Verbal Knobs
we owe the answers we proxy

Large language models (LLMs) are pre-trained by analysing patterns in massive datasets of billions of examples. Through additional fine-tuning, they learn to follow instructions and output helpful responses¹. But these models have fundamental limitations. Their knowledge freezes at a training cutoff date, and they generate text probabilistically rather than retrieving verified facts². This leads to what we call "hallucinations," plausible-sounding responses that aren't grounded in reality. It's not a bug in the training process. It's fundamental to how these systems work³. They're pattern-matching engines, not knowledge databases. And that distinction matters when we start relying on them for things that actually count.

My first interaction with generative AI was with an early ChatGPT model that hallucinated a solution to a programming problem I'd asked about, making up an NPM package that didn't exist. Fast forward to now, and Cursor and Claude sit at the top of my daily toolbox. It's easy to get caught up in the speed and scale these tools provide, to let trust creep in until something reminds you: for all their pattern recognition and processing power, they're still fundamentally dependent on you. Your critical thinking. Your judgment. Your creativity. Your ethical oversight. It’s easy to get caught up and accept eloquent B.S. as truth, but the truth is that, at its core, AI isn't a being. As good as it is at predicting answers from ambiguity, prompts are really just verbal knobs, configurations, and filters in a query. The more specific I am about what I want, how I want the AI to get outside information, and how I expect the output, the better I can tune those knobs toward something useful.

Once you accept the limitations of generative AI, and more importantly, your role and responsibility, the trade becomes clearer. You give up the illusion of a magic oracle, but you gain speed, scalability, and processing power that's genuinely useful when paired with your judgment. The tool becomes more trustworthy because you're not asking it to be something it isn't. What frustrates me isn't the hallucinations or the limitations themselves but how generative AI is being marketed and deployed. The promise that it'll make you rich by doing your job for you. The suggestion that it's coming to replace you rather than amplify what you can do. That framing doesn't just misrepresent the technology, but it encourages people to abdicate the very responsibilities that make these tools valuable. Critical thinking becomes optional. Verification becomes an afterthought. Ethical oversight gets automated away. And when things go wrong, when the output is factually incorrect or ethically questionable, the blame gets diffused to a lack of government oversight.

It's tempting to place all the responsibility on companies like Anthropic and OpenAI, which develop and sell these models. Don’t get me wrong, they should be transparent about limitations, build safeguards, and be held accountable. But there's another side to this: personal responsibility for how we use these tools. As much as there's responsibility on the company providing us with the tools and how these tools affect us, we also own what comes out through the other end and what we produce with it. The output we create, the work we put our name on, the solutions we proxy along to colleagues or clients—that's on us. We can't outsource critical thinking or delegate judgment calls⁴. We can't automate ethical oversight. Those remain fundamentally human responsibilities, regardless of how advanced the flirty pattern-matching engine becomes. The truth is, generative AI depends more on us than we depend on it.