LLM Authorship Is Obvious: What It Means for AI Agents

Bryan Cantrill's post lands a simple, uncomfortable observation: when someone uses an LLM to write something they're putting their name on, readers can usually tell. And the tell isn't flattering.

That's worth sitting with if you're building AI agents that produce outputs your business stands behind.

What Cantrill Is Actually Pointing At

The post isn't really about writing quality. It's about epistemic ownership — whether the person publishing something actually understands what they're publishing. LLMs produce fluent, confident prose that can outrun the author's actual knowledge. The gap between what got written and what the author can defend is visible to anyone paying attention.

For content, that's embarrassing. For a business deploying agents, it can be worse.

The Same Problem Shows Up in Agent Outputs

Most early agent deployments fail quietly for the same reason. The agent produces something plausible. Nobody checks it closely because it reads fine. Then it ships — a customer-facing summary, a support response, a generated report — and it's wrong in a way that matters.

The fluency is the trap. Output that sounds authoritative discourages scrutiny. Teams that would carefully review a junior hire's work let the agent's output through because it doesn't trigger the same skepticism reflex.

This is a known failure mode with a name in the evals literature: confident incorrectness. The model has no internal signal that it's about to hallucinate. It just generates the next token.

What Responsible Agent Deployment Actually Looks Like

The fix isn't to avoid agents. It's to build the right checks into the workflow.

A few mechanisms that work in practice:

Grounded retrieval over generation. If your agent is summarizing, have it pull from a specific document or dataset and cite the source. Fabrication drops sharply when the model has actual text to work from rather than training-time memory.

A second pass that grades the first. One agent generates, a second agent checks the output against a rubric or source document before it surfaces to a user. This adds latency but catches the obvious failures. It's how production evals are often structured.

Human-in-the-loop for consequential outputs. Not every agent output needs a human review step, but the ones that go to customers, get logged in a CRM, or inform a decision probably do — at least until you have enough eval data to trust the pipeline.

Narrow the task. Agents that do one specific, bounded thing fail less than agents asked to do something broad. "Extract the three action items from this transcript" is more reliable than "summarize this meeting and suggest next steps."

Our free AI Opportunity Audit is built around this principle — it identifies your three highest-impact automations from your website alone, so you're starting with something specific and bounded rather than trying to automate everything at once.

The Trust Problem Is a Design Problem

Cantrill's core point is about trust. Once readers clock that you're not really the author of your own posts, the relationship changes. The same dynamic plays out with agent outputs at scale.

If your customers or your team can't tell whether an agent's output is reliable, they'll either stop trusting it or stop checking it. Both outcomes are bad. The first kills adoption. The second accumulates invisible errors until something breaks visibly.

The agents worth building are the ones where you can explain, specifically, why the output is trustworthy: what data it draws from, what checks it passed, where a human reviewed it. That's a harder design problem than prompting an LLM and shipping the result. It's also the difference between an agent that helps your business and one that creates liability.

The Takeaway for Builders

Fluency is not accuracy. Confidence is not correctness. The most useful thing Cantrill's post does is name the tell: you can see when someone doesn't own what they published. Apply that same scrutiny to what your agents are producing, before your customers do.

Build It Right

If you want to deploy agents where the outputs are actually trustworthy, book a call — that's the kind of agent work we do.