Generative AI: How It Works, What It Does Well, and Where It Fails
When OpenAI released ChatGPT on 30 November 2022, it reportedly reached an estimated 100 million users within about two months, faster than any consumer app before it. The technology behind it was not new. Its core design, the transformer, was described by Google researchers in a 2017 paper titled "Attention Is All You Need". What changed was scale, and a chat interface that let anyone try it. Generative AI, systems that produce new text, images, audio, code or video, has since moved into offices, classrooms and courtrooms, with real benefits and some well-documented failures.

Illustration: generative AI as a creative collaborator.
How it works, briefly
Large language models such as GPT, Claude, Gemini and Llama are trained on very large collections of text to predict the next word fragment (token) in a sequence. Doing that well across billions of examples forces the model to capture grammar, facts and patterns of reasoning. They are then fine-tuned on examples of helpful answers and on human ratings of responses, a step that turns a raw text predictor into an assistant.
Image generators such as Midjourney, Stable Diffusion and DALL·E mostly use diffusion models. These are trained to remove noise from images step by step. To create a picture, they start from pure noise and gradually "denoise" it into an image that matches a text description.
Neither type looks up answers in a database. They generate output from learned patterns. That explains both what they are good at and how they fail.
What the evidence says about productivity
Two widely cited studies give concrete numbers:
- At a large software company's customer-support centre, agents given an AI assistant resolved about 14% more issues per hour. The least experienced workers improved the most, by about a third (Brynjolfsson, Li and Raymond).
- In an experiment with 453 professionals writing workplace documents, access to ChatGPT cut the time taken by about 40% and raised quality ratings by 18% (Science, 2023).
The gains tend to be largest for routine writing, summarising, drafting code and getting less experienced people up to speed. They are smaller, and sometimes negative, for tasks that need specialised judgement or fact-checking, where people may trust confident but wrong output.
Known failure modes
Making things up. Language models can produce fluent, specific and entirely false statements, often called "hallucinations". In a well-known 2023 US case, two lawyers were fined $5,000 after filing a legal brief that cited court decisions ChatGPT had invented. Rates have fallen in newer systems, but checking facts, figures, quotes and references remains essential.
Bias. Models learn from human-written data and can reproduce its stereotypes, in hiring tools, image generation or advice.
Deepfakes and fraud. Realistic fake voices and video are now cheap to make. In 2024, an employee of the engineering firm Arup in Hong Kong transferred about US$25 million after a video call in which the "colleagues", including the chief financial officer, were deepfakes. Verifying unusual requests through a separate channel is now a basic security habit.
Copyright. Whether training on copyrighted books, articles and images is lawful is being fought out in court. Cases include The New York Times against OpenAI and Microsoft, and Getty Images against Stability AI. Outcomes so far differ by country and by case.
Energy. Training and running large models uses a lot of electricity. The International Energy Agency's 2025 report projects that data-centre electricity demand will roughly double by 2030, driven largely by AI.

The EU's AI Act entered into force in August 2024. Photo: Profpcde / Wikimedia Commons, CC0
Rules are arriving
The European Union's AI Act entered into force in August 2024. Its rules for general-purpose AI models began applying in August 2025, including transparency about training data and labelling of AI-generated content in certain cases. Other countries are following with their own approaches, from voluntary codes to sector-specific rules.
Using it well
- Treat output as a first draft, not a finished answer. Check facts, numbers and sources yourself.
- Don't paste confidential, personal or client data into public tools unless your organisation allows it.
- Be open about where AI was used in work others rely on.
- Use it where it is strong, such as drafting, summarising, brainstorming and explaining code. Keep humans in charge of decisions that matter.
Generative AI is neither magic nor a passing fad. It is a powerful general-purpose tool whose value depends on how carefully people use it.
Sources
- arXiv:1706.03762, "Attention Is All You Need". https://arxiv.org/abs/1706.03762
- Brynjolfsson E, Li D, Raymond L, "Generative AI at Work", The Quarterly Journal of Economics 140 (2025): 889-942. https://doi.org/10.1093/qje/qjae044
- Noy S, Zhang W, "Experimental evidence on the productivity effects of generative artificial intelligence", Science 381 (2023): 187-192. https://doi.org/10.1126/science.adh2586
- IEA, "Energy and AI". https://www.iea.org/reports/energy-and-ai
Comments (0)
Leave a Reply
Comments are moderated. Please read our Comment Policy before posting.
Log in to post a comment.