DREAM / The Prime Minister resigns and wakes up under the Parisian bridge!

 Kliko këtu për të tjera
Voxpress:

MIT study: 95% of Artificial Intelligence enterprises fail

2025-08-26 17:59:20Biznes SHKRUAR NGA REDAKSIA VOX
MIT study: 95% of Artificial Intelligence enterprises fail
Illustrative photo

A new report from MIT (Massachusetts Institute of Technology) has shocked the artificial intelligence world by revealing that 95% of generative AI initiatives offer zero return on investment.

The findings, based on 300 public implementations, more than 150 executive interviews, and $30 to $40 billion invested in these pilot projects, suggest that billions of dollars have been spent on AI experiments that never scale, and that most organizations are locked into what MIT researchers call the "Generative AI Divide."

The figures revealed are grim, 40% of organizations said they have implemented AI tools, but only 5% have managed to integrate them into large-scale workflows. Most projects die in pilot purgatory. Meanwhile, the news warns of an "AI bubble", and some investors are lowering AI shares with the idea that the big moment of generative AI enterprise is already being delayed.

However, not everyone agrees with this interpretation.

"The problem is complete error," said Tanmai Gopal, co-founder and chief executive of PromptQL, an AI company that counts OpenAI, Airbus, Siemens, and NASA as its customers. "If the system is not always accurate, even in the smallest percentage of the time, you need to know when it is not. Otherwise, my minutes turn into hours; the return on investment disappears."

In his blog post, Gopal outlined what he calls the "verification tax."

"I don't know when I can get an incorrect answer from my AI. So I have to forensically check every response," he wrote.

This explains much of what MIT labeled as the GenAI split. Companies eagerly launch pilot projects, but employees end up spending so much time double-checking the results that the promised efficiency never materialize.

It's not that generative AI lacks raw power, models can be impressive. It is that their faith is uncalibrated. In regulated or high-risk industries, a bad response can be greater than 10 good. As Gopal puts it: "For a serious job, a high-confidence mistake costs more credibility than 10 successes earn."

MIT researchers formulated the same issue differently. They found that most enterprise AI tools don't store feedback, adapt to workflows, or improve over time. Without these qualities, they stagnate.

"Without high-quality, unreliable information, I don't know if a result is wrong due to ambiguity, lack of context, old data, or a pattern error. If I don't know why he's wrong," he said, "I'm not invested in his success," Gopal says.

This information matters because it reframes the entire conversation. If AI isn't failing because of a lack of skills, but because it's not designed to communicate its limitations and learn from corrections, then the solution is less about building bigger models, and more about building simpler models.


Në përputhje me standarded JTI

Video

SHIKO KANALIN NE YOUTUBE
×