brain is highly sensitive to inconsistencies in faces: tiny asymmetries, unnatural skin textures or an empty gaze. It adds up these micro-inconsistencies and raises the alarm.

Spotting AI-generated text: four indicators

Text also contains useful clues that can help identify AI-generated material:

1. Hardly any errors, but little originality: Flawlessly written text is not necessarily a sign of quality. A certain amount of “messiness” – unusual phrasing, awkward wordplay, unexpected turns – is part of human writing. It is also worth becoming suspicious when a text covers every aspect of a topic equally without ever becoming more concrete or adopting a distinct point of view.

2. A consistently optimistic underlying tone: The researchers from Imperial College London and Stanford University found that AI-generated texts resemble one another more closely than human-written texts and tend to display a conspicuously positive tone. Human writing, by contrast, shows a broader emotional range and occasionally creates friction.

3. Uniform sentence structure: Two metrics are often used when analysing AI-generated text: “perplexity” – how predictable is the text? – and “burstiness” – how strongly does sentence complexity vary? Humans spontaneously mix short and long sentences and shift rhythm more abruptly, whereas language models tend to favour more uniform structures.

4. Favourite vocabulary: Since the end of 2022, certain words have appeared with dramatically increasing frequency in online text. AI models have their preferred expressions too – words such as “crucial”, “innovation” and “transformative”, among many others.

Quick check: the ECHT model

If we bring together the observations in this article, we can build a compact framework for checking content. In German, we call it the “ECHT model” – “echt” meaning genuine or authentic:

E – Eigenheit / Distinctive feature: Is there a specific error or unexpected detail in the image or text that does not fit the overall picture?

C – Charakterliche Unstimmigkeit / Character inconsistency: What feels intuitively wrong? In images: the face, physics or shadows. In text: style, sentence structure or vocabulary.

H – Herkunftsnachweis / Provenance: Are there verifiable sources? For text, concrete source references are essential; for images, reverse image search and a look at the metadata can be worthwhile.

T – Ton und Zweck / Tone and purpose: Is the tone conspicuously optimistic and entirely free of conflict?

And what is the actual purpose of the content – to inform, to sell or merely to attract attention?

Conclusion: give your own voice enough room

AI-generated content is making the internet increasingly uniform and artificially cheerful. Genuine human content, by contrast, surprises, contradicts and carries the unmistakable signature of a thinking individual – including that person's weaknesses and reservations. AI-generated text, on the other hand, is often strikingly balanced, although this naturally also depends on the prompting.

AI slop is created when AI content is passed on without filtering – and that becomes problematic over time. Not because AI is inherently bad, but because in doing so we give too little room to our own voice – and our own judgement. Sharing AI-generated images and text is ultimately also a question of attitude: how much effort is my audience worth? And what is your view?