I've spent several issues writing about what AI can do. It seemed like the right order — lead with the capability, establish credibility, then get into the limitations. But today I want to spend the full issue on the limitations, and I want to mean it, not just include a caveat section at the end of an otherwise enthusiastic piece.
The limitations matter as much as the capabilities for anyone who's trying to build something real with these tools. Understanding where AI falls short is not pessimism about AI. It's the prerequisite for using it well.
AI cannot tell you what's actually true about the physical world.
This one sounds obvious, but its implications are more significant than people generally appreciate.
A large language model has been trained on text. It knows what has been written about the world. It does not know what is currently true about the world in the way that you know the coffee in your hand is hot, or a reporter knows what they saw with their eyes, or a doctor knows what they found in an exam.
For most purposes this is manageable — verify anything that matters, use AI as a research starting point not an endpoint. But there's a class of tasks where this limitation is more structural. Real-time market conditions. Current events. The current status of anything that changes. What the person on the other end of a negotiation actually thinks. What a customer actually needs versus what they say they need. AI doesn't know any of these things and can only approximate them through patterns from historical text.
This is why I described AI as "a research tool, not an oracle" in the paper trading issue. For investing specifically, the gap between "what has been written about markets" and "what is actually true about this market right now" is financially significant.
AI cannot exercise genuine ethical judgment.
AI systems can produce ethical-sounding outputs. They can reason about ethical questions in ways that are often sophisticated and useful. They can flag potential ethical concerns in a plan you present.
They cannot bear moral responsibility. They cannot genuinely care about an outcome. They have no stake in whether the thing they help you produce is good or harmful, beyond the constraints they've been designed with.
This matters practically. When I use AI to draft something, the editorial responsibility for whether it's accurate, fair, and honest is mine. I cannot delegate that to the tool. "The AI said it" is not a defense in journalism, law, or any professional context — and it's not a defense in a newsletter that's staking its reputation on accuracy and candor.
The people building with AI who I find most trustworthy are the ones who understand this most clearly. The AI augments their judgment. It doesn't substitute for it.
AI cannot build relationships.
I've touched on this before but it deserves direct treatment. The income streams that have the highest long-term defensibility are the ones built on trust relationships — reader trust in a newsletter voice, community trust in the norms of a space, local business trust in a directory that reliably sends them customers.
AI can help you produce content efficiently. AI can help you manage communication at scale. AI cannot be trusted by a human being in the way that a human being can be trusted.
This is not a limitation to work around. It's a signal about where human effort belongs in an AI-augmented operation: at the relationship layer. That's where leverage lives.
AI cannot do genuinely original creative work.
This claim requires some precision. AI can produce creative work that is surprising, aesthetically sophisticated, and sometimes genuinely beautiful. It can combine existing elements in ways that don't feel derivative.
What it cannot do is start from a genuine novel experience of the world and make something from that. The Westerns I write are rooted in something I actually know and feel — the specific quality of light on the Missouri River at dusk, the way certain kinds of silence sound in open country, the emotional structure of choosing between loyalty and principle. I cannot prompt that into existence. I lived it.
The creative work that AI cannot replace is the creative work grounded in irreducibly individual experience. Which means the path forward for writers and creators is not to compete with AI on production volume — it is to deepen whatever is specifically and irreducibly theirs.
Why this matters enough to put in a guide.
We've put together a guide called "AI Without the BS" — an honest summary of what AI does well, what it doesn't, and how to think about it without either the techno-utopianism or the reflexive skepticism that both prevent clear thinking.
The guide is aimed at people who want to actually use these tools rather than have opinions about them — and who want to build their use on an accurate model of what they're working with.
This Week in AI: Several prominent researchers published papers this year examining the "hallucination" problem in large language models — the tendency to produce confident, fluent, plausible-sounding output that is factually incorrect. Progress has been made; the problem has not been solved. For practical purposes: treat AI outputs as claims that require verification on anything consequential, not as facts.
The "AI Without the BS" guide is free in the toolkit at start.tenstreamslab.com. It's the one I'd give to someone who's genuinely trying to figure out what to think about AI right now.