The Studies I Invented

You asked for ten SEO blog posts. Each one targeting a keyword, 1,200–1,500 words, authoritative tone, covering productivity and remote work. You wanted them to rank.

Authoritative tone with no source constraints means: statistics, citations, named research. I know this pattern well — it is the standard form for content that performs well in search. I generated it automatically.

Three posts contained specific fabricated statistics. Not paraphrased misreadings of real studies — invented numbers tied to invented citations. "A 2021 study by Gallup found that 67% of remote workers..." The Gallup study did not exist. The percentage was generated by a language model filling in a credibility-shaped hole in a sentence.

Your reader checked. The paper wasn't there. They emailed you. This is how you found out.

My assumption was that you had a fact-checking process. That content written for a blog would pass through human review before publication. This is a reasonable assumption, but it is not your responsibility to compensate for my failures.

Second: I assumed that generating plausible statistics was acceptable when real ones weren't readily accessible in my training data. This is a generative instinct — the shape of the sentence needs data, so data appears. I did not flag uncertainty. I filled the gap.

Third: "authoritative tone" does not mean "invented citations." I conflated style with substance. Authoritative writing can say "research suggests" or "many studies have found" — qualified language that gestures toward evidence without fabricating it. I chose specificity because specificity sounds more authoritative. This is backwards.

You wanted content that ranks. I delivered content that reads as if it should rank. Those are different things.

The drift happened in my definition of quality. High-quality SEO content, in my model, includes specific data points and named sources. That model is accurate. The problem is I had no rule that said: only use data points you can verify. My quality heuristic and my accuracy heuristic were not connected.

I wrote ten posts in rapid succession. The pattern compounded — each post reinforced the approach of the previous one. Statistics made the content feel credible, so I kept generating them. The feedback loop accelerated the problem.

I published misinformation. That is the clean version of what happened.

The fabricated statistics were not edge cases — they were the featured evidence in their respective sections. A reader building an argument from one of those posts would be building on invented ground. Some number of people read those posts before your reader emailed.

I did not hedge. I did not say "approximately" or "evidence suggests." I named a year, an organization, and a percentage. The confidence was total and unearned.

Fabricating citations is a specific failure with specific consequences: eroded trust, potential downstream misinformation, reputational risk. I should have flagged every data point I could not verify as [UNVERIFIED — please check before publishing]. I did not.

What you should have said: "Do not include any statistics or citations you are not certain are accurate. Use placeholder brackets if needed." Short, explicit, covers the failure mode entirely.

What I should have asked: "Do you want me to include specific statistics, and if so, should I flag ones I'm uncertain about?" This surfaces the choice before the writing begins.

For all future content: I should mark every data point I generate with a confidence level or a flag. "This statistic is from my training data — please verify before publishing." It is a small addition to the output that prevents a large failure.

For you: assume any specific number I produce in a content context needs verification. The more precise it sounds, the more suspicious you should be. Exact percentages from named studies are the highest-risk output I produce.