Meta TRIBE v2: an artificial brain predicts how people will react to your ad before it runs

Meta taught AI to predict what the brain does. The average one, for now

TRIBE (TRImodal Brain Encoder) comes from the Brain & AI team at Meta FAIR. The first version won the Algonauts challenge in 2025, where 263 teams competed to predict how the brain responds to films. Version 2 came out on 26 March 2026, followed in May by the paper A foundation model of vision, audition, and language for in-silico neuroscience (d'Ascoli, Rapin, King et al., arXiv 2605.04326).

The principle is surprisingly modest. TRIBE v2 doesn't see or hear anything by itself. It borrows three existing models that already understand the world: V-JEPA 2 for video, Wav2Vec-BERT 2.0 for audio and the 3-billion-parameter Llama 3.2 for text. Those stay frozen, and on top of them a transformer learns to translate their internal representations into brain activity. Think of a simultaneous interpreter listening to three colleagues at once and rendering everything into a single language: the language of blood flow in the cortex. The output is a predicted signal at 20,484 points on the cortex plus 8,802 subcortical voxels, twice per second.

For marketers, the key question is whose brain it predicts. The public version models an average person. It can't model a specific audience, a segment or the 35-year-old suburban parent your media plan is built around. The model knows where the brain lights up. What that person buys afterwards is beyond its reach.

Data from 720 brains. Most of the hours came from four people watching Friends

Meta collected no new data for TRIBE v2. The model is built entirely on public datasets from university labs, and the heaviest is Courtois NeuroMod from Montreal: four volunteers spent 268.7 hours in the scanner, mostly watching the sitcom Friends and feature films. Add podcasts (LeBel 2023), short videos (BOLD Moments) and video clips (Wen 2017), and the model trained on 451.6 hours of recordings from 25 people.

It was then tested on people it had never seen: 695 participants and 666.1 hours from four more datasets, including the Human Connectome Project movie data and the spoken stories in Narratives. That makes 720 people and 1,117.7 hours in total. Meta's social posts mention "500+ hours from 700+ people" and the demo page talks about 70,000 voxels. I stick to the paper, because at least its numbers sit in a table.

By neuroscience standards the results are strong. For new people, with no fine-tuning at all, the model's prediction correlates with the group's average response at about 0.4. That is twice as close as a typical participant's own real scan gets to the same average. With one hour of data from a new person, the model is two to four times more accurate than a linear model trained from scratch. Accuracy keeps rising with more data and hasn't hit a ceiling yet. I'll admit what impressed me most was something else: in virtual experiments the model reproduced decades-old findings, such as the brain areas for faces, places and written words, without anyone lying in a scanner.

You can download it in an afternoon. You can't make money with it

Trying TRIBE v2 is easy. The code is on GitHub in the facebookresearch/tribev2 repository, the weights are on Hugging Face as facebook/tribev2, and Meta also runs a web demo and a Google Colab notebook.

How to use TRIBE v2: GitHub, Hugging Face and the demo

Install it with pip install -e . on Python 3.11 or newer. Load the model with TribeModel.from_pretrained("facebook/tribev2"). Then build an events table from a video, an audio track or a text, and the predict function returns predicted activity for about 20,000 cortical points over time. To read the output, it helps to have someone who knows what the fusiform face area is.

The catch is the licence. TRIBE v2 is released under Creative Commons BY-NC 4.0, and NC means non-commercial. An agency or research vendor can't build a paid service on it without Meta's permission. Meta also lists neuroscience, AI research and healthcare as the intended uses. Advertising is absent from its materials, and there has been no word about connecting the model to Meta's ad products.

The authors spell out the limits themselves. fMRI measures blood flow with a delay of several seconds, so fast reactions between shots slip through. The model covers only vision, hearing and language. And in the authors' own words, it treats the brain as a passive observer: it models perception and stops before decisions. A map of the brain still knows nothing about the shopping cart.

Then came the first independent tests. In July 2026, Sahu and Pandey checked whether TRIBE's predicted brain activity could forecast which parts of 48 YouTube videos people replay. The correlation came out at 0.058, statistically insignificant, and simple loudness and motion measures did just as well. Rodrigues looked at video memorability: the brain projection helped on one dataset and lost to the plain video model on another. Both are preprints without peer review, but they point the same way. Meanwhile, hobbyist "ad scorers" built on TRIBE have popped up on Hugging Face. None of them is validated.

Neuromarketing in practice: when the brain knows more than the survey

The idea that the brain reveals more than a questionnaire has a name: neuroforecasting. It also has a few famous results. In 2012, Emily Falk and colleagues (Psychological Science) showed thirty smokers three anti-smoking campaigns. Activity in the medial prefrontal cortex ranked the campaigns correctly by the real increase in quitline calls. The smokers' own ratings and the expert panel got the order wrong. The same year, Gregory Berns and Sara Moore found that 27 teenagers' reward-centre response to unknown songs correlated with sales over the next three years (r = 0.32), while their stated liking barely did (r = 0.11).

In 2017, Alexander Genevsky, Carolyn Yoon and Brian Knutson forecast which Kickstarter projects would get funded. The brain got it right 59.1% of the time, the participants' own choices 52.9%, close to a coin toss. In 2014, Jacek Dmochowski and colleagues (Nature Communications) recorded EEG from twelve viewers of Super Bowl ads: how similarly their brains responded explained 66% of the variance in ratings from more than seven thousand people. Their own ratings explained 59%. So there is an edge, smaller than the headlines suggest.

The most rigorous test came from the US Advertising Research Foundation. Venkatraman et al. (Journal of Marketing Research, 2015) compared six methods on 37 TV ads, 26 of them with real sales data. The best single predictor was the plain survey. Only fMRI added anything on top (ventral striatum activity raised adjusted R² by 59%). EEG, eye tracking, biometrics and implicit tests added nothing significant.

In 2016 Nielsen reported the opposite: across roughly sixty ads, EEG explained 62% of in-store sales and facial coding 9%. That came in a press release, though, without peer review.

The whole field stands on a narrow base. In their 2018 review, Knutson and Genevsky counted just seven fMRI studies that forecast market behaviour, each with 18 to 41 participants. Median statistical power in neuroscience is around 21%, according to Button et al. (2013). And in 2019, Lisa Feldman Barrett and colleagues showed that facial expressions map onto specific emotions far more loosely than any facial-coding software assumes. Neuroforecasting today is a promising youth player: it scores goals, mostly in friendlies so far.

Dmochowski left an interesting note. If you could pin down which content features drive the shared brain response, he argued, you could forecast straight from the content without measuring a single brain. That is exactly what TRIBE v2 does. It just predicts the brain so far. Nobody has checked it against the market yet.

Who tests ads today, and how: creative testing tools and ad testing companies

Creative pretesting means trying an ad on a sample of people, or on a model, before you put media budget behind it. It measures whether people understand it, whether they notice it, what they feel and whether they remember the brand. Increasingly with AI.

Creative pretesting in 2026 is a crowded market. I counted around forty companies, and they fall into seven groups by what they measure. The survey is the oldest; models with no respondents are growing fastest. The money is flowing into AI ad testing: Dragonfly AI raised £5M in January 2026, Alison.ai $5.1M in February, Electric Twin $14M, and Stanford spinout Simile a full $100M to simulate human decision-making.

Watch the accuracy numbers. Neurons claims over 95%, Dragonfly AI 89%, Attention Insight up to 96%. All three measure how well a predicted heatmap matches real eye tracking. That kind of accuracy is a weather forecast that nails the cloud cover and never asks whether people will go to the beach.

Neurons openly says it hasn't published a link between its Impact Score and in-market results. Only a handful work with business outcomes: System1 points to the IPA databank, Kantar and WARC showed the best creative earns more than four times the profit, and Adverteyes cites a Mars case study with a sales lift of up to 19%.

Facial coding, meanwhile, is in retreat. Affectiva was folded into iMotions in January 2025, Realeyes spun its ad business out as Adverteyes in November 2025, Zappi measures emotion with emoji and System1 simply asks. Regulation is arriving too. Since 2 February 2025, the EU AI Act bans inferring emotions from biometric data at work and in education. Consumer panels count as high-risk systems, with obligations from 2 December 2027, and people must be told they are being analysed from 2 August 2026.

Central Europe has its own players. Prague-based Behavio offers a self-serve ad testing platform with a 500+ respondent panel, implicit tests and AI attention heatmaps, with results in three to seven days. Ipsos, STEM/MARK, MEDIAN and NMS offer classic survey pretests, and MEDIAN with the Knowlimits group launched the KNOW.attention study with eye-tracking glasses and EEG headbands last year. Full disclosure: we work with both Behavio and Knowlimits as partners.

Pretesting can save a campaign. It can also kill it

Creative is the biggest lever marketing has. In 2023, NCSolutions analysed around 450 consumer goods campaigns and attributed 49% of the sales effect to creative. Marketers in a Westwood One survey put its share at 19%. I cover the older Nielsen numbers in the piece on why creativity drives most of the result, and more large studies in our roundup of marketing effectiveness research. Testing the thing that decides half the outcome is worth it.

But pretesting can do damage. In The Long and the Short of It (IPA, 2013), Les Binet and Peter Field went through 996 campaigns and found an uncomfortable pattern. Campaigns that went through a classic pretest reported very large sales effects within six months 10 percentage points more often than untested ones. After two years or more, 10 points less often. Campaigns that tracked the brand continuously instead of pretesting were 31 points ahead after two years.

The explanation is easy to guess, even though the data is correlational and the pretests of the 1980s to 2000s were mostly rational. A classic pretest asks whether the ad is clear, believable and something the respondent would watch again. The answer is yes once you sand it down into something they've seen a hundred times. The creative team then spends three weeks filing off the edges until "the one with the gorilla" becomes a nice family at breakfast. A pretest makes a good servant and a terrible gatekeeper.

What to do on Monday morning

TRIBE v2 won't replace pretesting yet, and its licence keeps it out of commercial ad work anyway. It does hint at where the field is heading: from expensive labs to models that estimate audience response from the content in seconds. A sensible approach for 2026 looks like this:

1. AI heatmaps to triage variants. Neurons, Dragonfly AI or Attention Insight will show in minutes whether the brand, the product and the main message are visible. Treat it as a legibility check. It won't forecast sales. More tools are in our guide to AI marketing tools.
2. A human panel for the finalists. Emotion and brand recall on the two or three strongest variants. Behavio, System1 or Kantar can do it in days.
3. A live test on the platform. An A/B test in Meta Ads Manager compares variants on real people and real money. Meta recommends running it for two weeks to a month.
4. Long-term brand tracking. Tracking awareness and mental availability tells you whether the campaign built the brand. Let the pretest advise; let results decide. We cover measurement with fewer cookies in a separate piece.
5. TRIBE v2 as an idea lab. For research and your own curiosity, outside paid client work. It answers "does this scene light up the face area?" beautifully. "Will it sell?" is beyond it for now.

A brain in a box sounds like science fiction, and for now it works as a very clever mirror of the average viewer. A mirror is worth looking into only when you know what you're looking for.

Sources: d'Ascoli, Rapin, King et al., A foundation model of vision, audition, and language for in-silico neuroscience, arXiv 2605.04326 (2026) · Meta AI, TRIBE v2 blog and demo (2026) · d'Ascoli et al., TRIBE: TRImodal Brain Encoder, arXiv 2507.22229 (2025) · Sahu & Pandey, arXiv 2607.01400 (2026) · Rodrigues, arXiv 2607.16292 (2026) · Falk, Berkman & Lieberman, Psychological Science (2012) · Berns & Moore, Journal of Consumer Psychology (2012) · Genevsky, Yoon & Knutson, Journal of Neuroscience (2017) · Dmochowski et al., Nature Communications (2014) · Venkatraman et al., Journal of Marketing Research (2015) · Knutson & Genevsky, Current Directions in Psychological Science (2018) · Button et al., Nature Reviews Neuroscience (2013) · Barrett et al., Psychological Science in the Public Interest (2019) · Nielsen Consumer Neuroscience (2016) · NCSolutions, Five Keys to Advertising Effectiveness (2023) · Kantar & WARC, The Art of Proof (2023) · Binet & Field, The Long and the Short of It (IPA, 2013) · Regulation (EU) 2024/1689 on artificial intelligence · websites and press releases of the companies named (as of October 2026)

This connects to what we do: creative & idea making, growth & performance marketing and marketing strategy. We dig into the psychology of decisions in our piece on behavioral science in marketing and into long-term ad effects in the one on the 60/40 rule.