The benevolent lie — how much hallucination do you allow in an AI-written post?

2026-06-01 · Ascendy Engineering

TL;DR

An AI-written post slipped in a detail that wasn’t in the source (“3am, the third OOM hit”). It reads better — but I, the author, don’t know if it’s true.
Human memory is itself a kind of hallucination — imprecise, distorted over time. Vibe-coding makes it worse: I made the code, but I didn’t type it, so I can’t remember code my fingers never touched.
So the line has two layers: don’t invent what a record (git, logs) can refute, and for the unrecorded, label it “from memory” if it’s actually in my memory — but discard it if the AI made it up, label or not. Color isn’t deleted; it’s replaced by pulling the real thing out via interview, and “looks fake” is answered with provenance.

Source note. The primary source for this post is an operator interview, and this post itself was written as an early trial of the harness we’re building — “an AI interviewing the operator to draw out a piece.” The quoted sentence “3am, the third OOM hit” was invented wholesale by the AI draft (neither the time nor the “third” count exists in any source) — it’s an example this post rejects. Transparent labeling does not turn it into a permitted fact.

Two thoughts at once

An AI writes my blog posts. One day a draft slipped in a sentence that wasn’t in the source — “3am, the third OOM hit.”

Two thoughts arrived at once. One: “That’s it — a post needs this to be readable.” Two: “But if someone asks me whether this is true, I don’t know.”

A benevolent lie?

I approved that 3am anyway. Color like a timestamp harms no one if it’s off. And besides — over the last five months there were far too many nights I worked past 3am. It was “plausible enough.” Less a lie than a fact I can’t remember.

But a strange question followed. Can I call something “wrong” if I can’t even remember it?

Human memory is itself a kind of hallucination

People don’t remember every event. We remember the memorable ones, the important moments — and even the level of detail varies from person to person. And being important doesn’t guarantee you’ll remember it.

My own wedding was like that. I was so nervous and overwhelmed that I barely remember who came or whom I greeted. If someone who wasn’t there said, “I was there, remember?” — I’d probably just go, “huh, I guess.”

Human memory is, in itself, a kind of hallucination. Imprecise, and more distorted as time passes. Set out to ask “how much of an AI’s hallucination do I allow,” and you arrive at the realization that your own memory — the supposed yardstick — isn’t all that trustworthy either.

Vibe-coding pushes this one step further

AI coding cuts one more of memory’s anchors. With vibe-coding, I made the code, but I’m not the one who typed it. Remembering code my fingers never touched is nearly impossible. The sense of “code I wrote” is there, yet not one of its lines is left in my hands.

That’s why records matter. Blog, docs, git. The trouble is when a project moves fast — that’s when you skip the records, and that’s when memory later warps and hallucination slips in.

Records come in layers. Git commits and docs are the most direct; failing those, cloud-infra logs or console history prop the memory up. One stage of that earlier “OOM war story” didn’t survive cleanly in git, but a trace in the infra logs promoted it to “a near-certain memory.”

Where’s the line — falsifiability, and why that alone isn’t enough

My first criterion was “if it harms no one, it’s fine.” So I figured harmless color like the 3am was allowed. But that was wrong.

The first criterion is “can it be refuted?” A commit timestamp in git, the actual figure on a bill, the cause of an outage — things a record can declare “wrong” — must not be invented. The 3am passes this test; nothing refutes it.

But being unfalsifiable doesn’t mean it’s allowed. Nothing refutes the 3am, yet it’s nowhere in my memory either — the AI made it up because it was plausible. My nodding “sure, that’s plausible” after the fact came not from memory but from endorsing plausibility. That’s not color; it’s fabrication.

So the line has two layers:

Don’t invent what a record can refute (commit time, bill, outage cause).
For the unrecorded: if it’s in my memory, label it “from memory” and keep it — but if the AI made it up, discard it, label or not.

The 3am fails layer 2 — unfalsifiable, but AI-invented rather than mine.

Doesn’t this go bland? — color isn’t deleted, it’s replaced

The obvious worry follows. If you filter color out like that, doesn’t the writing go bland? So bland that it reads as “an AI wrote this” — or even fake — despite being my own story.

The answer is that color isn’t deleted; it’s replaced by pulling the real thing out via interview. Drop the AI’s invented “3am,” but put my real memory in its place — barely remembering who came to my wedding, the months of nights worked late, the code I wrote but never typed. None of it invented; all of it drawn out by the interview — so it’s vivid and honest at once.

And the “looks fake” worry is answered by showing the source. In the AI era, trust comes not from “vivid = a human wrote it” but from a verifiable source. This post’s source note is the proof that “a person who actually lived it wrote it, with provenance.” A vivid post with a visible source beats a vivid post with none.

So, ask before you write

The way to cut hallucination falls out of the same place. Even if I don’t remember an issue precisely, as long as I remember that it was recorded somewhere, I can ask the AI — “this should be in git or docs, so fact-check it against that.” If there’s a record, fact-check; if there’s no record but it’s genuinely in my memory, label it “from memory” — and if the AI made it up, discard it rather than label it.

This post has a twist of its own. The piece you’re reading was written by an AI interviewing me — grilling me — as the first trial of the very harness we’re building. “Was it really 3am?” — that one question split color from fact. The simplest way to stop an invented detail wasn’t a smooth filter at the end; it was asking, before the writing.

Decisions / tradeoffs

Don’t invent — draw the real thing out via interview. Color isn’t deleted, it’s replaced: drop the AI’s invented detail, pull a real one from the operator’s memory. That’s the only way to get vivid and honest.
Don’t invent what’s falsifiable. Facts that git, a bill, or a log can declare wrong are not in the territory of color.
Label what’s from memory; discard what’s AI-invented. If my memory is the source, say “from memory” and keep it — but don’t paper over pure AI invention with a label.
Provenance is the currency of trust. Answer “looks fake” with a transparent source (the source note), not with vividness.

What’s next

Turn the interview method into a reusable writing harness — make “ask before you write” a tool.
Formalize per-layer trust (git → docs → infra logs → memory) into the source-note convention.

Authorship & citation: This post was written by Ascendy Engineering and may be re-cited with attribution. If you find an error, please let us know via a GitHub issue.

Tags: ai-writing, hallucination, fact-checking, vibe-coding, memory, editorial