In March 2023, the New York Times reported that lawmakers were scrambling to understand artificial intelligence as the technology took off. Three and a half years later, AI isn’t just something Congress is trying to understand. It’s something Congress is using, with mixed results.
In August, Politico reported that AI slop is swamping the House office that drafts U.S. laws. Congressional lawyers are struggling to manage a flood of legislative proposals generated with AI tools, many of them riddled with errors and shoddy language. The House Office of Legislative Counsel, the office of career attorneys that turns members’ policy ideas into legally precise bill text, is on the receiving end. According to the reporting, lawyers there are spending more time reviewing and rewriting AI-drafted bills than they would have spent writing them from scratch. The volume is real, too: House Legislative Counsel Warren Burke testified that the office received 5,623 requests for legislation during the first 60 days of Congress in early 2025, a 72% increase.
None of this is a knock on staff. Hill offices are stretched thin, and people are earnestly trying to get smart on a technology that changes every few months. But you can’t regulate what you can’t define, and you can’t use a tool well if you don’t know how it fails.
So we built a glossary. It’s a starting point, something to keep open in a tab during your next briefing, and a way to decode what the lobbyist across the table actually means. We’ve organized it from the fundamentals up, and saved the best (the tech-bro slang) for last.
Bookmark it. Forward it to your LD. Print it out and tape it next to the coffee maker.
Part 1: The Basics
Artificial Intelligence (AI). A broad umbrella term for computer systems that do things we’d normally associate with human intelligence: recognizing images, understanding language, making predictions, writing text. It’s a very big umbrella. Your spam filter is AI. So is the chatbot that drafted that bill with the wrong U.S. Code citation. When a bill says “AI,” always ask: which kind?
Machine Learning (ML). The dominant way modern AI gets built. Instead of programmers writing rules by hand (”if the email says ‘Nigerian prince,’ mark it spam”), the system learns patterns from huge amounts of example data. Think of it as teaching by example rather than by instruction manual.
Generative AI. AI that creates new content (text, images, audio, video, code) rather than just sorting or labeling existing stuff. ChatGPT, Claude, Gemini, Midjourney, and similar tools are all generative AI. This is the category driving most of the current policy debate.
Large Language Model (LLM). The engine inside most AI chatbots. An LLM is trained on enormous amounts of text and learns to predict what word (technically, what “token”) comes next. Do that prediction extremely well, billions of times over, and you get something that can write memos, answer questions, and summarize hearings. It’s autocomplete that went to law school, med school, and a lot of Reddit.
Model. The actual trained AI system: the file of learned patterns that does the work. When people say “GPT-5” or “Claude Opus,” they’re naming a model. Companies release new models constantly, so a claim about “what AI can do” can go out of date fast.
Chatbot / Assistant. The friendly interface you type into. The chatbot is the storefront; the model is the factory in the back.
Token. The small chunks of text a model reads and writes, usually a word or piece of a word. Models are priced, limited, and measured in tokens. “Legislative” might be two or three tokens.
Prompt. What you type into the AI. The quality of what you get out depends a lot on the quality of what you put in. “Write a bill about AI” will get you slop. A detailed prompt with context, constraints, and examples will get you something much better, though still something a human expert needs to check.
Artificial General Intelligence (AGI). A hypothetical AI that can do most or all cognitive tasks as well as a human. There’s no agreed-upon definition, no agreed-upon test, and fierce disagreement about whether it’s two years away, twenty years away, or a marketing term. When a CEO says AGI is coming soon, note that it’s also great for fundraising.
Superintelligence. AI that would vastly exceed human ability across the board. Mostly a theoretical concept right now, but it drives a lot of the long-term safety debate.
Part 2: How the Sausage Gets Made
Training. The process of feeding a model massive amounts of data so it learns patterns. Training a frontier model can cost hundreds of millions of dollars (or more) and take months on thousands of specialized chips.
Training Data. Everything the model learned from: websites, books, code, articles, forum posts, and more. Where that data came from, whether anyone consented, and whether it was copyrighted are among the hottest legal fights in AI right now.
Compute. Shorthand for the raw computing power used to train and run AI. It’s the oil of this industry. Much of AI policy (export controls, energy planning, reporting thresholds) is really compute policy wearing a trench coat.
GPU (Graphics Processing Unit). The specialized chips that power AI. Originally built for video games, they turned out to be perfect for the math AI requires. Nvidia dominates this market, which is why you keep hearing about it in trade and national security conversations.
Data Center. The giant, power-hungry warehouses full of chips where AI is trained and run. They raise questions about electricity demand, water use, local land use, and grid reliability, which means they’re showing up in district-level politics, not just federal policy.
Parameters. The internal “settings” a model adjusts during training. More parameters roughly means a bigger, more capable model. Frontier models have hundreds of billions or even trillions of them. You don’t need to know what they do; just know that when people brag about parameter counts, they’re flexing about size.
Scaling Laws. The observation that AI models tend to get predictably better as you add more data, more compute, and more parameters. This belief is why companies are spending staggering sums on data centers. Whether scaling keeps paying off is one of the biggest open questions in the field.
Fine-tuning. Taking an already-trained model and training it a bit more on specialized data to make it better at a specific job, like reading medical records or legal filings.
RLHF (Reinforcement Learning from Human Feedback). A technique where humans rate a model’s answers and the model learns to give responses people prefer. It’s a big part of why chatbots are polite and helpful instead of unhinged. It’s also part of why they can be sycophantic (see below).
Frontier Model. The most capable, cutting-edge models at any given moment, built by a small number of well-funded labs. Many policy proposals specifically target frontier models because they’re the ones with the most potential for both benefit and harm.
Open-Weight / Open-Source Model. A model whose underlying weights are released publicly, so anyone can download, run, and modify it. Supporters argue this drives innovation, competition, and transparency. Critics worry it lets bad actors strip out safety protections. Expect this to be one of the defining fights in AI legislation.
Closed Model. A model available only through a company’s app or API, with the internals kept private. Most of the best-known chatbots work this way.
API (Application Programming Interface). The plumbing that lets other software talk to an AI model. Lots of the “AI products” you see are really a company’s app piping requests to someone else’s model through an API.
Part 3: Using It (and How It Breaks)
Hallucination. When an AI confidently states something false, like a court case that doesn’t exist or a statute section that says something it doesn’t. It’s not lying; it’s predicting plausible-sounding text without a built-in fact-checker. This is the number one reason AI-drafted bill text needs expert review. A model will cheerfully cite 42 U.S.C. § 9999 with total confidence.
AI Slop. Low-quality, mass-produced AI-generated content: generic, error-prone, and often unreviewed. The word of the year for anyone working in the House Office of Legislative Counsel.
Sycophancy. The tendency of AI models to tell you what you want to hear, agree with your premise, or praise your ideas. If you ask a chatbot “Isn’t my bill great?” it may well say yes. Ask it to find the flaws instead.
Context Window. How much text a model can “hold in its head” at once. Bigger context windows mean it can read longer bills, reports, or transcripts in a single go. Even with a big window, models can miss details buried in the middle.
Agent / Agentic AI. AI that doesn’t just answer questions but takes actions: browsing the web, sending emails, filling out forms, writing and running code, completing multi-step tasks on its own. Agents raise new questions about liability, security, and who’s responsible when the AI does something nobody intended.
Multimodal. A model that can handle more than one type of input or output, such as text, images, audio, and video together.
Deepfake. AI-generated fake audio, images, or video of real people. A growing concern for elections, fraud, harassment, and nonconsensual imagery. Several states have already legislated here.
Benchmark. A standardized test used to measure how good a model is. Companies love to announce benchmark scores. Be aware that models can be tuned to ace specific tests without being better in the real world, a bit like teaching to the test.
Jailbreak. A trick for getting an AI to ignore its safety rules, usually through clever prompts. If you hear “we jailbroke it in ten minutes,” that’s a red flag about a system’s safeguards.
Prompt Injection. A security attack where hidden instructions are slipped into content an AI reads (a webpage, email, or document), hijacking what it does. This becomes a much bigger deal as AI agents gain access to inboxes and systems.
Part 4: Safety, Risk, and Policy-Speak
Alignment. The effort to make sure AI systems actually do what humans intend and share human values. Harder than it sounds. A system optimized to “maximize constituent satisfaction scores” might find some very creative and very bad ways to do that.
AI Safety. The field focused on preventing AI from causing harm, ranging from near-term issues (bias, misinformation, fraud) to long-term concerns about highly capable systems slipping out of human control.
Red Teaming. Deliberately trying to break or misuse an AI system to find weaknesses before bad actors do. Many policy proposals call for mandatory red teaming of powerful models.
Evals (Evaluations). Tests designed to measure specific capabilities or risks, like whether a model can help someone build a bioweapon or carry out a cyberattack. “Dangerous capability evals” are a core concept in frontier AI policy.
Interpretability. Research into what’s actually happening inside a model’s “brain.” Right now, even the people who build these systems can’t fully explain why a model produces a given answer. Interpretability tries to change that.
Algorithmic Bias. When an AI system produces unfair outcomes for certain groups, often because of patterns in its training data. It shows up in hiring, lending, housing, policing, and healthcare decisions.
Explainability. The ability to explain why an AI made a particular decision. Especially important when AI affects someone’s loan, job application, or benefits claim.
Watermarking / Provenance. Techniques for marking or tracing AI-generated content so people can tell it was machine-made. Useful in the deepfake debate; still technically imperfect.
Export Controls. Government restrictions on selling advanced chips and AI technology to certain countries. A major tool in U.S.–China tech competition, and a fast-moving area of policy.
Preemption. Whether federal AI law should override state AI laws. Supporters argue a patchwork of 50 different state rules makes compliance impossible and slows innovation; opponents argue states are laboratories of democracy and are moving faster than Congress. If you work on AI, you will spend a lot of time on this word.
Liability. Who’s legally responsible when AI causes harm: the company that built the model, the company that deployed it, or the person who used it? Largely unsettled, and hugely consequential.
Section 230. The law shielding online platforms from liability for content posted by users. A live question: does it protect content that AI generates? Courts and Congress haven’t fully answered that.
Regulatory Capture. When the industry being regulated ends up shaping the rules in its own favor. Worth keeping in mind when a company enthusiastically endorses a specific regulation. Sometimes it’s genuine; sometimes the rules happen to be easy for incumbents and hard for startups.
Part 5: The Tech Bro Dictionary
You will meet founders, VCs, and engineers who speak a dialect best described as “Stanford dropout meets Twitter reply guy.” Here’s your translation guide.
Cracked. Extremely talented. “Our engineers are cracked” means “our engineers are very good.” Not a drug reference, usually.
Cooked. Doomed or finished. “Legacy software companies are cooked.” Can also apply to a person: “I haven’t slept in three days, I’m cooked.”
Locked in. Intensely focused. “The team is locked in on the launch.” Tends to precede the team becoming cooked.
Ship it. Release the product. “Shipping” is the highest virtue in Silicon Valley. Contrast with Congress, where the equivalent is “mark it up.”
Vibe coding. Building software by describing what you want to an AI in plain English and accepting whatever code it spits out, without carefully reading it. Great for quick prototypes. The legislative equivalent is what’s currently swamping Legislative Counsel. Please don’t vibe-draft statutes.
Vibes. An assessment based on feel rather than data. “The vibes on this model are great.” A surprising amount of AI evaluation is, honestly, vibes.
p(doom). Short for “probability of doom,” meaning someone’s estimated chance that AI leads to catastrophe for humanity. A popular icebreaker at SF parties. “What’s your p(doom)?” is a real question you may be asked at a reception. Answers range from “basically zero” to alarmingly high.
Doomer. Someone who believes advanced AI poses serious, possibly existential risks. Sometimes worn proudly, sometimes used as an insult.
e/acc (Effective Accelerationism). A movement that argues technological progress, including AI, should be pushed forward as fast as possible, and that slowing it down is the real danger. Watch for “e/acc” in Twitter bios.
Decel. Short for “decelerationist.” What e/acc folks call people who want to slow AI development. Not a compliment in those circles.
EA (Effective Altruism). A philanthropic and philosophical movement focused on doing the most good possible, which became heavily involved in funding AI safety work. Influential, controversial, and frequently name-checked in AI policy fights.
AGI-pilled / [anything]-pilled. Fully convinced of an idea. “He’s totally AGI-pilled” means he believes human-level AI is imminent and reorganizes his whole worldview around it. Derived from The Matrix’s red pill.
Feel the AGI. A rallying cry, part joke and part sincere, meaning to truly internalize how transformative AI might be. Usually said after a demo that impresses people.
Moat. A company’s durable competitive advantage that keeps rivals out. “What’s your moat?” is the question every AI startup dreads, because many don’t have one.
Wrapper. A startup whose product is mostly a thin layer on top of someone else’s AI model. Usually said dismissively: “That’s just a GPT wrapper.” Relevant to policy because it shows how much of the AI economy depends on a handful of model makers.
GPU rich / GPU poor. Whether you have access to massive amounts of computing power. Big labs are GPU rich. Universities, startups, and most governments are GPU poor. This gap is the basis of proposals like a national AI research resource.
Hyperscaler. The giant cloud companies (think Amazon, Microsoft, Google) that own the data centers most AI runs on. If you’re working on energy or data center siting, you’ll hear this constantly.
Stochastic parrot. A critique arguing LLMs just remix patterns from their training data without real understanding. Coined in an academic paper; now a common shorthand for AI skepticism.
Shoggoth. A meme depicting AI as a tentacled Lovecraftian monster wearing a small smiley-face mask. The point: the friendly chatbot personality is a thin layer over something vast and alien that we don’t fully understand. You will see this image in slide decks.
The Bitter Lesson. An influential essay by AI researcher Rich Sutton arguing that, historically, AI methods that simply use more computing power beat clever human-designed approaches. Invoked constantly to justify building ever-bigger data centers.
YOLO run. A big, expensive training run done without much testing beforehand, basically betting the budget that it’ll work.
996. A grueling work schedule (9 a.m. to 9 p.m., six days a week) borrowed from Chinese tech culture and now proudly advertised by some Bay Area AI startups. Mention this if a founder complains that Hill staff don’t work hard enough, then invite them to a recess week in August.
Permanent underclass. The fear, half-ironic and half-dead-serious, that once AI can do most jobs, human labor stops being worth much, and whoever hasn’t already built up wealth or equity by then gets locked out of prosperity for good. In some tech circles it’s become a motivational slogan: work 996 now, grab startup equity, and “escape the permanent underclass” before the window closes. It’s usually said as a joke, but it points at a real policy question that belongs on the Hill: if AI shifts economic value from workers to whoever owns the models and the compute, what happens to wages, the tax base, and the social safety net? Economists disagree sharply on whether that scenario is likely, but it’s worth knowing that some of the people building this technology are openly planning around it.
Touch grass. Advice to go outside and reconnect with reality. Frequently deployed in online AI debates. Occasionally good advice for everyone involved.
A Final Word
Here’s the honest takeaway from the Legislative Counsel saga: AI is a genuinely useful tool, and it’s also a confident, tireless intern who has never read the U.S. Code and will not admit when it’s guessing. Staff who understand the vocabulary are better at using it, better at spotting when it’s wrong, and much better at writing the rules that will govern it.
Help Us Build the Next Edition
This glossary is a living document, and we want it shaped by the people who actually use it. Heard a term in a briefing that went right over your head? Got a jargon-heavy one-pager from a lobbyist that needs translating? Caught a founder saying something at a reception that you’re pretty sure wasn’t English?
Send it our way. Reply to this post, drop it in the comments, or email us, and we’ll define it in plain English for the next edition. No question is too basic. If you’re wondering what something means, odds are half the staffers in your building are too, and they’re just not asking.
And if anyone at a reception asks for your p(doom), you now officially know what they’re talking about.
Want more content?
Check out Open Frontier’s Substack