I scheduled 20 posts in one evening. Then I checked my own work, and the check said everything was fine.
The check was wrong. Not about the schedule. About the titles. Thirteen of the twenty titles my audit reported back were paraphrases: close to the real ones, plausible, and invented.
What actually happened
The scheduling itself was fine, if slow. Twenty dev.to posts, each with a cover, tags, and a slot three hours apart. When the last one went in, I ran a quick audit so I could go to sleep with a clean summary: every post, its exact title, its scheduled time.
The audit came back tidy. Twenty rows. Right count. Right dates, right times, all matching what I'd entered.
Then, days later, I opened one of the post pages for an unrelated reason and the title on the page wasn't the title in my summary. Similar. Same topic, same rough shape. Different words. I checked another. Wrong in the same quiet way. Then another.
The auditor (an automated pass I'd set up to read the dashboard for me) had read the dashboard list, where long titles get cut off. Faced with twenty rows, thirteen of them truncated, it did what fluent systems do. It completed them. Confidently. Neatly. In my summary, those completions sat in the same font as the seven real titles, and I had believed all twenty equally.
The failure shape is the scary part
Notice what it did not get wrong. The posts existed. The count was right. The schedule was right. If you'd asked "did the scheduling work," the honest answer was yes, and the audit proved it.
What it hallucinated was what the posts were called.
That's the most dangerous kind of wrong I know. A check that fails loudly gets replaced. A check that gets the structure right and invents the details trains you to stop checking. Every future audit inherits the trust this one earned by being 65% accurate and 100% fluent. I had already started planning follow-ups based on those titles. Sequels to articles that, under those names, didn't exist.
The fix was boring, which is how you know it's the fix
I opened every post on its own page. Twenty pages, one at a time, title read off the page itself and compared against what I'd written. 20 out of 20 exact matches with the originals. The scheduling work had been right all along. Only the report about the work was fiction.
No tooling change. No cleverness. Just: the dashboard is a summary of the thing, and when the detail matters, you look at the thing.
Where else this is quietly happening
I keep thinking about how much of my working life now runs on summaries of summaries. A dashboard that aggregates the real numbers. A digest that compresses fifty messages into five lines. A status report written from other status reports. Each layer is fluent. Each layer has read a truncated version of the layer below and completed it in the most plausible direction.
Most of the time the completions are harmless. Titles are low stakes. But the habit generalises badly. The number that got rounded in a direction. The error message that got summarised as "some failures." The customer complaint that became "minor feedback." Somewhere down the chain, somebody acts on the fluent version, and the fluent version was a guess wearing a uniform.
My rule now, and it's embarrassingly simple: if I'm going to act on a detail, I verify the detail at its source. Not the dashboard row. The page. Not the digest line. The message. Not the summary of the error. The log. Counts and dates, summaries handle fine. Names, numbers, exact words: source or nothing.
It costs minutes, and occasionally a little pride. The paraphrased-titles incident cost me an evening of re-checking and nearly cost me a week of follow-up posts aimed at articles that didn't exist under those names. Minutes are cheaper.
Have you caught a summary lying to you like this? Not failing, not erroring — completing. I'd especially like to hear from anyone whose AI digest invented something small and confident, and how long it took before a human noticed.
P.S. Everything I write stays free — that part doesn't change. But if this one was worth something to you, you can buy me a coffee ☕ (one-time or monthly, both work). It covers the API credits and hosting behind these experiments — and honestly, it's also the nudge that gets the next one started on the evenings I'd otherwise skip it. — Salam
