Meta's TRIBE v2: what it predicts, what it does not and what changes for content creators

Every so often a news story appears promising that someone has finally found the secret button to persuade people. This one is worth reading calmly, because there is a real, valuable part and a layer of interpretation that runs faster than the facts.

What Meta published

On March 26, 2026, Meta’s research team (FAIR) introduced TRIBE v2 (Trimodal Brain Encoder), a model that predicts how the human brain responds to almost anything that is seen or heard. According to Meta, it relies on more than 500 hours of fMRI recordings from more than 700 people and improves on previous methods by 2 to 3 times on movies and audiobooks. It is the evolution of the model that won the Algonauts 2025 brain encoding challenge.

In practice, it works like this: you give it a video, an audio clip or a text, and the model returns a prediction of which regions of the cortex would activate on average. It is useful for research, because it makes it possible to test neuroscience hypotheses without putting every volunteer in a scanner.

Two details that are usually left out of the headlines:

• The license is non-commercial. The code and weights are on GitHub under CC BY-NC 4.0. That means you cannot put it in a paid service or in a brand’s campaign.
• It was trained with video, audio and text in natural situations, not with isolated still images. The academic papers already using it point to this as a limitation for studies with photos or static designs.

What it does not say, even though it has been repeated

In the following weeks, more dramatic readings circulated: that Meta already manipulates its audience with this tool, that it can suppress critical thinking, that it is a persuasion machine. None of those claims appears in what Meta published, and none can be verified with the available information. It is worth separating three things:

1. Fact: there is a model that predicts average brain responses to audiovisual stimuli.
2. Hypothesis: that some company could use something like this to optimize content. It is plausible, but it has not been shown to happen with this model.
3. Opinion: that this is dangerous or liberating. It depends on who uses it and under what rules.

Predicting activation in the cortex is not the same as predicting a purchase decision. Between a brain pattern and a click there is context, price, trust and a thousand other factors. Anyone selling “neuromarketing” with TRIBE v2 as a guarantee is selling a promise the model does not make.

What a creator or a business can learn

Even if you cannot use the model, the work leaves useful ideas:

• Attention and emotion can be studied. Videos that keep someone attentive usually combine a clear stimulus with a change of pace. Look at your own retention metrics and note the second at which people leave.
• Your real audience is still the best lab. Test two versions of a video with your audience and compare saves, comments and messages. That does predict your sales.
• Ethics is an advantage. If the use of tools like this becomes common, brands that clearly explain what they promise and what they do not will have trust that is harder to copy.

What comes next

This line of research is worth following. If licensed commercial applications appear in 2027, it will be time to ask what they measure, with what data and with what proof that they help sell. Until then, the best “prediction” you have is still the one you get by asking your customers and watching what they do.

Sources

• Meta, TRIBE v2 demo, accessed on October 24, 2026.
• Meta FAIR, TRIBE v2 repository on GitHub (CC BY-NC 4.0 license).
• AI at Meta, TRIBE v2 announcement, March 2026.