Scientists have developed an artificial intelligence tool called Concept2Brain that predicts how human brain waves respond to photographs or written text. In a study published in Nature Communications, the researchers report that the open-access platform generates synthetic electroencephalography (EEG) recordings that resemble real human brain activity during visual and emotional perception. The system could allow researchers to simulate brain-wave experiments on a computer before recruiting human volunteers.
Electroencephalography, or EEG, uses sensors placed on the head to measure tiny electrical voltage changes at the scalp produced by brain activity. Because this technique tracks neural activity millisecond by millisecond, scientists often repeat an image or event many times and average the time-locked EEG traces to isolate event-related potentials, which are characteristic wave patterns tied to seeing, thinking, or feeling.
In recent years, artificial intelligence has increasingly been paired with brain imaging to interpret what a person is experiencing. Much of that effort has focused on decoding, such as an fMRI study in which volunteers listened to stories, watched silent movies, or imagined narratives while a computer model reconstructed the general meaning of those stories from their brain scans.
Far less attention has gone to the reverse problem, known as encoding, in which a computer starts with an image or phrase and predicts the fast-moving scalp electrical waves an observer would produce. To make such predictions possible, neuroscientists first built large reference libraries linking thousands of everyday object concepts to human brain activity, including a 2023 multimodal neuroimaging project that cataloged brain scans and similarity ratings across 1,854 object categories.
Don’t miss the next big psychology finding
Every morning, PsyPost sends a short briefing on the most interesting new peer-reviewed research in psychology and neuroscience.
Building on large visual datasets, a 2022 study showed that computer encoding models could simulate millisecond-by-millisecond EEG responses directly from object photographs. Other research has established that scalp electrical waves reflect not only what an object is, but also how emotionally intense it feels, including a 2008 EEG study co-authored by one of the same researchers that linked emotional arousal to the size of electrical brain responses.
Because recording real EEG data from human participants is costly and time-consuming, Alejandro Santos-Mayo, then at the University of Florida and the Complutense University of Madrid, and his colleagues set out to create a generative tool that bridges conceptual meaning and scalp electrical activity. By connecting a brain-wave compression network with a multimodal AI system, the team aimed to capture complex, multi-variable patterns in brain waves and predict participant-specific EEG responses from both pictures and written descriptions.
“The evolution of artificial intelligence (AI) tools has opened a new, non-linear, multivariate approach to analyzing and predicting brain data,” Santos-Mayo, the study’s lead author, now at NeuroSpin, a brain-imaging center at Université Paris-Saclay, and the Department of Psychology at the University of Florida, and author of Emociones con sentido, told PsyPost. “A fundamental objective of science is not merely to observe nature, but to understand it—and understanding any phenomenon requires being able to predict it.”
“Here, we aimed to create a new tool for predicting electroencephalography (EEG) brain signals,” he said. “By predicting these signals, we can generate synthetic EEG samples without relying on slow, costly data collection, while gaining a deeper understanding of how different neurophysiological components manifest in the brain. The ultimate goal of Concept2Brain is to offer a free, open-access tool for anyone interested in simulating EEG activity.”
To train Concept2Brain, the researchers used EEG recordings from two earlier passive picture-viewing experiments conducted by members of the same research group, encompassing 88 university volunteers receiving course credit. In the first experiment, 45 participants viewed 120 naturalistic photographs twice for one second each against a black background, while in the second, 43 participants viewed 240 different photographs once for two seconds each against a gray background.
Across both datasets, the 360 unique photographs were divided evenly among pleasant, neutral, and unpleasant scenes, and varying the screen layouts was intended to keep the model focused on picture content rather than visual borders. Brain activity was recorded from 128 scalp sensors plus a reference sensor at 500 samples per second, providing 13,640 clean 1.5-second trials after filtering out blinks and movement noise.
Concept2Brain links two deep neural networks to convert real-world concepts into brain waves. The first network, called a conditional variational autoencoder, compresses each 129-channel EEG recording into a compact summary of 256 numbers while tagging the recording with a participant identifier so it preserves individual differences in brain-wave shape. Once trained, the decoder half of this network can turn any 256-number summary back into a smooth, trial-averaged EEG pattern for any of the 88 virtual participants.
The second network serves as a translator built around OpenAI’s CLIP system, an existing AI model that maps both pictures and written words into a shared 512-number conceptual space. A three-layer neural network learns to translate those conceptual numbers into the 256-number brain-wave summary, enabling users on the public Concept2Brain website to upload an image or type text and download simulated brain responses. The authors also posted the model’s code, trained weights, and underlying EEG datasets in public repositories.
“I would like to invite readers to try the model for themselves,” Santos-Mayo said. “Our primary goal has always been to create an open, accessible tool for everyone, which is why we developed the online platform at concept2brain.brainbodybehavior.science. No installation is required, and the model can be run directly in your browser.”
To test how well the system worked, the team first ran a 10-round cross-validation check, repeatedly training the model on 80 percent of the photographs, monitoring overfitting on 10 percent, and predicting brain waves for the remaining 10 percent held out of training. They also tested whether the system reproduced classic wave landmarks, rhythmic brain oscillations, and individual participant traits, and evaluated it against independent public brain-imaging datasets and written text prompts.
When predicting responses to held-out photographs, Concept2Brain produced synthetic brain waves that closely matched real human recordings across most images, with alignment emerging around 100 milliseconds after picture onset, when visual signals reach the cortex, and remaining strong until roughly 700 milliseconds. One notable exception was an erotic photograph that showed a negative correlation with the real brain data, which the authors suggest may have occurred because CLIP’s built-in content-safety guardrails distorted how that image was represented.
When fed the 360 photographs it had been trained on, the model’s simulated waveforms closely tracked the empirical recordings across all scalp sensors and time points, reproducing early visual waves as well as the late positive potential, a slow wave over parietal sensors between 400 and 800 milliseconds linked to emotional engagement. The synthetic data also reproduced rhythmic brain changes, such as a drop in 10-hertz alpha waves after 300 milliseconds, and matched each of the 88 virtual participants most closely to their own real-world human counterpart.
Next, the researchers tested Concept2Brain on 1,078 emotional photographs from a standardized picture library that were not included in training, finding a modest positive association between the size of the simulated late positive potential and standard human ratings of emotional arousal. When given photographs of fearful and neutral expressions from 140 actors in a facial-expression database, the model also generated larger face-processing waves around 170 to 200 milliseconds and larger late emotional waves for fearful faces than for neutral faces, even though it had never been trained on facial-expression labels.
To check whether predictions generalized to other laboratories, the team tested Concept2Brain against an external multimodal dataset of everyday objects. In that dataset, 4 volunteers viewed 1,854 object concepts during magnetoencephalography, which records the tiny magnetic fields produced by brain activity, and 3 volunteers viewed 720 object concepts during functional MRI. Concept2Brain’s synthetic waves aligned with early visual and later semantic phases in the magnetic recordings, and correlated with functional MRI activity in early visual cortex around 75 milliseconds and in a face-processing region around 190 milliseconds.
In a second external test, the researchers entered stimuli from an independent EEG dataset of 40 people viewing 40 faces and 40 cars. Concept2Brain reproduced the spatial scalp map distinguishing faces from cars, though the authors note that overall associations across both external datasets were weaker than within-dataset tests.
“The model was trained in an entirely unsupervised manner—we never explicitly labeled the visual content of the images used during training,” Santos-Mayo noted. “One of the first pleasant surprises was the model’s ability to simulate synthetic EEG signals for images containing faces.”
In brain research, the face-processing region from the fMRI comparison is known as the right fusiform face area, and its rapid electrical signature is called the N170 component. “We know from existing literature that facial stimuli activate the right fusiform face area around 170 ms (the N170 component),” he said. “Despite the unsupervised setup, the model spontaneously and accurately reproduced this 170 ms component.”
Because CLIP handles both images and text, the researchers also tested Concept2Brain with 120 short written descriptions sorted into pleasant, neutral, and unpleasant categories, as well as short example prompts such as “a picture of an alligator.” Across all 88 virtual participants, pleasant and unpleasant descriptions prompted larger late emotional waves between 400 and 800 milliseconds than neutral descriptions.
As an illustration using just two statements about climate change, the model also produced stronger simulated brain responses to a statement emphasizing harmful consequences than to one describing mitigation strategies. The authors note that this pattern is consistent with negativity bias, the tendency for negative information to draw a stronger response. The paper reports the emotion, face, and text comparisons descriptively, without formal statistical tests.
“Another major surprise involved using written text to test the model,” Santos-Mayo added. “By leveraging CLIP—an OpenAI network capable of processing both images and text—we tested Concept2Brain’s performance using simple written descriptions instead of visual images. The resulting performance was remarkably similar to that achieved with real pictures, highlighting that our model opens a promising avenue for cross-modal brain signal generation.”
“From a cognitive neuroscience perspective, one of the key findings of Concept2Brain is the suitability of conditioned variational autoencoders (cVAEs)—a type of AI architecture—for analyzing and extracting the non-linear latent space of EEG data,” he explained. “The use of these advanced methods needs to become more widespread in our field. Outside our discipline, the average reader can view the Concept2Brain model as an exciting milestone in brain data simulation. Despite the current limitations of the model, this work marks the beginning of a new era in using AI techniques not only to observe the brain, but also to generate synthetic reproductions of its signals.”
There are a few important things to keep in mind when interpreting these simulations. Although Concept2Brain accepts written prompts, the model was trained exclusively on brain recordings from people viewing photographs rather than reading sentences, meaning text inputs estimate how a brain might respond to a related visual concept rather than simulating actual language comprehension. The authors caution that text-prompted predictions are less validated and may be noisier than picture-based outputs.
“While this model demonstrates the ability of AI to predict and simulate brain signals, users must interpret synthetic data with caution,” Santos-Mayo cautioned. “Synthetic data cannot replace empirical data collected from human participants, nor can it be used to draw definitive scientific conclusions. Instead, its purpose is to provide a fast method for testing hypotheses or estimating sample sizes prior to real-world data collection.”
“The current version of Concept2Brain was trained on 88 human participants across 360 distinct images,” he said. “This sample size established the viability of the AI architecture, though future iterations will require larger datasets to achieve higher fidelity.”
In addition, Concept2Brain produces smoothed waveforms that resemble averages across many repeated trials, so it cannot capture the trial-to-trial variability of raw single EEG recordings or model underlying cellular mechanisms. Its correspondence with real brain activity also weakened when tested against external datasets collected with different recording hardware and tasks.
“It is natural to feel apprehensive about rapid AI advances, especially when they touch on something as personal and complex as the human brain. However, I want to clarify that this model cannot replace a real human brain, nor can it generate a synthetic ‘brain-like’ mind. The current state of the field is nowhere near that capability,” Santos-Mayo emphasized.
“Developing such a model would require virtually impossible amounts of data from a single individual, with no guarantee of success,” he added. “Concept2Brain makes much more modest claims: it serves as a solid proof of concept and a technical milestone in brain signal simulation, but it poses no threat to society, nor should it trigger dystopian fears.”
Finally, the current model learned its concept-to-brain mapping from 88 volunteers viewing 360 photographs, which limits both the diversity of virtual brains and the range of visual concepts it represents. Because the current architecture depends on OpenAI’s CLIP to interpret stimuli, any blind spots, biases, or content filters inside that third-party system can carry into the generated brain waves, though the authors designed the framework so CLIP can be swapped out for newer AI models in future versions.
“This research opens new horizons not only for predicting brain signals, but also for enhancing our broader toolkit in cognitive neuroscience,” Santos-Mayo shared. “Currently, I am working on developing new non-linear methods based on neural networks to decode brain activity. Concept2Brain proved the suitability of cVAE architectures for extracting the latent dimensionality of EEG signals—meaning we can isolate and distill pure informational features from raw brain activity to interpret, understand, and predict cognitive variables. I am now applying similar methods to other cognitive functions and building new tools for the research community.”
“Regarding Concept2Brain specifically, we plan to release future upgrades trained on larger datasets incorporating more human participants, diverse stimulus samples, additional neuroimaging modalities, and updated AI components,” he continued. “We warmly welcome any researchers interested in collaborating on future versions.”
The study, “Concept2Brain: an AI model for predicting neurophysiological responses to text and pictures,” was authored by Alejandro Santos-Mayo, Faith Gilbert, Arash Mirifar, Anna-Lena Tebbe, Ruogu Fang, Mingzhou Ding, and Andreas Keil.
