Unmasking machines: how VerifyLabs detects AI-generated content

Artificial intelligence can sometimes feel like magic, right? Whether’s its black magic or the good type depends on who’s using it. At VerifyLabs, we believe that understanding the basics of this technology builds understanding of the power it brings. So, let’s explore how AI detection, particularly with VerifyLabs.AI, helps us differentiate between what’s human-made and what’s a Bot Special.

AI: not a human mind but a sharp pattern spotter

It’s helpful to remember that AI doesn’t have a human brain. It doesn’t “think” or “understand” in the way we do. Instead, it’s incredibly good at pattern recognition. It’s like a super-smart detective that can find tiny clues in data that humans wouldn’t see.

Imagine teaching a child to recognise a cat. You show them thousands of pictures of cats – big ones, small ones (ooh kittens!), fluffy ones, short-haired ones, cats in different poses, different lighting. Eventually, the child learns to identify the common features that make a “cat” a “cat.” AI learning works similarly, but on a much grander, faster scale.

Training AI: differentiating between human and synthetic

For VerifyLabs to detect AI-generated content, our AI models undergo extensive “training.” This involves feeding them enormous amounts of data that are clearly labelled:

• “This is human-made content” (images, videos, audio, text from real people).
• “This is AI-generated content” (deepfakes, AI-written articles, synthetic voices, etc.).

During this training, the AI doesn’t just look at the surface. It dives deep, analysing countless tiny features and patterns.

What does AI “look for” in content?

The specific “clues” an AI detector looks for vary depending on the type of media (image, video, audio, text), but here are some common principles and examples:

In text:

• Predictability (“low perplexity”): human writing displays variety and inconsistency – variations in sentence length, structure, and word choice. We use long sentences, short sentences, and sometimes throw in unexpected words or phrases. AI, especially early versions, often tried to predict the most probable next word, leading to more predictable, uniform, and less surprising text. Modern AI is getting better at mimicking human weirdnesses, but subtle patterns can still exist.
• Vocabulary and phrasing: AI models are trained on massive datasets of human text. While they can generate coherent sentences, they usually exhibit preferred ways of phrasing or using certain words more consistently than humans might. They might lack the natural flow, idioms, or quirky expressions that give human writing “flavour”. That’s right: rest easy Dickens, Shelley, Byron.
• Factual accuracy and depth: AI can sometimes “hallucinate” or present plausible-sounding but incorrect information. To be fair, anyone working in marketing knows what that feels like, but human writing, especially well-researched pieces, typically shows a clearer, logical flow and deeper insight.
• Absence of “human errors”: perfection in life is usually a clue something’s off. No different in AI. Human oeuvres often contain slight imperfections, stylistic quirks, or even minor typos. AI-generated text is often “too perfect” in its grammar and structure, or it might have an overly formal or generic tone.

In images and videos:

• Subtle “artefacts” and imperfections: even highly realistic deepfakes can leave behind tiny digital fingerprints or “artefacts” that are invisible to the human eye but detectable by AI. These could be subtle inconsistencies in pixel patterns, compression artefacts, or noise that don’t match genuine media.
• Inconsistent lighting: AI still struggles to perfectly replicate complex lighting conditions, leading to unnatural shadows or highlights on the manipulated subject compared to the background.
• Unusual eye movements or blinking patterns: humans blink in fairly consistent ways. Some deepfakes might show too much or too little blinking, or unnatural eye movements.
• Facial anomalies: while sophisticated, some deepfakes might have subtle distortions around the edges of the face, unnatural skin textures, or a slight blurriness.
• Physiological irregularities (harder to differentiate now): earlier deepfakes sometimes had issues with things like a person’s pulse showing in the skin, or consistent blood flow in the face. While more advanced models are better at this, these subtle physiological cues can still be targets for detection.
• Lip synchronisation and mouth shapes: if audio is added to a video, the AI checks if the lip movements perfectly match the spoken words and if the mouth shapes look natural for the sounds being made.

In audio:

• Spectral analysis: AI can analyse the unique sound frequencies and patterns within a voice. Human voices have natural variations and imperfections that AI-generated voices might lack.
• Prosody and intonation: this refers to the rhythm, stress, and intonation of speech. While AI can mimic these, subtle unnaturalness in pauses, emphasis, or emotional expression can be detected.
• Background noise consistency: AI-generated audio might sound unnaturally “clean”, or the background noise might be inconsistent with the simulated environment.

Classifiers and confidence scores

At its heart, VerifyLabs.AI uses what are called “classifiers” – these are the AI models that have been trained to distinguish between humans and AI. When you upload content to VerifyLabsAI, our system performs the following steps:

1. Analysis: the content is broken down and meticulously analysed for all the subtle patterns and clues mentioned above.
2. Comparison: these patterns are then compared against the vast knowledge base the AI gained during its training.
3. Confidence score: the AI then assigns a “confidence score” – essentially, how sure it is that the content belongs to the “human” category or the “AI-generated” category.
4. Clear result: VerifyLabs.AI translates findings into our easy-to-understand “green circle” (human), “red square” (AI-generated) or “grey bar” (more investigation advised) indicator. We believe this clear, simple visual helps you make informed decisions without needing to be an AI expert yourself.

The arms race: why constant innovation is crucial

AI technology is advancing at lightning speed. This means that just as AI models get better at creating deepfakes and synthetic content, our AI detection models must also evolve to keep pace. VerifyLabs.AI is part of a highly-skilled team dedicated to continuous research and development.

Our goal is to be your trusted partner in the age of AI. By understanding a little more about how AI detection works, you can appreciate the sophisticated technology working behind the scenes to show you what’s real and protect your digital peace of mind.

Trust and Verify.