So ChatGPT walked into the newsroom, and now everyone’s acting like they’ve discovered fire. Here’s what I wish every journalist (and everyone else) knew about AI. This is my practical introduction, without the hype.
- When we say “AI,” we’re often talking about ChatGPT and its digital siblings. Large language models that have basically eaten the entire internet and can spit out surprisingly human-sounding text. Before these general-purpose machines showed up, AI was like that one friend who’s really good at exactly one thing (recognizing cats in photos, maybe?) but completely useless at everything else. To distinguish between the different kinds, and because this new type generated text, it’s called generative AI.
- Think of them as the world’s most expensive compression algorithm. They’ve taken all this text that “somehow” ended up in their training data and turned it into pure math. Probabilistic math. Then humans had to teach this math monster how to chat like a normal person instead of just word-vomiting statistical predictions.
- Every conversation is like chatting with someone who has short-term memory loss. When you’re typing away to ChatGPT, that’s called inference—you lob a prompt at it, it lobs an answer back. If it seems to remember stuff from earlier in the chat or previous conversations, that’s just because the chat history gets fed back into it every single time. (The company behind it could store your data and use it to train its next big model. You can sometimes opt out of training, especially with enterprise accounts.)
- Don’t get too distracted by all the talk about AGI or artificial general intelligence. It’s mostly a marketing ploy to raise money and avoid regulation. Nobody really knows what it actually means.
- Academic papers are living in the past while the rest of us are in 2025. When you see studies, note the actual models they used for testing. Academia moves slower than tech—some papers make wild claims based on ChatGPT 3.5, which was released in November 2022 and has long been superseded.
- Yes, they’re “stochastic parrots”—and so what? If you hear stochastic parrots and someone tells you how dumb the machines actually are—well, yes. But millions of people are using them anyway. (And parrots aren’t necessarily bad authors?!) Take it from tech analyst Ben Evans: “Some people are so pleased with themselves for discovering that LLMs are statistical systems with error rates that they forget to notice that everyone in the field knows this, and more importantly, forget to notice that this doesn’t make them useless.”
- Breaking: AI discovers the internet exists. For AI chatbots to give accurate and up-to-date answers to specific questions, they need access to external knowledge. This could be the internet—ChatGPT and similar systems can now access websites in real-time and use the information they find to generate responses. (A bot visits the website, not a human.) Alternatively, you can connect a vector database. When the answer generation draws from these external sources, it’s called retrieval-augmented generation, or RAG.
- Garbage in, garbage out. The quality of your output is proportional to the quality of your input. If you want decent results, you need to actually think about your prompts. Tell it what role to play, give it clear instructions, show it examples of the desired output. There are many frameworks for this prompt engineering. But because humans are trained to think in Google keywords and are slow to adapt, newer models try to help us. This is especially true for “reasoning” models like o3, where short questions are artificially expanded, or AI search engines that broaden queries to provide better results.
- Chatbots are just the opening act. Interfacing with AI through chatbots is just the beginning—and may turn out to be an aberration. More and more, AI features will hide behind simple buttons. Or we’ll use our voice to interact with machines. More importantly, AI will try to solve complex tasks on its own. Let’s talk agents. We can give AI access to systems (connect via MCP to APIs). If we replicate a predefined workflow (using tools like n8n or Langflow), we might call that “agentic,” even though it’s constrained and limited in what it can do. But the more degrees of freedom we give the machine, the more of an “agent” it becomes. There are programming agents like OpenAI’s Codex that work for many minutes on complex tasks, testing and iterating on their own.
- The elephant in the server room: energy consumption. Data centers are being built left and right, and it’s fair to assume that model training and generating images and videos cost a lot of energy. In 2024, data centers consumed 1.5 percent of the world’s electricity. That figure could double by 2030. Then there’s water usage for cooling the servers. How much does a single text prompt use? Hard to say. Models are getting more efficient—OpenAI CEO Sam Altman claims that an average query today consumes “roughly one-fifteenth of a teaspoon” of water and “0.34 watt-hours” of energy. For context, that’s… actually, I have no idea. How much energy does watching Netflix use? 80 watt-hours for an hour? Cue the whataboutism.
Want more perspectives?
In my newsletter THEFUTURE, journalist and technologist Jaemark Tordecilla emphasizes: “Most of the time, AI solutions are really workflow solutions.” Think-tank director Dirk von Gehlen argues that “often the processes are just as valuable as the products that end up being created. The specifically human aspect is often only created through the journey.” And Laurens Vreekamp, founder of Future Journalism Today, warns: “The architecture and infrastructure needed to build and maintain LLMs have serious limitations – and that results in problematic flaws.”
One final reality check:
