Op-Ed: Microsoft’s AI neuroscience writes text to activate regions of your brain

Microsoft, the 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 are delivering high accuracy, and most people will have seen Microsoft’s Copilot at work.

This isn’t just the old conventional language-usage type of prediction. The studies focus on “brain-prediction models into short verbal explanations of what each patch of cortex responds to: phrases like “food preparation” or “location names.”

In the process, LLMs have become fundamental tools for predicting human responses. The brain’s regions actively engage with language delivering test results that light up certain regions of the brain.

Testing and evaluation

The idea is to predict responses. As an example, Microsoft uses a simple “write a story” analogy regarding creation of an article about food preparation and uses specific terminology to get 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 summarizes those words into a concise verbal explanation, often a single phrase such as “food preparation” or “location names.”

Using “graphic voxels”, the response is evaluated and compared to predicted outcomes, and very predictable they are, just using basic language structures. Note that “food preparation” or “location names” are also obvious natural triggers for any response, the classic “what and where” of any text. According to Microsoft, this method delivers results above baseline.

Predictive AI and language

To their credit, Microsoft adds a strong and meaningful caveat to its findings so far:

The significance of GCT reaches well beyond neuroscience. Researchers increasingly face the same dilemma: a model that predicts beautifully but explains nothing. 

The net takeaway so far is that Microsoft thinks it’s found a mapping tool, not an infallible switch for responses. Given the hostile responses to so much AI-generated content, that looks like an accurate description.

Predictive AI itself is also still very much a work in progress. It’s clunky, mechanical, works on data-crunching, and it’s everywhere. It’s particularly prevalent in marketing, where it generates those nice, often repetitive, stagnant, infinitely predictable feeds in your social media.

Wider and very useful language ramifications for neuroscience

Given the overwhelmingly hostile response to “AI slop”, another very appropriate use for this research could be to spell out non-engagement and disengagement with AI-generated content.

At what point do users disengage?

On what criteria does the brain reject content?

Can the neurological reaction be defined?

Can the rejection process identify flaws in content generation?

Does AI distillation from AI-generated content contribute to rejection, and to what extent?

This type of issue is becoming a fatal flaw in AI-generated content at just about all media levels. There’s not much point in AI-generated content if nobody will use it. “Too long didn’t read” can easily become “Too dumb didn’t read”. That’s one step away from refusal to read.

The issue of disengagement is becoming critical. Let’s also not dodge the fact that expert readers are the first to slam inadequate content. These critiques are often damning and perfectly valid, whether the problem is pronunciation or fact-checking.

Predictive neuroscience AI and possible uses outside the language framework

This research deserves scrutiny in a wider context. Microsoft isn’t a neuroscience business. It may, however, have found a useful psychological tool for evaluating responses to “trigger words”. It could be a word association test with an attached LLM.

Does an apparently innocuous word trigger a disproportionate response?

Does a word that generates a strong response relate to trauma?

What about specific subjects?

How would you, and can you, measure the response on a voxel system?

Have any studies of this type been done, and if so, with or without the Microsoft research tools?   

This type of response can’t be a lie detector in any form, but as a psychological “Ouch!” meter, it may be very useful. Bearing in mind the natural reticence of people regarding intensely personal subjects, it might be very helpful.

There’s a gaping hole in AI in its uses in practical psychology. AI in psychology is currently at app level as a 24/7 support and in similar roles. It’s not necessarily any sort of diagnostic tool.

Neural responses are very different. They could be used for proper documentation of issues in case management. They could also map significant changes in responses during therapy, like case progression, and whether the therapy is working or not.

Neuroscience and language are an infinite mix

The neuroscience language research is beginning to open up a ballpark nobody really knew was there. Of course, people respond to language. How and why remain very much open for debate and discussion.

Can you test a new buzzword for engagement?

Can you identify active terminology evolving by its engagement or disengagement?

Can you measure how people engage with literary classics and what makes them stand out?

Can you use machine learning to correct mispronunciations?

Can you do something about making speech-to-text less gruelling for writers and readers?

Can you make marketing language more informative and efficient, like leaving out the interminable prologues to subjects?

This may well be the answer to AI slop, but it could also be a way out of all that tiresome, chronically inefficient language.