Microsoft, University of California at Berkeley, San Francisco and Columbia University are working on neuroscientific studies to predict the brain’s response to language. These studies provide high accuracy and most people will have seen Microsoft’s Copilot at work.
This is not just the old kind of prediction using conventional language. Studies focus on “The brain models prediction in short verbal explanations of what each part of the cortex reacts to: phrases like “food preparation” or “names of locations.”
In the process, LLMs have become fundamental tools for predicting human responses. Brain regions are actively engaged with the results of tests that provide language that illuminates certain brain regions.
Testing and evaluation
The idea is to predict the answers. As an example, Microsoft uses a simple “write a story” analogy about creating a food preparation item and uses specific terminology to elicit responses. These responses are checked against the prediction.
This is called Generative Causal Testing, as Microsoft explains:
GCT has two steps: explanation, then verification. To generate an explanation, the method starts from a predictive model for a single voxel or region and identifies the short phrases that most strongly drive its predicted response. An LLM then condenses those words into a concise verbal explanation, often a single phrase such as “food preparation” or “place names.”
Using “graphic voxels”, the response is evaluated and compared to the predicted results, and they are very predictable, just using basic language structures. Note that “food preparation” or “names of places” are also obvious natural triggers for any answer, the classic “what and where” of any text. According to Microsoft, this method provides results based on
Predictive AI and language
To their credit, Microsoft adds a strong and meaningful caveat to its findings so far:
The importance of GCT reaches far beyond neuroscience. Researchers are increasingly faced with the same dilemma: a model that predicts beautifully but explains nothing.
The main takeaway so far is that Microsoft thinks it has found a mapping tool, not an infallible answer key. Given the hostile responses to so much AI-generated content, this seems like an accurate description.
Predictive AI itself it is also very much a work in progress. It’s clumsy, mechanical, data-gathering, and everywhere. It’s especially prevalent in marketing, where it generates those cheesy, often repetitive, stagnant, endlessly predictable feeds on your social media.
Broader linguistic ramifications and very useful for neuroscience
Given the overwhelmingly hostile response to the “AI slant,” another very appropriate use for this research might be to describe disengagement and disengagement with AI-generated content.
At what point do users disconnect?
By what criteria does the brain reject content?
Can the neurological reaction be determined?
Can the rejection process identify flaws in content generation?
Does the distillation of AI from AI-generated content contribute to rejection and to what extent?
This type of problem is becoming a fatal flaw in AI-generated content at almost all levels of media. There’s not much point in AI-generated content if no one is going to use it. “Long time no read” can easily become “Too dumb no read”. This is one step away from refusing to read.
The issue of secession is becoming critical. Let’s also not shy away from the fact that expert readers are the first to criticize inappropriate content. These criticisms are often damning and perfectly valid, regardless of whether the problem is pronunciation or fact-checking.
Predictive neuroscience AI and potential uses outside of language
This research deserves consideration in a broader context. Microsoft is not a neuroscience business. However, he may have found a useful psychological tool for assessing responses to “trigger words.” It could be a word association test with an LLM attached.
Does a seemingly innocuous word cause a disproportionate response?
Is a word that generates a strong response associated with trauma?
What about specific topics?
How would you measure, and can you, the response in a voxel system?
Has any research of this type been done, and if so, with or without Microsoft’s search tools?
This type of reaction cannot be a lie detector in any form, but as an “Oh!” meter, can be very useful. Given people’s natural reticence about intensely personal topics, it can be very helpful.
There is a gaping hole in AI in its uses in practical psychology. AI in psychology is currently at the application level as a 24/7 support and in similar roles. It is not necessarily some kind of diagnostic tool.
The neural responses are very different. They can be used for proper documentation of issues in case management. They can also chart important changes in responses during therapy, such as case progress and whether therapy is working or not.
Neuroscience and language are an endless mix
Neuroscience language research is beginning to open up a space no one knew was there. Of course, people respond to language. How and why remain very much open to debate and discussion.
Can you try a new word for engagement?
Can you identify the active terminology that evolves from its engagement or disengagement?
Can you gauge how people engage with literary classics and what makes them stand out?
Can you use machine learning to correct mispronunciations?
Is there anything you can do to make word of mouth less exhausting for writers and readers?
Can you make your marketing language more informative and efficient, such as cutting out endless subject prologues?
This may very well be the answer to stripping AI, but it may also be a way out of all that tiresome, chronically inefficient language.
