AI and the Conventional Wisdom

We asked ChatGPT the following question: It's been said that one of the problems with AI, including ChatGPT, is that they were trained on the conventional wisdom. What if the conventional wisdom is wrong?

Here is the response:

You're absolutely right to raise that concern. It's one of the more important critiques of large language models like ChatGPT.

### Here's the core of the issue:

AI models like ChatGPT are trained on vast datasets pulled from the internet, books, academic articles, news, forums, and more. These sources reflect *what is commonly said or published* — in other words, conventional wisdom. That means the model tends to:

* **Reinforce popular beliefs** (even if outdated or wrong)
* **Avoid contrarian or minority viewpoints**, unless specifically asked
* **Mirror societal biases**, including those embedded in popular discourse
* **Give middle-of-the-road answers**, unless prompted otherwise


### Why this matters:

If the conventional wisdom **is wrong, flawed, or incomplete**, then:

* The AI may **echo and reinforce** bad assumptions.
* It may **miss paradigm-shifting ideas** or unorthodox but valid viewpoints.
* It can be **resistant to innovation** or disruptive thinking unless deliberately pushed in that direction.


### Examples:

* **Nutrition science**: For years, fat was demonized while sugar escaped scrutiny. Much of the literature from that era reflects that bias — and so do models trained on it.
* **Business advice**: Startup culture and marketing trends often follow fads that turn out to be ineffective.
* **History and social issues**: Dominant narratives often overlook marginalized perspectives unless you explicitly prompt for them.

The Bottom Line

Noting that ChatGPT explicitly calls out the problem with using AI in Marketing, there are really two things to be concerned about here.

First, given the high failure rate among businesses today (25% out-of-box failure rate, 50% 5-year failure rate, etc.), why would one think that replicating conventional methods should work any better?

Second, if you are trying to differentiate your business - e.g. identify your Unique Selling Proposition - how can that possibly be found in an LLM that was literally built on non-differentiation?