The Real Issues with AI
The concerns worth taking seriously before you adopt these tools. None of these are reasons to never use AI — they're things to carry with you when you do.
Environmental impact
Two things are true at once: per-prompt impact is small and falling fast, and aggregate impact is rising sharply because adoption is scaling faster than efficiency. Personal use looks negligible. Collective use is anything but.
When you use the tool
One typical text prompt to a frontier model in 2025 used roughly:
And these per-prompt numbers are dropping fast. Google reports a 33× reduction in energy per median prompt over twelve months. The often-repeated "AI chat uses 10× as much as a web search" claim is contested in 2025 literature — for simple chat queries, the two are now in the same order of magnitude.
Reasoning queries, long contexts, and image or video generation are materially higher — by orders of magnitude in some cases. Top-end model prompts can hit ~29 Wh. The "small individual impact" framing breaks for those workloads.
When everyone uses the tool
At the individual level, a prompt looks negligible. At the system level, billions of prompts a day add up to something else entirely.
Per-query efficiency keeps improving. Aggregate consumption keeps rising. Each 10% efficiency gain has historically driven a 20–30% increase in deployment. More efficient does not mean less impact at scale. The hopeful per-prompt curve and the alarming aggregate curve are the same story.
The resource these tools consume is real. The resource they can free up — human ingenuity, ideas, time — is also real. We are not resolving this trade-off for you. We are naming it plainly. Every figure above is drawn from publications between mid-2025 and early 2026; both efficiency and adoption are changing rapidly, so re-check before quoting any of these numbers.
Copyright & ownership of ideas
LLMs are a compression of an enormous amount of training data that originated with humans. The line between ownership, attribution, and general knowledge has become even fuzzier. This matters for anyone producing original materials, translations, and educational content. We don't have a clean answer. We have an awareness.
Hallucinations & getting things wrong
- The latest models get factual claims wrong far less often than a year ago. This is improving fast.
- Not improving as quickly: they do not natively apply a framework like Holistic Management correctly.
- The question worth asking: do we reduce our effectiveness by using these tools when applying our framework?
Atrophy of holistic thinking
An analogy many practitioners already understand: doing your own EOV monitoring vs. hiring professional monitors. When you walk the land yourself, the act of monitoring is the act of learning. When you outsource it, you may lose the discipline.
The same risk applies to AI. Lean on it as a crutch for the hard work of holistic planning, and you may atrophy the mental muscles you need to do that work well over time.
Concentration of power
Using AI tools injects an outside decision-maker's influence into your operations and workflows. A few companies offer these services, which means a few companies hold significant leverage over how the tools behave. This isn't new — but the magnitude is.