Régis closes his laptop past midnight. The eight-hundred-word article on industry trends just came out of ChatGPT after four prompts and two quick read-throughs. He reads his headline one last time, smiles, and schedules the post for seven in the morning. He is proud. He feels he has written something good, maybe his best piece of the month.
Three weeks later, the article sits at forty-two views, twelve of them from his own colleagues. The client asks why traffic hasn’t moved. Régis can’t see the problem. He did everything right, or at least that’s how it feels.
Régis is neither lazy nor incompetent. He is caught in a trap that psychology has documented for decades, one that generative AI has just made far more dangerous for businesses. This trap has a name, or rather three: the IKEA effect, NIH syndrome, and the Dunning-Kruger effect. Taken separately, each of these biases is already well documented by research. Combined around a ChatGPT prompt, they form a mechanism that pushes thousands of professionals to publish mediocre content while being convinced of the opposite.
The IKEA effect: why you love what you just assembled
In 2011, researchers Michael Norton, Daniel Mochon, and Dan Ariely asked participants to assemble IKEA boxes, fold origami, and build Lego sets (Harvard Business School). The result: participants valued their own creations nearly as much as an expert’s, even when their work was visibly clumsy. The mere act of putting in effort was enough to create attachment. The researchers named this phenomenon the IKEA effect.
The effect disappeared under one specific condition: when the task was left unfinished, or when the creation was destroyed before being evaluated. Labor alone does not create the illusion of value, it needs a sense of completion.
A ChatGPT prompt checks that exact box. You type an instruction, you get a finished text in fifteen seconds, and you feel like you accomplished something. The sense of competence documented by Mochon, Norton, and Ariely in their follow-up work activates the same way it does with a piece of flat-pack furniture: you worked, so you like the result, regardless of its actual quality. This is one of the mechanisms I detail in my book about the illusion of mastery created by any tool that delivers a fast result (My book, chapter 6).
The problem isn’t the tool. It’s the speed at which it turns a simple prompt into a feeling of intellectual ownership. An employee who used to spend three days drafting a white paper knew, from the sheer duration of the work, that they were allowed to be tired and obligated to get a second read. An employee who gets the same white paper in twelve minutes no longer has that natural safeguard. The tool’s speed removes the one signal that used to prompt someone to ask for a second opinion.
NIH syndrome: your best feedback came from outside
NIH syndrome, short for Not Invented Here, was formalized as early as 1982 by researchers Ralph Katz and Thomas Allen, who studied fifty research and development teams at MIT. Their finding: teams grow increasingly resistant to outside information as their tenure lengthens. A more recent study cited by MIT Sloan Management Review tracked five hundred sixty-five innovation projects worldwide. Only sixteen percent of them remained entirely unaffected by this bias, and its presence correlated with a reduced likelihood of project success (MIT Sloan Management Review).
NIH syndrome isn’t limited to rejecting a competitor’s or a consultant’s ideas. It also shows up in how a professional relates to their own AI-generated content. Once Régis produced his article alone, at midnight, with his own prompt, outside feedback becomes an obstacle rather than a help. Input from a professional writer, a client, or even a colleague trained in SEO runs into a quiet resistance. The content already feels like his, it carries his mark, and correcting it deeply would mean admitting that working alone wasn’t enough.
The irony is that generative AI was supposed to democratize access to proven external practices, tested models, expertise once reserved for a few specialists. It sometimes ends up reinforcing the opposite: the idea that producing content alone, fast, without outside eyes, is sufficient. In the organizations where I work, this reflex shows up in one specific sentence, spoken by an employee who just received criticism on an AI-generated text: “but I already spent time on it.” That sentence, harmless as it sounds, signals that the conversation has left the territory of content and entered the territory of personal identity.
Dunning-Kruger: why you are the worst judge of your own prompt
In 1999, psychologists Justin Kruger and David Dunning, then researchers at Cornell University, published a study that has since become a reference point in social psychology (Journal of Personality and Social Psychology). Their protocol: submit participants to tests of logic, grammar, and humor, then ask them to estimate their own score. The result: the weakest performers grossly overestimated their performance, while the strongest tended to slightly underestimate theirs. The authors’ conclusion fits in one sentence: the skills needed to do a task well are often the same skills needed to judge whether you did it well. Without them, you remain unable to measure your own failure.
Applied to AI-generated content, this mechanism takes on a particular weight. An executive, a communications manager, or a consultant who has no grounding in SEO, editorial structure, or the conventions of their industry simply lacks the mental tools to judge whether a ChatGPT text is good. They judge it on how smoothly it reads, whether the expected keywords appear, and the absence of typos. Those are exactly the criteria a language model excels at, and exactly the criteria that are insufficient to judge a piece of content’s strategic relevance.
The paradox is unforgiving: the less editorial expertise a contributor has, the more satisfied they tend to feel with their AI output, and the fewer reasons they see to ask for an outside opinion. The first two biases end up reinforcing the third instead of correcting it.
The convergence: a triangle flooding the web with content its author alone adores
Taken individually, each of these three biases explains an old, well-documented human behavior. Combined around AI-generated content production, they create a self-feeding loop. You produce a text in a few minutes, the IKEA effect. You immediately become its owner and close yourself off to outside feedback, NIH syndrome. You lack the skills to evaluate its quality objectively, the Dunning-Kruger effect.
The numbers confirm how widespread this is. A 2026 survey by Ahrefs of eight hundred seventy-nine marketing professionals found that only fourteen percent of them consider AI-generated content superior in quality to human-written content, even as most keep producing it at scale (Ahrefs). A study led by researchers at MIT Sloan revealed something telling. When evaluators don’t know a text’s origin, they show no preference for human content, and sometimes even prefer AI-generated content. But the moment they learn a human was involved, their rating rises (MIT Sloan). Perceived quality, in other words, depends less on the text itself than on the story people tell themselves about who produced it and how.
That disillusionment already carries a real cost. According to a report published by Forrester in late 2025, a global consulting firm recently had to refund several hundred thousand dollars to a client after delivering a document riddled with AI hallucinations (Forrester). The same report predicts that by the end of 2026, centralized content teams will no longer produce two-thirds of content inside B2B organizations, replaced by isolated contributors, each convinced their own prompt is enough.
Breaking the triangle: measure instead of feel
The good news is that this triangle can be dismantled. Not with willpower or good intentions, but with precise mechanisms, the same ones I use in the Innovational Intelligence™ System for any technology adoption project in a company (My book, chapter 4).
Three levers work.
First, separate production from evaluation. The person who wrote the prompt should never be the only one judging the result. An outside, trained eye, with no emotional stake in the text produced, mechanically breaks the IKEA effect. That eye doesn’t need to be hostile, it simply needs to owe nothing to the effort already invested.
Second, measure with indicators external to how you feel about the work. Actual traffic, conversion rate, time on page, client feedback. These numbers don’t know how much effort you put into your prompt, and that’s exactly what makes them useful. An article that took you fifteen minutes and one that took you three hours are worth exactly the same to a reader: zero, if the content doesn’t meet their need.
Third, organize regular confrontation with outside sources: agencies, peers, competitor data. NIH syndrome recedes the moment a team accepts comparing its work to others’ rather than shielding it. It’s also a matter of psychological safety: a team afraid of outside judgment would rather produce fast and alone than expose itself to criticism.
Régis eventually understood his mistake, not by working harder, but by agreeing to have a senior writer review his next articles before publication. The first round of feedback was rough. The second article tripled his traffic. Nothing changed in the tool he used. Everything changed in the room he left it.
The essentials
Remember three things:
- The effort you put into a prompt creates an illusion of value, regardless of the actual quality of the text produced.
- Content you generate alone quickly becomes territory you defend rather than a draft you improve.
- Without editorial or strategic expertise, you lack the tools to judge your own AI output objectively.
The answer isn’t to give up generative AI. It’s to refuse to be the author, the sole reader, and the sole judge of what it produces for you, all at once.
References
- (My book) https://philippeboulanger.com/book/
- (Harvard Business School) https://www.hbs.edu/faculty/Pages/item.aspx?num=41121
- (MIT Sloan Management Review) https://sloanreview.mit.edu/article/beating-not-invented-here-syndrome
- (Journal of Personality and Social Psychology) https://doi.org/10.1037/0022-3514.77.6.1121
- (MIT Sloan) https://mitsloan.mit.edu/ideas-made-to-matter/study-gauges-how-people-perceive-ai-created-content
- (Ahrefs) https://ahrefs.com/blog/?p=188567
- (Forrester) https://www.forrester.com/blogs/predictions-2026-trust-gets-tested-for-b2b-marketing-sales-and-product-leaders
