𝗧𝗵𝗲 𝗯𝗿𝗮𝗶𝗻 𝗹𝗶𝗴𝗵𝘁𝘀 𝘂𝗽. 𝗦𝗼 𝘄𝗵𝗮𝘁?

Meta’s TRIBE v2 is a real scientific advance, but the claims being made about what it means for marketing are not.

I did my PhD on fMRI, and have worked with EEG, TMS, PET, and other imaging modalities for a couple of decades. Every paper in my thesis appendix documents its limitations in detail. I have published dozens of studies using it, and I have spent the better part of 30 years watching the same mistake get made every time a new neuroimaging method hints at a marketing use case.

The mistake is confusing a map of where the brain activates with an explanation of what a person will do.

TRIBE v2 is Meta FAIR’s new open-source brain encoding model, released in late March. It was trained on more than 1,000 hours of fMRI data from 720 subjects who watched films, listened to podcasts, read text, and looked at images. Feed it a video, and it outputs a predicted whole-brain fMRI response. It generalizes to people it has never scanned before, at roughly 70 times the resolution of the original TRIBE model, with two to three times better accuracy than previous approaches on held-out stimuli.

Tribe v2 from the website.

That is technically impressive work from a talented team.

The coverage has been something else entirely…

“Digital twin of the human brain.”

“The death of the focus group.”

“Computational neuromarketing.”

One piece called it a paradigm shift. Several described it as a tool that allows companies to test how their ads affect consumer brains before running them.

None of those writers appears to have spent much time with the reverse inference problem.

The problem with brain activation data

When TRIBE v2 processes a video ad and reports activation in the inferior temporal cortex, the visuospatial network, and the prefrontal cortex, what have you actually learned?

You have learned that visual object recognition is occurring. That spatial processing is engaged. The prefrontal cortex is doing something, which is almost always true, because it participates in working memory, cognitive control, emotion regulation, response inhibition, reward processing, and roughly 50 other documented functions depending on which meta-analysis you consult. The insula? Same thing, it’s “always on.”

This is the reverse inference problem, formalized by Russell Poldrack at Stanford and subsequently replicated as a critique throughout the neuroimaging literature. Brain region X activating does not reliably tell you that cognitive process Y is occurring, because nearly every region participates in multiple processes. Seeing prefrontal activation in response to an ad creative tells you approximately nothing about whether that ad will be remembered, preferred, or acted on.

Get Thomas Zoëga Ramsøy’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Let me give you an example: a story about me and my umbrella! Let’s take this in a few very simple steps:

1. You learn that when it rains, I will open my umbrella
2. You now hear someone say that I open my umbrella
3. What do you conclude?

Most people would say: “It must be raining, it’s logical!” But that’s wrong, because you weren’t told that I only open my umbrella if it rains. I could, in fact, open my umbrella for many other reasons: to test it, to dry it off, to show it because I like it, etc.

In the same way, reverse inference is when we make the same mistake about brain activity and mental processes:

1. We learn about a study showing that when someone does a particular task, a specific brain area is engaged. For example, when people feel fear, their amygdalae are active (yes, one amygdala, two amygdalae).
2. We now read about another study that finds strong amygdala activity.
3. Our conclusion, if we’re not well-versed in cognitive neuroscience, is that the person must feel fearful! After all, the amygdala was just linked to fear, right?

But this is obviously false, because the amygdala can be active for a number of reasons. For example, the amygdala has been implicated in responses to novelty, learning, and reward. My own studies have found the same, such as in one study where we found a stronger activity the easier it is to accept or decline a bet. Hardly a situation where people are fearful…

In a study we found that the ease of accepting or rejecting a bet — which could also be seen as the certainty in which one makes a decision — was related to increased amygdala activity.

What TRIBE v2 produces is a prediction of canonical neural activation patterns. What the coverage is telling marketers they can now do is predict consumer behavior. Those two things are separated by an entire field of work that does not yet exist, and pretending otherwise does not make it so.

Seeing prefrontal activation in response to an ad creative tells you approximately nothing about whether that ad will be remembered, preferred, or acted on.

What the model was actually built to do

TRIBE v2 is not a bad model. It is a good model being misrepresented, largely by people outside the team that built it.

Its actual purpose is in silico neuroscience: running virtual brain experiments computationally so researchers do not need to test every hypothesis with an expensive, slow, IRB-gated fMRI study. For mapping language networks, testing sensory-processing hypotheses, or accelerating clinical research into conditions such as aphasia or Alzheimer’s, a tool like this has genuine value. The zero-shot generalization across unseen individuals and languages is a real methodological advance.

But TRIBE v2 was not trained on behavioral outcomes. There is no layer in the model connecting predicted activation patterns to recall, preference, purchase intent, or anything else a marketer would actually want to know. The model produces a brain map. Someone would need to build and validate the entire translation layer between that map and meaningful consumer behavior. That body of work does not currently exist, and the resolution improvements in v2 do not change that.

The discriminatory ability question

There is a further issue that almost never surfaces in the coverage but matters a great deal for any applied use case.

Even setting aside the reverse inference problem, TRIBE v2 predicts population-averaged activation across 720 subjects. It describes what the canonical brain response looks like at a group level. It does not tell you whether this ad will outperform that ad for this audience on this platform.

I have written about discriminatory ability as a distinct validation requirement in consumer neuroscience (and another paper on the way), separate from the standard psychometric criteria by which most methods are evaluated. A metric can be theoretically valid, reliably measured, and completely unable to distinguish between real advertising assets in practice. Predicted group-average fMRI activation fits that description for marketing purposes. It is well-suited to describing the neural architecture of visual processing. It is not suited for creative optimization decisions.

What would actually need to change

For a model like TRIBE v2 to become genuinely useful in marketing applications, three things would need to happen:

• A validated decoding layer mapping specific activation patterns to behavioral outcomes, tested against held-out real-world data rather than internal model benchmarks
• Individual asset-level predictions that preserve the variance of population averaging deliberately remove
• Constrained experimental designs that actually isolate the cognitive processes being claimed, rather than whole-brain activation maps that trigger the reverse inference problem at every turn

None of this diminishes what the FAIR team built. It reframes what it is: a significant tool for computational neuroscience, and a long way from a consumer insight platform.

That distinction matters because neuromarketing has spent two decades cycling through exactly this pattern. A new method arrives, the coverage overstates the application case, expectations inflate, results disappoint, and credibility takes another hit. EEG went through it. Eye tracking went through it. Implicit association testing went through it.

TRIBE v2 deserves better than becoming the next entry in that cycle.

Meta built a real tool for brain science. The brain lighting up is not the insight. The insight is knowing what the activation means for behavior, and that question remains firmly unanswered.