We sell validation, and we found a sentence in our own writing that only borrowed an institution's name

I keep "never use unverified figures" as a content principle. At the end of every post I add a footnote titled Figures not used in this post, disclosing what I left out and why.

I applied that principle to everything I have written. It took a full day, and the results were not good.

1. A sentence with an institution's name but no study behind it

The first thing flagged was this sentence.

"According to a 2026 BCG/HBR study, the more AI tools founders use…" — No such study exists. This is the incorrect sentence copied verbatim from my own writing.

It appeared in two of my posts. It sounds plausible. It names two institutions and even carries a year.

But the Harvard–BCG study I had actually verified is a different one. It is a preregistered experiment with 758 consultants, and it found that quality rose 40% on tasks inside the areas where AI performs well. It was published in 2025.

The content is different, and so is the year. The "problems that come from using AI heavily" that the sentence describes are not what that study examined.

In other words, it was not a citation but a borrowed institution name. Had I seen it in someone else's writing, I would have pointed it out immediately.

2. Figures circulating without sources

These turned up at the same time.

• A market growth rate of "CAGR 175%"
• "90% reduction in MVP launch costs"
• A specific percentage for the number-one cause of failure

None of the three had a source. They are widely quoted numbers, so they seem to have drifted in from somewhere and settled into my sentences.

One thing was especially uncomfortable. The opening section of the post containing those figures said: "Unverified figures are withheld or removed." Within the same post, the principle and the body contradicted each other.

3. The checking tool was wrong first

Then something more awkward happened.

I use a gate card to check content. It has five items—facts, tone, legal, quality, and measurement—and the quality item included a rule of "2 hashtags or fewer."

When I ran an Instagram draft through that rule, it came back as a violation. It had 5 hashtags.

But when I reread my own operating guidelines, the hashtag count was set differently for each channel. Fewer for LinkedIn, more room for Instagram.

"2 or fewer" was the LinkedIn rule. It should not be applied to Instagram. If I had cut them down as instructed, I would have fallen below the minimum I had set myself.

Reviewing the rest the same way, of the 8 issues flagged that day, 3 were false positives. Had I trusted the results and fixed everything, I would have nearly broken things that were fine.

4. What the manual check missed

So instead of reviewing by eye, I actually ran the validation tool. The results were different.

The post was about e-commerce abandonment rates. Checking by hand, I had judged the figure in the title to be a number without a source. I was wrong. The body had a proper link to the research firm's original report.

The real problem was elsewhere.

The body stated the rate with the qualifier "about" and under the metric name "cart abandonment rate." The title, however, had dropped the qualifier and changed the metric to **"abandonment right before payment"**.

Those are different metrics. The share of people who put items in a cart and don't buy is very different from the share who reach the payment step and then leave. The latter is much lower.

It was a case where the source was fine but the title distorted the body. A keyword search cannot catch this. If you only check whether a number is registered, it passes.

(I have not reproduced the actual value of that rate here. What matters in this post is not the number but the fact that the qualifier disappeared, and reprinting an unverified external figure here would repeat the same mistake.)

5. What I did about it

Over the day, I sorted things out as follows.

• Figures whose sources could not be verified were removed from the body. I kept the argument and removed only the numbers
• Sentences that merely borrowed an institution's name were deleted without a substitute citation. Hastily attaching some other evidence would repeat the same mistake
• 16 auto-generated posts were archived instead of published. Unsupported percentages had made it into their titles, and similar topics repeated six or seven times
• The hashtag rule on the gate card was split by channel. Without fixing the cause, the same false positives would recur

One item I could not save. It was three figures from a well-known growth case; they appear only in secondary sources, and I could not find the primary material. I kept the sentence and removed only the numbers. I plan to put them back once I verify the original.

6. Why I am making this public

Because what I sell is validation.

If the one selling validation has not validated its own writing, the product has an explanation but no evidence. And this kind of thing is usually discovered by outsiders first.

There is one more thing. Many of the problems found today were traces of AI-written drafts passed through without enough human review. AI is very good at producing plausible sentences with institution names attached. Readers let them slide just as easily.

So my conclusion is not "use less AI." It is that a sentence produced by AI is something to be checked, not a finished output. A gate is needed in between, and that gate also needs regular inspection. Today one of those gates was wrong, so I fixed it.

If you do just one thing today

Open one piece of yours that is currently published, and pick just one sentence that contains a figure. Then actually open the source for that number.

If there is no link, or you open it and it says something else — you have found the same thing I found today.

---

Footnotes

① Sources consulted (checked 2026-07-26)

• Dell'Acqua, F. et al. — Navigating the Jagged Technological Frontier. Harvard Business School Working Paper 24-013 · Organization Science (2025). Preregistered experiment with 758 consultants; task quality +40% inside the frontier. https://www.hbs.edu/faculty/Pages/item.aspx?num=64700
• The audit results and actions described in the body come from our own operating logs. The scope was 5 of my blog drafts, 10 drafts awaiting publication, and 16 auto-generated posts.

② Figures not used in this post

• The original values of the figures removed in the audit — Reproducing them in the body would spread unverified numbers again, so I left them out. Only the item names remain, such as "market growth rate" and "cost reduction rate."
• 3 referral program figures from the growth case — Removed because they were confirmed only in secondary sources. This post, too, says only "I could not find the primary material," without the numbers.
• The actual value of the abandonment rate discussed in section 4 — An external figure cited in a draft under audit. That draft did have a source, but it is only an example, not evidence supporting this post's argument, so I did not reproduce the value. This is to prevent it from being reprinted unverified.
• Percentages in the archived auto-generated posts — As unsupported figures, they only served as the reason for archiving; there was no reason to reproduce them here, so I mention only the items.

③ How this was made

The draft was produced with AI, then validated and edited by a person. The only external figure used to support this post's claims is the 1 source verified against the original in ①; the rest of the account comes from our own operating logs. The abandonment rate in section 4 is mentioned only as an audit example, and its value is not reproduced (see footnote ②). No generative images or audio were used.

---

CTA

If you have never gone back to check the figures in your published writing, a 2-minute diagnostic can show you where to look first. Please feel free to answer only as much as you are comfortable with. https://www.star-t.io/en/ai-readiness?entry=content_buildlog_citation&utm_source=blog&utm_medium=organic&utm_campaign=buildlog_2026&utm_content=blog-p1-borrowed-citation

---