How to identify AI-generated content: a step-by-step guide
After the release of ChatGPT in late 2022, artificial intelligence (AI) powered tools, especially those used to generate different types of content, have spread rapidly.
Image by Cybernews.
After the release of ChatGPT in late 2022, artificial intelligence (AI) powered tools, especially those used to generate different types of content, have spread rapidly.
Individuals and businesses are leveraging generative AI (Gen AI) tools to produce various kinds of content, such as text, images, audio and videos, to fulfill different needs. However, as with any new technology, this comes with risks.
For instance, threat actors use Gen AI to create phishing emails that are difficult for humans and machines to detect.
Political opponents in different countries (e.g., examples of what happened in recent US elections) are using deepfake content to impact public opinion. These are just some examples of the negative side of generative AI.
In this article, I will discuss the different ways we can detect whether a piece of content was produced using AI or not. However, before we begin, let's explore the various AI techniques used to create content.
Understanding AI-Generated Content
AI tools can generate content that spans the following four categories:
- Text
- Image
- Video
- Audio
While there are dedicated tools to produce each type of content, as technology advances, many AI tools have become multi-models; hence, they can generate more than one content type. For example, ChatGPT can produce both images and text.
When talking about generating text content, AI tools can produce the following types of text:
- Articles
- Blog posts
- Research papers
- Marketing copy – such as social media posts and copywriting brochures
- Source code – AI can assist software programmers to create code
- Creative writing – including stories, poems, and scripts
- Academic essays
- Technical documentation – for example, those used to describe software products functionality
- Translations between languages
Here are links to the most popular AI text generator tools:
These represent the main large-scale, general-purpose AI text generators.
There are plenty of other AI tools that provide specific text-generation services.
However, many of these are based on the already mentioned AI systems' large language models (LLMs) or use similar technologies.
For image generation, AI tools can create the following image types:
- Photorealistic images – You can describe the image you want to generate in text and the AI program will create it for you.
- Digital art and illustrations – such as cartoon images.
- Product designs.
- Architectural visualization – AI tools can quickly generate visualization without the need to have extensive knowledge of 3D modeling and rendering software.
- Logo designs.
Here are links to AI tools for generating images:
- DALL·E 3 – the most famous AI image generator tool, the latest version can be used via ChatGPT or Microsoft Bing's AI Copilot.
- Midjourney – you need to access it via the Discord platform (see Figure 1).
- Stable Diffusion – an open-source alternative to DALLE and Midjourney. It is commonly used to generate artistic portraits, historical portraits and architectural renders with ease.
- Adobe Firefly – Adobe integrates Gen AI of images into its widespread Photoshop program. Users can generate images by providing detailed text descriptions. You can test the functionality of Adobe Firefly online for free by visiting Adobe Express online service.
In the context of video generation, AI is advancing rapidly, allowing users to generate:
- Text-to-video generates video and other visual effects from text description.
- Generate short videos specialized to be published on specific social media platforms, such as Facebook and Instagram.
- Face-swapping in existing videos.
- Video editing and enhancement.
Here are some AI video generation tools:
- Faceswapper - allows users to swap faces and clothes.
- Deepswap – an online tool for swapping faces in videos.
- Vidnoz AI Face Swap tool - Swap faces in both images and videos.
- Descript - allows editing videos by editing their transcript.
- Fliki – creates videos for social media platforms.
- Invideo – text-to-video converter.
- VIDEO GPT – creates videos from text prompts (powered by ChatGPT).
- Gen-2 - text-to-video and image-to-video generation.
For audio generation, AI tools can produce different audio content, such as:
- Text-to-speech
- Voice-overs for videos and podcasts
- Generate audio responses for customers' inquiries
- Music composition
- Music backgrounds for videos
- Sound effects – such as generating natural sounds like rainfalls or thunder
- Generate sounds to be used in applications and websites
- Audio enhancement, such as noise reduction in recordings and separation of audio elements in music tracks
- Voice cloning – this is used to generate a voice that resembles any person. This technique is used to generate deepfake audio content. An example of a cyberattack carried out using this technique is what happened in Hong Kong at the beginning of this year. A finance worker at a multinational firm was deceived into paying out $25 million to fraudsters who used deepfake technology to impersonate the company's chief financial officer in a video conference call.
How do we discover AI-generated content?
Now, let us discuss how we can discover content generated using AI technology.
Text analysis
Text content generated using AI tools can be detected by checking the following characteristics:
- AI-generated text tends to use a consistent tone throughout the text, even in long pieces. Human writing tends to utilize natural variations in style and tone.
- Lack of human experience. For instance, text generated using AI tools lacks personal experience and specific real-world examples. AI models cannot give authentic, detailed examples when generating text about topics like "the dangers of deepfake technology in businesses".
- AI tools often produce general concepts and find it challenging to dive deep into specific subjects, typically giving only broad ideas about a topic.
- AI tools may use the same word or phrase in a small text multiple times. For example, some AI models – such as ChatGPT- are known to overuse certain words like "realm" or "indeed" , "Moreover" repeatedly in addition to phrases such as: "Delve into"," It is important to know", "Delving into the intricacies of" and "Furthermore".
- AI systems are often trained on outdated data. This makes them unable to incorporate very recent events or the latest advancements in their generated text, making it look absolute.
- AI sometimes uses words that human writers do not commonly use. These words could be technically correct but are rarely used by humans to describe something in the same context.
There are different online AI detectors:
Image analysis
Different signs and techniques can be used to spot images generated using AI tools.
Inconsistent details
AI tools still find problems generating 100% accurate details in images. For instance, always look at the details in a photo, especially when the suspicious image contains many elements. For example, look a person’s fingers when checking a picture depicting people. It is common for AI tools to add more fingers to human hands or draw them strangely (see Figure 3).
To check for small details appearing in suspicious photos, you can use enlarging tools:
Problems in displaying texture
AI tools find it difficult to generate texture, especially those using complex patterns. For example, always check textures in images. If they look extra smooth or unnaturally sharp, this could be considered a sign of an AI-generated image. Pay close attention to hair and skin textures.
Wrong element positions
AI-generated tools may place some elements in an image in the wrong place. For example, if the image shows a dining table, misplaced items on the table could indicate an AI-generated image.
Image metadata
Check image metadata. For instance, AI tools may not create metadata for their generated images or could insert random metadata into these images. You can use the following tools to check for image metadata:
Inconsistent backgrounds
AI-generated images may use different backgrounds assembled from various images. Such inconsistency could indicate that he image is AI-generated. Look for mismatched lighting, shadows or sudden changes in style or detail level.
Check text content
If the image contains text, make sure to spell-check it thoroughly to see if it contains grammatical or spelling errors. If so, there is a high chance that this image is AI generated.
Check for watermarks
Many AI tools watermark their generated images. Some watermarks could be drawn using a transparent font, so check the image thoroughly for the existence of watermarks.
Lighting and shadow inconsistencies
AI generated content may feature inconsistent shadows when generating images of objects. For example, these shadows may appear in the wrong place.
Finally, we can use AI detection tools to detect AI-generated images; here are a few tools:
- Is it AI (AI-Generated Image Detector)
- AI or Not - detects AI-generated images & audio
- Content at Scale (Advanced AI Image Detector)
- Fakeimagedetector – detects manipulated images using advanced techniques like metadata analysis and ELA analysis
- Winston AI - detects AI-generated images and deepfakes made with Midjourney, DALL-E, Stable Diffusion, Meta AI and more
Video analysis
A video is a series of moving images in a specific order. So the same indicators applied to detecting images can be used to indicate if a video is created using AI technology.
However, there are still some indicators specific to videos, such as:
- Strange facial movements: Always look at the person’s facial movements. If you find inconsistent movements, such as abnormal lip or eye movements, this could point to an AI-generated video.
- Check audio and lip movements: Many AI-generated videos have inconsistencies between lip movements and speech. For example, many deepfake videos of politicians were discovered due to poor lip-synchronization.
- Unordinary scene transition: Many AI videos suddenly move from one scene to another. Here, there is no natural flow from one scene to another.
- Check for consistent quality: AI-generated videos commonly have the same quality throughout the entire video, and this consistency does not appear in regular videos.
Here are some tools to discover AI-generated videos:
Audio analysis
In the audio field, detecting AI-generated content can be achieved by using the following tools:
- Resemblyzer - it allows you to derive a high-level representation of voice through a deep learning model (referred to as the voice encoder). Given an audio file of speech, it creates a summary vector of 256 values (an embedding, often shortened to "embed" in this repo) that encapsulates the characteristics of the voice.
- AI Speech Classifier – allows you to detect whether an audio clip was created using ElevenLabs.
- Detect AI voices from play.ht – an online tool to detect if it is synthetic or real.
- Aivoicedetector - identify whether an audio was generated by an AI voice or by a human voice. You can access this tool either through a browser extension or by uploading the audio file online.