Generative artificial intelligence

• 1. The First Algorithmic Voice

  ›

  In 1906, Russian mathematician Andrey Markov introduced a concept that would eventually birth generative artificial intelligence. He analyzed vowel and consonant patterns in the novel Eugeny Onegin to model natural language. This early work laid the foundation for systems that could generate probabilistic text after training on a corpus of words. By the early 1970s, artists began using computers to extend these techniques beyond simple Markov models. Harold Cohen developed AARON, a pioneering computer program designed to autonomously create paintings. These works were exhibited and marked an early attempt at machine creativity.

  The terms generative AI planning or generative planning emerged in the 1980s and 1990s to describe systems generating sequences of actions. Military planners used these tools to generate crisis action plans while manufacturing sectors utilized them for process planning. By the early 1990s, this technology was considered relatively mature within its specific domains. The field then shifted focus toward deep learning technologies beginning in the late 2000s. Neural networks trained as discriminative models dominated the era due to the difficulty of generative modeling.

  192021

• 2. Architectures That Generate Reality

  ›

  Advancements in 2014 produced the first practical deep neural networks capable of learning generative models for complex data like images. Variational autoencoders and generative adversarial networks enabled systems to output entire images rather than just class labels. DeepDream became one of the earliest examples of such deep generative models producing visual content. In 2017, the Transformer network replaced older long short-term memory models to enable advancements in generative capabilities. This architecture allowed models to process entire sequences simultaneously and capture long-range dependencies more efficiently.

  OpenAI developed the first generative pre-trained transformer known as GPT-1 in 2018 following the Transformer breakthrough. Generative adversarial networks consist of two neural networks trained simultaneously in a competitive setting. A generator creates synthetic data by transforming random noise into samples resembling training datasets. A discriminator learns to distinguish authentic data from the synthetic data produced by the generator. These two models engage in a minimax game where the generator aims to fool the discriminator while the discriminator improves its detection abilities. Transformers allow models to capture the significance of every word in a sequence when predicting subsequent words.

  31325658

• 3. Adoption Across Global Markets

  ›

  In March 2020, the release of 15.ai marked one of the earliest popular use cases of generative AI. An anonymous MIT researcher created this free web application to generate convincing character voices using minimal training data. The platform became the first mainstream service for AI voice cloning used in memes and content creation. By 2021, DALL-E emerged as a transformer-based model advancing AI-generated imagery. Open-source projects like VQGAN+CLIP and DALL-E Mini became widely used public text-to-image generation tools before mobile applications allowed users to generate art from simple prompts.

  Midjourney and Stable Diffusion launched in 2022 to democratize access to artificial intelligence art creation. These systems generated photorealistic images based on text descriptions leading to widespread adoption among artists and designers. November 2022 saw the public release of ChatGPT which popularized generative AI for general-purpose text tasks. A 2024 survey revealed that Asia-Pacific countries were significantly more optimistic about generative AI than Western societies. China emerged as a global leader with 83% of respondents using the technology compared to a global average of 54%. Chinese entities filed over 38,000 generative AI patents between 2014 and 2023 surpassing United States patent applications.

  46477578

• 4. The Copyright War Begins

  ›

  Generative AI systems such as ChatGPT and Midjourney are trained on large publicly available datasets containing copyrighted works. AI developers argue this training is protected under fair use while copyright holders claim it infringes their rights. Getty Images sued Stability AI over the use of its images to train Stable Diffusion. The Authors Guild and The New York Times sued Microsoft and OpenAI regarding the use of their works to train ChatGPT.

  As of 2024 several lawsuits related to the use of copyrighted material in training remained ongoing. The United States Copyright Office ruled that works created by artificial intelligence without human input cannot be copyrighted due to lack of human authorship. In January 2025 the office released extensive guidance establishing that users can exert control over selection and placement of creative elements. This guidance allowed for case-by-case determination of whether modifications rise to the minimum standard of originality required under Feist. Subsequently the USCO registered the first visual artwork composed entirely of AI-generated materials titled A Single Piece of American Cheese.

  1710093

• 5. Societal Fears And Labor Shifts

  ›

  In April 2023 reports indicated image generation AI resulted in 70% of jobs for video game illustrators in China being lost. Fran Drescher declared during the 2023 SAG-AFTRA strike that artificial intelligence poses an existential threat to creative professions. Voice generation AI emerged as a potential challenge to the voice acting sector while developments contributed to Hollywood labor disputes. However a 2025 study concluded the US labor market had not experienced discernible disruption from generative AI so far.

  Generative models reflect and amplify cultural biases present in underlying data. A language model might assume doctors are male and secretaries are female if those biases exist in training data. An image model prompted with text describing a CEO might disproportionately generate images of white male CEOs when trained on racially biased datasets. Deepfakes have garnered widespread attention for uses including revenge porn fake news hoaxes health disinformation financial fraud and covert foreign election interference. In July 2023 fact-checking company Logically found popular models would produce plausible disinformation images when prompted to do so.

  13144449

• 6. The Carbon Footprint Of Intelligence

  ›

  Scientists and journalists expressed concerns about the environmental impact of developing and deploying generative models. High CO2 emissions large amounts of freshwater used for data centers and high electricity usage characterize the industry. The carbon footprint globally is estimated to be growing steadily with potential annual emissions ranging from 18.21 to 245.94 million tons of CO2 by 2035. Highest estimates for 2035 near the impact of the United States beef industry which currently emits 257.5 million tons annually as of 2024.

  Proposed mitigation strategies include increasing efficiency of data centers to reduce energy usage and building more efficient machine learning models. Regulating transparency of these models and their energy and water usage remains a key focus area. Researchers are encouraged to publish data on their models' carbon footprint while subject matter experts understanding both machine learning and climate science become more numerous. Large-scale data centers consume fresh water for cooling and generate significant e-waste alongside high energy consumption that grows steadily.

  18