From avoid-ai-writing (opens in a new tab)

AI writing, evaluated

The tool you came from flags the patterns that make text read as machine-written. Those are fixable. The harder problem sits under them: a model trained to produce plausible text sounds right whether or not it is right.

Built by the show, for the show. Chain of Thought host Conor Bronsdon wrote avoid-ai-writing to check the podcast's own show notes, essays and posts, then open-sourced it. Listen to the show Get the newsletter

Our new retrieval layer marks a pivotal moment in the evolution of enterprise AI^{(pattern 1: Significance inflation)}. Experts agree^{(pattern 2: Vague attributions)} it plays a crucial role in reducing hallucinations. It’s not just a search index, it’s a foundation for trust.^{(pattern 3: Sentence structure)} The future looks bright.^{(pattern 4: Generic conclusions)}

1 Significance inflation Claims historic weight without saying what changed.
2 Vague attributions Which experts? With no source, nobody can check it.
3 Sentence structure The “not X, it’s Y” frame rebuts a claim nobody made.
4 Generic conclusions Could end any paragraph about anything.

A made-up paragraph with four patterns from the avoid-ai-writing catalog marked. Fixing them makes it read better. It can't tell you whether the retrieval layer reduces hallucinations at all, and that second question is what the essay and episodes below are about.

Essay Conor Bronsdon On the newsletter I Paid an AI Agent $8 to Write About its ‘Life’ I paid an AI agent $8 to write about what it costs to be an agent. Its first draft runs with formatting fixes only, so you can read it for tells and decide whether it was worth the money.

Three episodes on the gap between plausible and true

EP 54 Dan Klein, Scaled Cognition

Why LLMs Are Plausibility Engines, Not Truth Engines

“If it’s plausible enough, it’s going to be indistinguishable from the truth.”

The clearest statement of the problem the detector exists for. A model trained to produce likely text is not trained to produce true text, and past a certain quality the two stop being separable by reading.

Show notes and full transcript →
EP 55 Sudhir Hasbe, Neo4j

Hallucinations Are a Data Architecture Problem

The case that hallucination is a failure in the retrieval layer rather than a flaw in the model, which changes where you go looking when the output is confidently wrong.

Show notes and full transcript →
EP 57 Alex Ratner, Snorkel AI

Every AI Agent Has an Evaluation Gap

Everyone says their system works; almost nobody can say what “works” means in a number. The same reason a detector you can run beats a vibe check.

Show notes and full transcript →

The newsletter

The same arguments, in writing

Up to one email a week: essays drawn from these conversations, plus a note when a new episode is out. Unsubscribe anytime.

Form not loading? Get the newsletter →

What this is

Chain of Thought is where builders reason through what's changing in AI and software infrastructure. Every episode carries its full transcript on the page, so you can read one instead of listening to it.

Recent guests