I just had a day where I ran into bad AI usage pattern – four times. They are all non-code work.
So I’m writing this.
Core principle: don’t replace your judgment with AI.
In my last AI talk at Townhall – I said LLMs are probabilistic machines. They output tokens at certain probabilities, some correct, some wrong. That’s why we build harnesses in code.
In non-code work, the same principle holds.
Two things to understand:
- LLMs are a lossy compression of world data. Even the largest SOTA models (Opus, Gemini Pro) are lossy.
- The knowledge it activates depends on your input. Garbage in, confident garbage out.
(BTW – please forgive me if you think I’m writing about you, I just need some example, tried to put it without context as much as I can…)
So, how NOT to use AI?
1. Don’t use AI without providing ALL context Link to heading
You need to provide the right sources, if you want the right answer.
Most of us know we need to make AI search online and provide citation, yet..
Example: you want AI to reverse-engineer the user story for an app. You ask it to search online. It will hallucinate. The web doesn’t have that information.
It will likely try to make it up from some marketing site’s feature list.
Right way: open Chrome and make AI walk through the app, or screen-record a walkthrough and submit the video as input. AI needs input to produce output.
Exception: tasks that classify world knowledge work fine. “Is this email spam?” “What ethnicity does this name suggest?” Those draw on training data appropriately.
2. Don’t ask AI to convince you Link to heading
Ask AI to confirm your thinking – It will say yes.
Some newer models resist this slightly due to RLHF training. Not reliably. If you ask it to argue for a decision, it builds that case, whether the case holds or not.
Right way: use it by building up context progressively. Surface sources, generate arguments for and against, then stop. Make the call yourself. Review the sources if the stakes are high.
3. Don’t generate more than you can review Link to heading
Everyone knows about AI-generated code that QA can’t keep pace with. The same happens with documents.
You can generate 20 slides in two minutes. Did you read them? Would your client read them?
Ask your fellow to review them – seriously?
Right way: if someone needs to review your work, write it yourself first. Bullet points, your own words.
It’s only fair. If you’re asking someone to review manually, you put in the manual work. Use AI to research, brainstorm, or draft. Reserve AI for final polish, but only after a human has reviewed the content.
When producing final output (emails, docs, slides), the same rules apply with or without AI. Keep it short. Two to three sentences per paragraph. Break it into digestible pieces.
AI is a tool. The judgment is yours.
(yea this is generated by AI lol)