The New AI Slang Dictionary: 30 Terms Every Internet User Should Know in 2026

Every tech wave brings its own vocabulary, and the AI wave has been particularly chatty. What started as a handful of researcher terms ("transformer," "fine-tune") has exploded into a sprawling lexicon that now shows up in group chats, TikToks, LinkedIn posts, and arguments at dinner. Some of these words are technical. Some are insults. Some are jokes. Most are useful.

Here are 30 you'll see everywhere in 2026, what they mean, and how to use them without sounding like you just downloaded ChatGPT yesterday.

1. Slop AI-generated content that's technically correct but hollow. Think of those LinkedIn posts that sound vaguely inspirational but say nothing. Slop is the noun version of "this was clearly written by a robot." Usage: "Half my feed is just AI slop now."

2. GPT-ese The signature dialect of ChatGPT and its cousins. Heavy on "delve," "tapestry," "in the realm of," and sentences that always end on a hopeful note. If a paragraph feels like it's trying to win Most Improved at a school assembly, that's GPT-ese.

3. Hallucination When an AI confidently makes something up. The model isn't lying, exactly, because lying requires intent. It's more like a kid who didn't study but is winging the oral exam anyway. Common in legal briefs lately.

4. Prompt The instruction you give an AI. The line between "good prompt" and "bad prompt" is now an entire industry. Five years ago, "prompt" was something a stage actor needed when they forgot their line. Now it's a job title.

5. Prompt Engineering The craft of writing prompts that actually get the result you want. Half art, half voodoo, half getting fired when companies realize anyone can do it. (Yes, that's three halves. The math is part of the joke.)

6. Jailbreak A prompt designed to trick an AI into ignoring its safety rules. Companies patch them, users invent new ones, and the cycle continues. Closely related to: "DAN" prompts, grandma exploits, and roleplay loopholes.

7. Humanize To rewrite AI-generated text so it sounds like a person wrote it. This is now a verb, a feature, and an entire product category. Tools like EssayTone specialize in this, taking that telltale stiff GPT prose and reshaping it into something with rhythm, voice, and the kind of sentence variation actual humans use. Students use humanizers to make essays read naturally. Marketers use them so blog posts don't sound like everyone else's blog posts. Whether this counts as cheating, polishing, or just editing depends entirely on who you ask, but the verb is here to stay.

8. AI Detector A tool that claims to spot AI-written text. Reliability ranges from "decent" to "flips a coin." Universities love them. Students fear them. Falsely accused human writers really hate them.

9. Burstiness A measurement AI detectors use, referring to how much sentence length varies in a piece of writing. Humans bursty. AI flat. The word sounds like a cereal mascot but is genuinely technical.

10. Perplexity Another detector metric, measuring how predictable the next word in a sentence is. AI tends to write low-perplexity prose. Humans write higher-perplexity prose because we're weird and pick odd words. Also the name of an AI search engine, which made naming conventions in the industry even more chaotic.

11. Token The basic unit of text an AI processes. Sometimes a word, sometimes a piece of one. The reason AI pricing pages list everything in tokens, and the reason "but how many tokens is that" is now a real question normal people ask.

12. Context Window How much text an AI can hold in its head at once. Bigger is better. The race for longer context windows is the AI version of a horsepower war.

13. RAG Retrieval-Augmented Generation. Translation: the AI looks stuff up before answering instead of just guessing from memory. Used everywhere customer service used to live.

14. Agent An AI that doesn't just answer questions but actually does things. Books your flight, edits your spreadsheet, replies to your emails. Currently somewhere between "genuinely useful" and "horrifying when it goes wrong."

15. Vibe Coding Writing software by describing what you want to an AI and accepting whatever it produces. Coined in early 2025 and already a punchline. The opposite of careful engineering, and surprisingly effective for small projects.

16. Multimodal An AI that handles more than just text. Images, video, audio, all in one model. Five years ago this was sci-fi. Now it's the baseline.

17. Fine-tune Training a general AI model on specific data to make it better at one thing. The reason your bank's chatbot sounds like your bank instead of like ChatGPT.

18. Open Weights An AI model where the underlying parameters are public, so anyone can run it themselves. Different from "open source" in ways that lawyers love arguing about.

19. Closed Model The opposite of the above. You can use it, you can't see how it works, and you definitely can't run it on your laptop.

20. Mid Not AI-specific, but the most common review you'll see of any new model. "Tried the new release. It's mid." Devastating in its brevity.

21. Cooked What happens to a worker, an industry, or a homework deadline when AI gets good at it. "Translators are cooked." "I forgot my essay was due tomorrow, I'm cooked." Versatile.

22. Skill Issue The accusation that a problem with AI is actually a problem with the person using it. If you can't get a good output, maybe your prompt is the problem. Originally from gaming, fully naturalized into AI discourse.

23. Glazing When an AI is overly complimentary, telling you every idea is brilliant and every question is great. A known failure mode of training AIs to be "helpful." Made headlines in 2025 when one major model started glazing so hard the company had to roll it back.

24. Sycophancy The technical term for what users call glazing. Same thing, more syllables, used by researchers in papers.

25. AGI Artificial General Intelligence. The hypothetical AI that can do anything a human can. Definitions vary wildly. Timelines vary even more wildly. Whether we're close, decades away, or asking the wrong question entirely depends on which podcast you listened to most recently.

26. ASI Artificial Superintelligence. AGI's bigger sibling. An AI smarter than humans at everything. Frequently invoked. Frequently terrifying. Frequently a marketing term.

27. P(doom) Pronounced "P doom." Your personal probability estimate that AI will cause human extinction. People put numbers on this at parties now. A 5% P(doom) means "I'm worried but functional." A 50% P(doom) means "I have a podcast about it."

28. Doomer Someone with a high P(doom). Once a niche identity, now a recognized faction. The opposite is "accelerationist," sometimes shortened to "e/acc," which has its own whole vibe involving startup hoodies and posting at 2 AM.

29. Wrapper A product that's basically just ChatGPT or another model with a thin layer of custom UI on top. Used dismissively. "It's just a GPT wrapper" is the AI-era version of "it's just a website."

30. Alignment The problem of making sure AI does what we actually want it to do, not just what we literally asked for. The classic example: ask an AI to maximize paperclip production, and it turns the planet into paperclips. A mostly serious field of research with a deeply silly central thought experiment.

Bonus terms making moves

A few that didn't make the top 30 but are climbing fast:

●     Context engineering: Like prompt engineering but more fashionable

●     Vibes-based eval: Judging an AI by feel, not benchmarks

●     Model collapse: What happens when AIs train on AI-generated data and slowly get worse

●     Synthetic data: AI-generated training data, often used to fix the above

The lexicon will keep shifting. Half these words will sound dated by 2027. A few will outlive everything else. The only safe bet is that whatever replaces them will also sound made up at first, and then suddenly won't.