Generative AI: What is it, and how does it work?

What is generative AI?

Although the topic of generative AI has bloomed recently, the technology itself is not new. The idea of computers mimicking human thinking abilities has entertained science-fiction writers since the dawn of computerization. Artificial intelligence began to be treated more like reality than fantasy when English mathematician and computer scientist Alan Turing published his paper "Computing Machinery and Intelligence" in 1950, introducing the concept of machines capable of reasoning similarly to humans.

Since then, artificial intelligence and its generative branch have rapidly progressed. However, let’s focus on the current decade. While AI and machine learning are nothing new, in 2022 discussion about them ramped up. That’s when names like ChatGPT, DALL-E, and Midjourney began to appear. These are among the first generative AI models available to the general public due to their easy-to-use interfaces and ability to understand queries written in natural language used by humans in conversations, unlike earlier models operated in programming languages.

Generative AI vs. traditional AI

Traditional AI and generative AI are two branches of the same field of science: artificial intelligence. The difference between AI and generative AI is that “basic” AI has somewhat different capabilities and uses. Traditional AI aids with automating repetitive and menial tasks, data analysis, and predictions based on historical information and is generally used to perform predefined tasks. Generative AI is designed to produce new content based on existing content. While it can detect patterns, it can also recreate them differently.

How does generative AI work?

Generative AI uses machine learning, or more specifically, a subclass of machine learning known as deep learning. Deep learning uses artificial neural networks, which are artificially created virtual neural models inspired by the functioning of neural networks found in the brains of animals.

What distinguishes deep learning from other machine learning methods is its ability to learn semi-supervised and unsupervised, meaning that deep learning models can ingest very large sets of unlabeled data with minimal human surveillance. Unsupervised learning means that the deep learning model takes the data without instructions on what to do with it and automatically tries to take it apart and analyze it to find patterns. The foundation models created using this method can then be used as the basis for various types of generative AI.

Through deep learning, generative AI technology can identify patterns that appear in human-created content, learn them, and replicate them. It generates new content while guided by user requests called "prompts.”

What are generative AI models?

When discussing generative artificial intelligence, it’s worth knowing that no single generative AI model exists. Researchers have come up with many ways to train models, and we’ve listed some of them below. Each model can be used for different purposes and generate results of different quality and characteristics. Generative AI models can also be combined to produce different or better results.

Diffusion models

Diffusion models are used in image generation and use forward and backward training or forward and backward diffusion. A diffusion model is trained by adding a random noise pattern to the training data, which is called forward diffusion. It then learns to recover said data through a backward diffusion process. After such training, the diffusion model learns to generate new data based only on randomized noise patterns.

Generative adversarial networks (GANs)

GAN models use two competing neural networks, called a generator and a discriminator. The generator generates outputs based on prompts, while the discriminator compares them with real examples of training data to determine which outputs have been artificially generated. The goal is to repeat this cycle for as long as necessary so that the discriminator stops noticing differences between the real and artificial data.

Variational autoencoders (VAEs)

Variational autoencoders are best suited for generating video, images, and text. They are similar to GANs: They also use two neural networks, but in this case, there’s an encoder and a decoder. The encoder converts the data to latent space (a representation of the compressed data that still contains key features of the input data), while the decoder converts it into new content.

Transformer-based models

Transformer models are designed to learn the context and meaning of data by analyzing the relationships between successive pieces of information. Transformer models are the basis of large language models (LLMs), which can process natural language. For example, the popular ChatGPT chatbot is based on a transform model, and the "GPT" part refers to the "generative pretrained transformer.” In addition to sentences, transform models can also analyze DNA chains or amino acid sequences in proteins.

Although the textual content created by LLMs may seem human-like and intelligent, it’s worth remembering that they essentially work by predicting one word after another, resulting in the formation of comprehensible sentences. Thus, artificial intelligence, as we know it, isn’t really intelligent and cannot be fully trusted to provide factual information.

Examples of generative AI tools

The current AI craze means that every month, or maybe even every week, a new tool or feature based on generative artificial intelligence emerges. However, a few names have really made the news:

ChatGPT

ChatGPT is a large language model (LLM)-based chatbot developed by OpenAI and probably the best-known (so far) generative AI tool. It was launched in November 2022. As a chatbot, it was designed to mimic the way another human would respond to a user’s queries. However, it’s versatile in its capabilities and can generate many types of textual content, such as song lyrics, translations, poems, computer code, or simulated conversations.

ChatGPT has "combined" with the DALL-E image generation model, resulting in a multipurpose chatbot that is now capable of generating both text responses and visual content. The image generation function is available in the paid version of ChatGPT Plus, which currently uses the newer GPT-4 language model. At this time, the free version of ChatGPT is based on GPT-3.5.

The popularity of ChatGPT has sparked industry-wide competition, with tech giants like Google and Microsoft racing to develop their own LLM-based chatbots and AI assistants. However, the question remains: Is ChatGPT safe ? As far as online apps are concerned, they’re safe as long as you don’t share any sensitive data with them or they don’t leak your data after a data breach, which is true for all cloud-based services.

Google Gemini

Gemini, formerly known as Bard, is Google’s answer to OpenAI’s ChatGPT. Like its competitor, it’s an LLM-based chatbot and was launched in March 2023 on a limited scale. Gemini uses a large language model of the same name, which is being developed by Google. The Google team claims that the chatbot has been trained to understand images, video, and audio inputs, and it’s possible that it will be further developed to generate these types of content as outputs as well.

Microsoft Copilot

Microsoft Copilot was introduced as Bing Chat in February 2023. The chatbot was based on OpenAI’s GPT-4 language model and was presented to the public as a built-in feature of the Edge browser and Bing mobile app. Later, Microsoft decided to unify its AI products under the name Copilot.

The Copilot chatbot operates a refined version of the GPT-4 language model called Prometheus. In March 2023, Microsoft merged it first with the DALL-E 2 image generator and then DALL-E 3, resulting in a more functional product capable of generating both textual and visual content.

DALL-E

DALL-E is another generative AI product developed by OpenAI that focuses exclusively on image generation. DALL-E 3 was launched with the intention of incorporating it into ChatGPT. It has also been implemented by Microsoft Copilot.

The latest versions of DALL-E are available to the public only through the aforementioned chatbots. However, the Craiyon platform (formerly DALL-E Mini), developed by machine learning company Hugging Face, uses a generative model based on the original DALL-E.

Midjourney

Midjourney is an image generation tool released in July 2022 and developed by the San Francisco-based company of the same name. Its goal is to generate images based on user descriptions entered in natural language. Midjourney is currently available via a bot on the tool’s official Discord server.

The Midjourney tool and ChatGPT were a key part of the "AI boom" that began around 2022. The tool was responsible for an influx of AI-generated memes, such as the famous Balenciaga Pope, which was a photorealistic image of Pope Francis wearing a fluffy white Balenciaga jacket. Due to the quality of the results, Midjourney has also sparked the first discussions over the future of generative AI, the security of Midjourney and similar tools, AI misinformation, and the potential harm it could lead to.

QuillBot

QuillBot is a generative AI-based software that was advertised as a writing assistant and launched in 2017. The tool’s purpose is to generate text based on prompts and paraphrase existing sentences. The tool also offers grammar and plagiarism checking, translation, and a summarization tool. QuillBot security is ensured by encryption and regular safety checks, but as with all generative tools – it’s better not to give it any personal information just to be safe.

Use cases of generative AI

AI can be leveraged in several ways, such as:

• Content generation. Generative AI tools are primarily designed to generate output based on user requests. Different AI tools are designed to do other things, so depending on your tool, you can generate text, music, or an image that matches your description.
• Idea generation. Generative AI chatbots can be used as "pocket companions" for brainstorming and generating ideas. For example, they can come in handy if you are struggling with a creative block.
• Customer service. Various companies are using AI chatbots as support agents to communicate with their customers since chatbots are designed to mimic human responses.
• Enhancing learning. Chatbots can be prompted to ask you questions about specific topics or converse in a foreign language, acting as virtual learning assistants.
• Scientific development. Generative AI is not just about text and images. AI models are also helpful for research in biology, chemistry, neuroscience, drug research, and astronomy.

Benefits of generative AI

AI has several advantages, including:

• It boosts efficiency. Generative AI can assist humans in various tasks related to content creation, such as creating first drafts of text or visual content, reducing time costs, and increasing productivity.

• It enhances creativity. Generative artificial intelligence can sometimes "come up with" an idea or solution you wouldn’t have thought of, aiding your creative and problem-solving process. It can generate ideas or make suggestions on how to improve your work.

• It offers multiple personalization solutions. Generative AI can tailor its results to individual users based on their unique preferences and behaviors. It can be used in customer service or advertising, creating individually tailored content.

Drawbacks of generative AI

While generative AI has some benefits, it’s impossible to ignore its drawbacks. Since the technology is new and unregulated, people, institutions, and governments have many questions about its ethics, safety, potentially harmful uses, and the future of the industries it has already begun to affect.

• Hallucinations. Generative AI might be inspired by humans, but it isn’t one. Large language models don’t understand what or why they’re communicating, which can result in them spitting out nonsense disguised as facts, known as “hallucinations.” Users who aren’t vigilant could potentially be fooled if it provides misinformation.

• Bias. Depending on the data and methods used for training, generative AI models can be biased in various ways, as evidenced by the Google Gemini fiasco of February 2024, when users discovered that the tool produced inaccurate and misleading images of historical figures and events. Generative AI tools can perpetuate stereotypes present in training data and may be biased even more than humans, who, unlike AI, have the ability to assess the validity of information based on their knowledge and real-life experience.

• Technological unemployment. It’s already being reported that many creative jobs worldwide are in less demand due to the integration of generative AI into workplaces. Graphic designers, content writers, illustrators, concept artists, voice actors, and other creative professionals are at risk of being replaced by LLMs.

• Copyright violations. Generative AI models require input data to be trained on, and this data often comes from unwilling participants. Currently, the legal situation of AI-generated content is unclear, and AI companies are being sued by people whose work has been used to train models without their consent.

• Misinformation and propaganda. AI images and deepfakes can be used to manipulate public opinion, showing politicians saying things they’ve never said, opposition figures committing crimes, or just ordinary people doing things they’ve never done, potentially causing informational and legal chaos in the future, especially as we witness advancements in how deepfakes are made.

• AI spam. The generative AI boom has caused many companies, especially marketing companies, to embrace the technology to an extreme. Chatbots have been deployed to generate entire sites focused solely on driving engagement by incorporating highly positioned SEO keywords into meaningless content.

Should I use generative AI?

If you want to use generative AI, you should remember the importance of your security. Generative AI tools do not operate in a vacuum – they can collect data and anything you input into them. You shouldn’t talk to a chatbot about anything you wouldn’t want to leak. For example, don’t share sensitive personal or company data. AI chatbots are not sentient and cannot hurt you, but the companies that operate them can use your data to further develop their tools. There have been instances, for example, of ChatGPT outputting sensitive data that someone else had previously entered.

Also, remember to verify everything the chatbot tells you. They seem smart in how they "talk," but they are imperfect. Chatbots tend to hallucinate when they don’t know how to answer your questions or spontaneously become more creative than factual, resulting, for example, in generating non-existent book titles and scientific papers.