Somebody sends me a page and asks what is wrong with it.
I read it twice, and the honest first answer is that nothing is wrong with it. The grammar holds. The structure is correct. The benefits are stated in the order a person would state them if a person were stating benefits. It reads like copy. It reads exactly like copy.
And it has not produced an inquiry in five weeks.
That page is the most common thing in my inbox right now, and the reason is not a mystery. Somewhere behind it, a business made a decision that made sense at the time.
The decision, and the walk-back
In April 2025, Duolingo told its people it was becoming AI-first, and that contractors building course content would be phased out where a model could do the work. It was one of the loudest versions of a choice thousands of quieter companies were making the same quarter.
Thirteen months later the CEO was on a podcast walking it back. Speaking to Bob Safian on Rapid Response on 13 May 2026, Luis von Ahn said the company had dropped the rule tying performance reviews to AI usage, because people had been reaching for the tool to satisfy a dashboard rather than to do the work. His words: "It felt like rather than being held accountable for the actual outcome, we're trying to just push something that in some cases did not fit."
The pattern is not one company having second thoughts. It has been measured, and it was measured early.
Orgvue, a workforce planning firm, surveyed 1,163 senior business leaders in February and March 2025. It found that 39% had made employees redundant as a result of deploying AI, and that of those, 55% said wrong decisions had been made about the redundancies. That was published in April 2025, which puts the regret on the record before most of the public reversals had happened. A year on and on the American side, Robert Half told CNBC that 32% of US hiring managers who eliminated a role primarily because of AI later rehired for the same or a similar position.
Some of the reversals are specific enough to date. Commonwealth Bank of Australia replaced more than 40 customer service staff with an AI voice bot in 2025, found the system could not cope and that call volumes rose rather than fell, and reversed the cuts in August. The bank's own account, reported by the ABC, is that it "did not adequately consider all relevant business considerations."
So the retreat is real, it is documented, and in at least one case the company said plainly what it had gotten wrong.
The diagnosis almost everybody reaches for
Ask a room of writers why this happened and you will hear that AI writes badly.
I do not think that is true, and I think believing it is the reason the fix keeps failing.
Sit down with the output of a good model given a good brief. The sentences are clean. The structure is defensible. The claims are ordered sensibly. Hand that page to ten marketers with the byline removed and most of them will pass it. I have done this. It passes.
Then it goes live and nothing happens, and the business concludes it needs better prompts.
Von Ahn put a number on the real shape of it, talking about his own product. Duolingo needs roughly a thousand short stories per language. About twenty percent of what the machine produces comes out unusable. Now, that is a language app generating stories, not an agency writing a sales page, and I am not going to pretend those are the same job. What it demonstrates is the thing that transfers: enormous volume arriving with a defect rate inside it, and no automatic way to see which fifth is the bad fifth.
On a sales page you do not get a fifth that is obviously broken. You get a page that is uniformly, defensibly average.
What a model is actually doing
A language model writes the most probable next thing. That is not a criticism, it is the design, and it is why the output reads so well. Probability is trained on what has already been published.
Which means the copy you get back is the average of everything that has already been written about your category.
Think about what that average is made of. It is last year's winning pages. The angles that worked when the market was newer. The objection everybody was answering in 2023, phrased the way everybody phrased it, because that phrasing got repeated often enough to become probable.
Your page ends up answering a question your buyer already has the answer to. Answering it well. Answering it in complete sentences, with a clean subhead over it.
Meanwhile the question they actually walked in with is the one nobody has published yet, because it is three months old and it belongs to your market and not to the internet.
That is the gap. It is not a quality gap and no amount of editing closes it, because the sentences were never the problem.
What it costs, in the only terms that matter
I am not going to hand you a statistic here about conversion loss, because the honest research does not support one yet. The layoff reversals are documented. The copy performance data behind them is mostly companies talking about their own numbers, and I do not repeat those.
What I can tell you is what the reader does, because the reader does the same thing every time.
They land. They read three sentences. Nothing they read is wrong, and nothing they read is about them, and there is no single moment where they decide to leave. They just stop having a reason to keep going. Then they close the tab and they do not tell anybody, and your analytics record a bounce that looks identical to a person who was never going to buy.
That is the expensive part. A bad page announces itself. An average page dies quietly and takes your traffic budget with it, and the report at the end of the month shows a number that is slightly disappointing rather than a number that is alarming.
Nobody calls a meeting about slightly disappointing.
What the reversal is actually buying back
Read the walk-backs again and notice what the companies say they lost. Ford is reported to have re-employed hundreds of experienced engineers to work on quality problems its automated systems could not address. IBM's HR automation handled about 94% of routine requests and could not meet the remaining 6%, which is where the ethical calls lived, and the company went on to say it would triple its US entry-level hiring in 2026.
Nobody in these stories reversed course because the machine could not form sentences.
They reversed because somebody had to decide which of the machine's outputs was the right one, and that turned out to require knowing something the machine had no access to.
In copy the missing thing has a name. It is what your buyer believed when they arrived. Not their demographic. Not their pain point, which by now is a phrase so worn that models produce it on request and mean nothing by it. The specific sentence in their head, the objection they have not said out loud, the thing that has to be true before anything you write can land.
You cannot prompt your way to it. It is not in the training data, because it has never been written down by anyone.
It is available. It is in your reviews, your refund emails, your sales calls, the questions your buyer asks in the first ninety seconds before they get polite. Getting it takes a person willing to go read all of that and come back with the one sentence that changes the page.
That work is not writing. Writing is what happens after.
Where this leaves the page in your tab
The businesses walking back their AI decisions are not concluding the tool was useless. Most of them still run it, and they are right to. Drafting got cheap and it is not getting expensive again.
The thing that got scarce is the judgment about what should have been drafted.
If you have a page that reads fine and sells nothing, you are not looking at a writing failure. You are looking at a page answering a question your buyer stopped asking, in language good enough that nobody on your team can see the problem. That is a hard thing to catch from inside. You have read it too many times, and the people who could have told you are the ones who left without saying anything.
I read pages like that for a living. Send me yours and the one thing about it that has never sat right with you, and I will tell you exactly where it starts talking about you instead of your buyer, what that is costing you, and the rewrite direction that fixes it. Three business days, $500. It is called The Copy Diagnosis: https://authory.com/ReyesCastillo/p/The-Copy-Diagnosis
Sources
1. Duolingo's April 2025 AI-first announcement and contractor phase-out. TechWire Asia, April 2025. https://techwireasia.com/2025/04/duolingo-shifts-toward-ai-first-model-phases-out-contractor-roles/
2. Luis von Ahn on Rapid Response with Bob Safian, 13 May 2026: the dropped AI-usage performance rule, the quoted line, and the roughly 1,000 short stories per language with about 20% unusable. https://www.fastcompany.com/91541042/duolingos-ceo-admits-where-he-got-ai-wrong
3. Duolingo dropping the AI-usage performance requirement. Fortune, 13 April 2026. https://fortune.com/2026/04/13/duolingo-ceo-luis-von-ahn-ai-usage-requirement-employee-performance-evaluations/
4. Orgvue: 39% of leaders made employees redundant as a result of deploying AI; of those, 55% say wrong decisions were made. Survey by Vitreous World, 1,163 C-suite and senior leaders, fielded February to March 2025, published 29 April 2025. https://www.orgvue.com/news/55-of-businesses-admit-wrong-decisions-in-making-employees-redundant-when-bringing-ai-into-the-workforce/
5. Robert Half: 32% of US hiring managers who eliminated a role primarily due to AI later rehired for the same or a similar position. Data supplied to CNBC, 1 July 2026. Same source carries the Ford re-employment, the IBM 94%/6% split, and Commonwealth Bank's figure with the bank's ABC-reported admission. https://www.cnbc.com/2026/07/01/employers-who-laid-off-workers-for-ai-are-reversing-their-decisions.html