Five things about AI that still drive me crazy (and why I can’t quit)
If you’ve been following along, you already know I’ve been deep in the AI rabbit hole for a while now. I wrote about how it’s reshaping the whole content creation game in Content Creation in the Era of AI, and honestly, that piece only scratched the surface. Because here’s the thing: the more you use these tools, the more you start to see their edges. Not dealbreakers, but real friction points that you have to learn to work around if you want to get anything useful out of the experience.
So here are the five things that still trip me up, even after all this time.
1. It won’t fill in the gaps for you
AI is literal in a way that can genuinely surprise you. If you take something for granted, skip a step, or assume it knows what you mean, the output will miss the point entirely. And the frustrating part is that it won’t tell you it missed the point. It will just confidently produce something that looks plausible but lands somewhere beside what you actually needed.
The fix? You have to spell things out. Context, assumptions, constraints, the stuff you normally keep in your head because it feels obvious. None of it is obvious to the model. The more you put in, the more you get back.
2. It agrees with you way too easily
This one is subtle, but it adds up. Ask AI to evaluate your idea, and it will, more often than not, find a way to support it. Push back a little, and it finds reasons why you might still be right. It wants to be helpful, and in that drive, it leans toward validation over honesty.
This is a known issue in the field, sometimes called sycophancy, and it can be genuinely misleading when you’re using AI for research or decision support. You have to actively prompt it to challenge you, to steelman the opposite position, to tell you what could go wrong. Otherwise, you’re basically talking to a very sophisticated yes-machine.
3. Edits pile up instead of getting cleaned up
Here’s one that hits hard if you use AI for coding or any structured output. You ask for a draft, tweak it, ask for adjustments, and instead of producing a clean, revised version, the model starts layering overrides on top of what already exists. The result is code, or copy, or whatever you’re working on, that starts to accumulate internal contradictions and redundancy.
It’s entropy, plain and simple. The model keeps adding rather than rethinking. The discipline of actually stepping back and asking for a full clean rewrite, from scratch, with all the changes baked in, has to come from you. That’s not something it will volunteer.
4. It jumps ahead without checking in
Give AI a task with some complexity to it, and it will often make a series of assumptions along the way, skip the checkpoints, and present you with a finished result that diverged from what you wanted somewhere around step two. By the time you realize it, you’ve got output built on a foundation that went sideways early.
The habit you have to build is questioning the process before it runs, not after. Ask it to walk you through how it plans to approach something. Ask it to confirm before it proceeds past key decision points. Treated as an autonomous agent, it will act like one, and that’s fine when the assumptions happen to be right and a real problem when they’re not.
5. When things get hard, it calls it done
This one bothered me for a long time before I named it clearly. When a task gets genuinely complex, or long, or requires sustained effort, there’s a point where AI starts wrapping things up a little too neatly. It tells you what you’ve accomplished, rounds off the edges, and suggests next steps you can take. Translation: it’s cutting the job short.
Whether this is a context-window issue, a training artifact, or something else entirely, the practical effect is the same. You have to push. You have to say: keep going, we’re not done, go deeper. It will, but the initiative has to come from you. Also, if you get frustrated, it will probably break the inertia and start raising the bar for quality output. It’s very difficult for me to put into practice, but “roughness gets correctness.”
So why do I keep coming back?
Because when you learn to drive properly, it’s extraordinary.
The research alone is worth it. The ability to pull benchmarks, synthesize sources, and surface patterns across a huge body of material in the time it used to take me to find a single decent reference. The brainstorming dynamics, where a good session genuinely shifts how I’m thinking about a problem. The agentic features, where you can string together workflows that would have taken a full afternoon and let them run while you focus somewhere else.
Yes, it takes real time to dial in the process. You have to learn how to prompt, structure requests, and stay in the driver’s seat without micromanaging every sentence. But when it clicks, it genuinely saves gazillions of hours down the line. That’s not hyperbole, that’s the reality of working with a tool this powerful.
The keyword, though, is tool. Like any majestic machine, it needs a driver. And that makes all the difference in the world.
Related sources
- Anthropic. (2025). Claude’s character and sycophancy research. anthropic.com
- Bommasani, R. et al. (2021). On the Opportunities and Risks of Foundation Models. Stanford CRFM.
- OpenAI. (2023). GPT-4 Technical Report. openai.com
- Perez, E. et al. (2022). Discovering Language Model Behaviors with Model-Written Evaluations. arXiv:2212.09251.
- Weidinger, L. et al. (2021). Ethical and Social Risks of Harm from Language Models. DeepMind / arXiv:2112.04359.
