Generative AI

• 1. Introduction

  ›

  Generative artificial intelligence arrived in public view not with a single announcement but with a moment of recognition. In November 2022, ChatGPT became available to anyone with an internet connection, and within weeks millions of people were typing questions into it and watching it write back in full sentences. What they were encountering had a history stretching back more than a century, a technical architecture built over years of quiet research, and a set of consequences that governments, courts, artists, and scientists are still trying to understand.

  The field is called generative AI because its systems do not merely classify or predict. They produce: text, images, audio, video, software code, three-dimensional objects, even candidate mathematical proofs. To make that possible, the models learn the underlying patterns of their training data and then use those patterns to generate new material in response to a prompt. Understanding where that capacity came from, what it can do, and what it costs the world to run at scale is what this documentary explores.

  154243

• 2. A Century of Precursors

  ›

  Russian mathematician Andrey Markov introduced the mathematical tool that would become generative AI's earliest ancestor back in 1906. His analysis of vowel and consonant sequences in the poem Eugeny Onegin gave rise to what became known as the Markov chain. A Markov chain, once trained on a body of text, can generate new probabilistic sequences; it was the first formal method for modeling the patterns of natural language by machine.

  By the early 1970s, artists were pushing these ideas into visual territory. Harold Cohen developed a computer program called AARON, which was designed to generate paintings autonomously. Cohen exhibited works produced by AARON, introducing audiences to the idea that a machine could make something that looked like art.

  In the 1980s and 1990s, the phrase generative AI planning described something quite different from today's chatbots: symbolic AI systems used to generate sequences of actions toward a goal. Practitioners applied these tools to military crisis action plans, manufacturing process plans, and early autonomous spacecraft. Researchers described the technology as relatively mature by the early 1990s, long before the neural network era.

  19202123

• 3. The Architecture That Changed Everything

  ›

  In 2014, two new model types arrived almost simultaneously. The variational autoencoder, or VAE, and the generative adversarial network, or GAN, made it possible for the first time to build deep generative models of complex data like images. A GAN works by pitting two neural networks against each other: a generator that tries to produce convincing synthetic data and a discriminator that tries to spot the fakes. The continuous competition between them pushes the generator toward increasingly realistic outputs.

  The year 2016 brought another advance specifically in audio. DeepMind's WaveNet demonstrated that a deep neural network could generate raw audio waveforms directly, enabling speech synthesis more realistic than anything available before it. A subsequent system called Tacotron 2 extended this to full end-to-end neural text-to-speech generation.

  Then in 2017 came the transformer architecture, which proved more capable than the earlier long short-term memory networks for generative tasks. Transformers process tokens in parallel rather than sequentially, improving both training efficiency and scalability. That architecture became the foundation for OpenAI's generative pre-trained transformer series, beginning with GPT-1 in 2018. Every major language model since has built on the same core idea.

  3031323356

• 4. Public Adoption and Global Divides

  ›

  In March 2020, a free web application called 15.ai, created by an anonymous MIT researcher, offered users the ability to generate convincing character voices from minimal training data. It is credited as the first mainstream service for audio deepfakes. The following year, DALL-E drew widespread public attention to the possibility of turning a text description into a generated image.

  The full public reckoning came in November 2022 with the release of ChatGPT. Within months, generative AI was inside Microsoft Office as Copilot, inside Adobe's suite as Firefly, and inside Google Photos. Open-source alternatives including Stable Diffusion and the LLaMA language model made the technology available outside corporate platforms entirely.

  Adoption was not uniform across the world. A survey by marketing research firm Ipsos found that Asia-Pacific populations were significantly more optimistic about generative AI than Western societies. A separate survey by SAS and Coleman Parkes Research found that 83% of Chinese respondents were using the technology as of 2023, against a global average of 54% and a U.S. rate of 65%. A UN report documented that Chinese entities filed over 38,000 generative AI patents between 2014 and 2023, more than any other country. By mid-2025, however, many companies were walking back pilot projects, citing integration difficulties, data quality problems, and unmet returns. Analysts at Gartner described this moment as the hype cycle's trough of disillusionment.

  3534374648

• 5. Deepfakes, Disinformation, and Cybercrime

  ›

  In July 2023, the fact-checking company Logically tested Midjourney, DALL-E 2, and Stable Diffusion by prompting them to produce disinformation images. The models obliged, generating images depicting electoral fraud in the United States and Muslim women supporting India's Bharatiya Janata Party. The test illustrated a structural problem: systems built to respond to natural language requests could generate politically destabilizing imagery without any specialized knowledge on the part of the requester.

  Deepfakes, a portmanteau of deep learning and fake, extend this problem to video and audio. They have been used in celebrity pornography, revenge porn, financial fraud, and covert foreign election interference. In April 2023, the German publication Die Aktuelle published an AI-generated fake interview attributed to Michael Schumacher. Voice-cloning technology has been used to create fake versions of musicians' songs and to mimic the voices of public officials.

  In 2025, a report by Graphika documented how generative AI was being used to launder articles from Chinese state media through social media, obscuring their origin. That same year, a report from the American Sunlight Project found that the Pravda network was publishing as many as 10,000 articles a day, with the apparent goal of pushing particular narratives into the training data of large language models. Also in 2025, Israel signed a six-million-dollar contract with a U.S. firm called Clock Tower X to spread pro-Israel content across platforms in order to influence what ChatGPT, Gemini, and Grok would say.

  16369192

• 6. Copyright, Labor, and the Law

  ›

  ChatGPT and Midjourney are trained on large publicly available datasets that include copyrighted works. AI developers have argued that this constitutes fair use; copyright holders have argued the opposite. Getty Images sued Stability AI over the training of Stable Diffusion on its images. Both the Authors Guild and The New York Times sued Microsoft and OpenAI over the use of their works to train ChatGPT. As of 2024, several of these lawsuits remain ongoing.

  In April 2023, it was reported that image generation AI had caused 70% of video game illustration jobs in China to disappear. The same month, developments in generative AI contributed to the Hollywood labor disputes of 2023. Fran Drescher, president of the Screen Actors Guild, stated during the SAG-AFTRA strike that artificial intelligence poses an existential threat to creative professions. A 2025 study examining the U.S. labor market found no discernible disruption so far. A study of Danish workers found that those who used chatbots saved an average of 2.8% of their working time, with no significant change in earnings or hours worked.

  In January 2025, the United States Copyright Office released detailed guidance concluding that AI-generated works can qualify for copyright protection when the human user exerts sufficient creative control over the output. The same office subsequently registered the first visual artwork composed entirely of AI-generated materials, titled A Single Piece of American Cheese. In the European Union, the Artificial Intelligence Act requires disclosure of copyrighted training material and labeling of AI-generated output.

  17131299100

• 7. The Environmental Price

  ›

  Training a large language model and then running it at the scale of a popular product requires enormous data centers, and those data centers carry real-world costs. Scientists and journalists have raised concerns about CO2 emissions, electronic waste, diesel generator exhaust, and the freshwater consumed for cooling. The carbon footprint of generative AI globally is estimated to be growing steadily. Potential annual emissions range from 18.21 to 245.94 million tons of CO2 by 2035, with the higher end of that estimate approaching the annual emissions of the U.S. beef industry, currently pegged at 257.5 million tons.

  Training frontier models requires computing power on a scale that only large technology companies can afford. Smaller firms such as Cohere and OpenAI have ended up buying access to data centers from Google and Microsoft respectively. Export controls add a geopolitical dimension: in 2022, the United States imposed restrictions on the sale of GPU and AI accelerator chips to China. New chip designs including NVIDIA's A800 and Biren Technology's BR104 were developed specifically to work within the limits of those sanctions.

  Smaller models can run locally. LLaMA-7B, a version with 7 billion parameters, can run on a Raspberry Pi 4. One version of Stable Diffusion can run on an iPhone 11. For those larger models with hundreds of billions of parameters, such as GPT-4, datacenter-scale hardware is required, typically arrays of GPUs such as NVIDIA's H100 or AI accelerators such as Google's TPU chips.

  1887888983

• 8. What Comes Next

  ›

  Google DeepMind introduced three systems in 2023 that suggest the direction the field is heading. FunSearch generated computer programs that solved problems in pure mathematics, including new constructions in the cap set problem and the bin packing problem. AlphaDev discovered small sorting algorithms that outperformed known human benchmarks and were subsequently integrated into the LLVM standard C++ sorting library. GNoME proposed candidate inorganic crystal structures for testing in materials science.

  In 2025, Google DeepMind followed with AlphaEvolve, a system for general-purpose algorithm discovery that uses large language models to propose code changes, automated evaluators to assess each candidate, and an evolutionary process to iteratively improve results. These applications suggest that generative AI is moving beyond producing prose and images into producing scientific and mathematical knowledge that researchers can verify and use.

  Detection remains a parallel challenge. In May 2025, Google deployed SynthID, its watermarking tool for output from Gemini, Imagen, and Veo. Detection requires using Google's own SynthID detector portal. OpenAI developed a comparable watermarking tool for ChatGPT in 2023 but did not release it, citing concerns that users would switch to competitors and that the tool could be circumvented simply by asking another model to rephrase the output. That gap between generating content and reliably identifying it remains one of the field's most consequential open problems.

  6367666870