In the last year alone, you’ve probably seen headlines like:
“ChatGPT rots your brain!”
“95% of AI deployments fail!”
Provocative? Yes. Accurate? Not quite.
Both claims come from legitimate research efforts from teams associated with MIT. But what the public has absorbed from these studies isn’t evidence. It’s a hyped, out-of-context narrative. And that is often more damaging than the technology it’s trying to explain.
The issue isn’t the researchers themselves. Both studies include useful insights and thoughtful caveats. But in the AI hype cycle, nuance is often the first casualty. What remains is a kind of statistical gossip: numbers stripped of context, findings distorted into policy, and preprints recirculated as gospel.
Let’s look at two of the most misinterpreted studies of 2025 and what we can learn from the way they were misunderstood.
Study #1: Your Brain On GPT
This project made headlines for allegedly showing that ChatGPT damages your brain. Some coverage even invoked the ghost of 1980s anti-drug PSAs: “This is your brain. This is your brain on ChatGPT.”
What did the study actually do?
The team ran an experiment with 54 students over four months. Participants were split into three groups: one used GPT to write short essays, another used a traditional search engine, and a third wrote everything from scratch. The researchers used EEG to measure brain activity and memory recall, comparing levels of neural engagement during writing tasks.
The results weren’t shocking. The more help students received from external tools, the less active their brains appeared. The GPT group showed the lowest engagement, weakest memory retention, and least sense of authorship. In follow-up interviews, they also struggled to recall content they had just “written.”
In other words: passive use of automation leads to lower cognitive engagement.
But that’s hardly breaking news. It’s the same with copy-paste, formulaic templates, or letting someone else do your thinking for you. The real insight isn’t about the danger of GPT, it’s about the importance of how we use tools.
And yet, headlines that leapt to dramatic conclusions, in the vein of “LLMs dull your brains,” and, “AI is hurting student learning,” flooded the internet. The study was non-peer-reviewed, limited in scope (54 participants, one task, short time frame), and focused on one narrow educational setting. But these caveats were quickly forgotten. At best, the study raised interesting questions. It didn’t provide sweeping answers.
Study #2: “95% of AI Deployments Fail”
This figure has become a staple of conference keynotes, LinkedIn hot takes, and AI think-pieces. It originally comes from an MIT Sloan report that explores how companies are (and aren’t) succeeding with GenAI.
Here’s what few people seem to realize: the “95%” number isn’t a robust finding.
The report itself is thoughtful and useful as exploratory research. It includes interviews with representatives from 52 organizations, 153 surveys, and 300 self-reported case studies. But it lacks the kind of rigor required to support a definitive claim like “95% of deployments fail.”
The failure rate refers specifically to pilots that didn’t deliver ROI within six months. That’s a bizarrely short time frame for technologies with long adoption curves. It also skews the results against capital-intensive or strategic implementations that take years to pay off.
The methodology is also fragile. There’s no financial audit. No clear KPIs. No response rate. No reproducible dataset. The sample is a classic convenience sample. And many of the success/failure judgments rely on subjective coding rather than validated metrics (meaning, for the non-academics out there, that someone simply decided whether a project “succeeded” or “failed” based on their own judgment, not on hard numbers or agreed-upon criteria).
Most curiously, the report is also used to promote MIT’s own initiative—NANDA—which purports to help solve the very problems the report diagnoses. That doesn’t invalidate the findings. But it does raise questions about motivation and messaging.
Still, the number is out there. It’s catchy. It spreads. And now, “95% of AI deployments fail” is being treated as a known fact—despite being more slogan than science.
What We Can Learn
These aren’t isolated incidents. The same pattern repeats again and again: a study makes a cautious claim, the media strips out the nuance, and the resulting narrative becomes the new conventional wisdom.
This isn’t just an academic issue. Misinterpreted research leads to:
- Panic in classrooms
- Poorly designed corporate strategies
- Misguided regulation
- Bad public discourse
It would seem then that up until now, AI’s biggest risk isn’t sentient robots or job loss. It’s bad and over-hyped thinking about how the technology actually works.
So what’s the solution? We don’t need less research; we need better interpretation. That means:
- Teaching literacy in how to read studies
- Paying attention to context and methodology
- Holding media accountable for accuracy
- Rewarding researchers who resist overstatement
In short: it’s time to stop treating every AI headline as gospel (or garbage) and start building a more thoughtful relationship with the science behind the hype.
Because if we don’t, the real cognitive decline won’t be caused by AI. It’ll be caused by us.