Yes, under tightly controlled laboratory conditions, AI has inferred aspects of what a person is viewing from functional MRI (fMRI) signals. Depending on the experiment, a model may select a viewed image from a candidate set, classify an object, estimate visual features, generate an approximate scene, or retrieve a caption with similar meaning. Those results are not the same as reading anyone’s thoughts or producing a pixel-perfect copy of private experience.
The distinction matters: these systems use specialized scanners, participant-specific training data and a precisely defined task. fMRI measures blood-oxygenation changes indirectly, and generative models can add plausible details that were not uniquely present in the measured signal.
As an Amazon Associate I earn from qualifying purchases.
What “infer what you’re looking at” means
Brain-decoding papers use several different tasks. Calling all of them “image reading” hides important differences in what the model is allowed to do and how success is measured.
Recommended Free Tools
| Task | What the model produces | What a successful result establishes | What it does not establish |
|---|---|---|---|
| Candidate identification | One choice from a supplied image set | The brain response contains enough information to distinguish the viewed item from alternatives under that test | Unconstrained recovery of any image a person might see |
| Feature or category decoding | Visual attributes, such as shape, scene properties or object category | Some information about visual content can be predicted from selected brain areas and representations | Recovery of the exact photograph or the viewer’s complete perception |
| Image reconstruction | A newly generated image resembling the viewed scene | Decoded signals can guide a generative model toward a semantically or visually similar output | A photograph-like record in which every pixel came from the brain signal |
| Caption retrieval | A text description selected from a large caption corpus | Visual brain activity can be mapped to language representations closely enough to retrieve related descriptions | Verbatim access to a person’s private thoughts or an exact sentence they silently formed |
In other words, “infer” usually means statistical prediction against a known target. It does not mean that an algorithm opens a universal visual feed from the brain.
#1 Best Overall
- Real-Time EEG Neurofeedback Headband for Brainwave Monitoring: Monitor your brain activity in real time using advanced EEG sensors. Track key brainwave patterns such as alpha, beta and theta waves to understand how your brain responds during meditation, focus sessions, relaxation and sleep preparation.
- Smart App with Guided Meditation & Brain Training – No Subscription Fees: The companion app offers guided meditation and neurofeedback exercises with real-time brainwave feedback. As your mind calms and focus sharpens, visuals and sound become clearer, helping you practice mindfulness, enhance focus, improve sleep, and train mental control.
- Track Your Brain Training Progress: The app records your sessions and brainwave data, letting you monitor meditation duration, focus levels, and training history. Download your data freely to track progress and understand your brain performance.
- Train Focus and Calmness with Real-Time Neurofeedback: Turn brain activity into meaningful feedback that guides your mind toward deeper calm and concentration. Neurofeedback training helps make meditation more effective while strengthening awareness, focus and relaxation.
- Soft Hydrogel Sensors for Better Comfort & Signal Stability: Flexible hydrogel skin-contact sensors adapt naturally to your forehead, improving comfort and maintaining stable signal transmission. The low-impedance hydrogel interface enhances EEG signal quality for more reliable brainwave analysis.
How an fMRI decoder works
The experiments follow a pipeline that is much more constrained than a headline about mind reading suggests.
1. Record brain responses while the person views known stimuli
The participant lies in an MRI scanner while images, videos or other stimuli are shown. The visual cortex responds, and the scanner records changes in blood oxygenation over time. This blood-oxygen-level-dependent signal is an indirect and relatively slow proxy for neural activity; it is not a direct measurement of individual neurons firing.
Some visual reconstruction work uses high-field 7-tesla fMRI. The scanner, sequence, visual presentation and alignment procedures are part of the experimental setup. Results from that setup should not be generalized to an arbitrary hospital MRI, a phone camera or another type of brain scan.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →2. Fit a model to that participant’s responses
Researchers pair each measured response with the image or label that produced it. A decoder learns statistical relationships between patterns in visual areas and representations of the stimulus. Depending on the study, those representations may be receptive-field features, hierarchical computer-vision features, latent variables from a generative model or embeddings from a language model.
This training step is why participant-specific data matters. A model trained on one person’s responses does not automatically work for another person. Even when researchers align people using a shared neural-code procedure, the cross-individual mapping itself requires corresponding fMRI data and validation.
3. Decode a held-out response
After training, the participant views new material, and the model receives only the resulting brain signal. It then predicts a candidate, feature vector, category, image representation or caption. A fair test keeps the new stimulus separate from the examples used to fit the model.
4. Evaluate against a defined target
Evaluation can ask whether the correct image was selected from a set, whether a category was classified correctly, whether a generated picture resembles the target in human or machine measures, or whether a retrieved caption matches the scene’s meaning. These scores are not interchangeable. A high candidate-selection rate cannot be compared directly with a semantic-similarity score from a caption task.
What major experiments have demonstrated
Identifying a viewed image from many alternatives
Nishimoto and colleagues used receptive-field models of visual cortex to identify which natural image a participant had seen from a large collection of novel images. The authors wrote: “We show that these receptive-field models make it possible to identify, from a large set of completely novel natural images, which specific image was seen by an observer.”
Rank #2
- manufacturer: Neurosky
- Item Trademark: NEUROSKY
- Item Weight: kilograms pounds 03968320716 018
- Item Category: Headphones
This is a strong demonstration that visual information in fMRI can support recognition among candidates. It is still a selection problem: the algorithm is choosing the best match from options, not generating an arbitrary unseen picture from nothing. The candidate set, participant, visual area and model all form part of the result.
Reconstructing a scene with Brain-Diffuser
The 2023 Scientific Reports paper by Furkan Özçelik and Rufin VanRullen introduced Brain-Diffuser, a two-stage framework that combines decoded fMRI features with the generation capabilities of Versatile Diffusion. The system creates a natural-scene image that is consistent with information extracted from the brain response.
The output can look convincing because the generative model supplies texture, layout and other details that make a coherent image. That plausibility is not proof that each visible detail was uniquely recovered from the participant’s cortex. The correct description is an inferred or reconstructed image, not an exact photograph of subjective experience.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Decoding object categories from seen or imagined items
Other work has decoded object categories from fMRI responses to objects that participants saw or imagined. These studies use hierarchical visual features and examine how information differs across brain areas. A category-level result can indicate that the signal supports distinctions such as one class of object versus another, while saying little about the precise instance, viewpoint, color or personal associations.
Showing the effect of attention
In an experiment with two superimposed images, reconstructions changed depending on which image the participant attended to. The model had been trained on fMRI responses to single natural images, then applied to the overlapping-image situation. This provides evidence that attention alters the information available to a particular reconstruction pipeline. It does not show that an unrestricted decoder can recover every item in a cluttered visual field or determine all of a person’s thoughts about it.
Retrieving captions from visual brain activity
A 2025 Nature Machine Intelligence study mapped visual fMRI activity to language-model embeddings and used nearest-neighbor search in a corpus of 3.1 million Google Conceptual Captions to describe viewed images. The corpus size is a property of the study’s retrieval system, not a decoding-accuracy percentage.
Caption retrieval demonstrates a meaning-level bridge from visual brain activity to language. A returned caption can describe the scene’s gist while differing from any sentence the participant would have spoken or written. The result therefore belongs to semantic decoding, not verbatim thought transcription.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Why a generated reconstruction is not a brain photograph
Generative reconstruction has two information sources: the measured fMRI pattern and the model’s learned knowledge about what images normally look like. If the signal indicates a road, trees and a building but does not determine the exact texture of the leaves, a diffusion model can fill in a likely texture. The image may be visually persuasive while containing details selected by the prior learned from other images.
Rank #3
- Train like a Jedi! Use your own mind to control the movements of a tablet-projected hologram. Works with the same innovative brain-sensing EEG technology used in medical and military devices.
- Feel the force awaken in you with the ultimate experience for Star Wars fans of all ages. Bring the world of Luke, Yoda and Darth Vader into reality more fully than other games or action figures.
- See if you can become a Jedi Master! Like your own Jedi academy, train your powers of concentration through 25+ levels from battling droids to levitating an X-Wing with real Star Wars music, sound effects and instruction from Master Yoda.
- More than a toy or game, this interactive science-based experience can be enjoyed together for hours by families and fans of all ages. Inspires STEM-based learning and education through Star Wars.
- Includes Bluetooth brain-sensing headset, training base, app (download on App Store or Google Play) and science learning poster. Requires iPad or Android tablet (not included). Does not work with Kindle or iPad Mini.
That is why researchers distinguish decoded features from generated content. A reconstruction should be judged by the task’s stated metric—such as category accuracy, similarity to the target or human identification—not by whether it appears subjectively realistic. Realism alone can reward a model for producing a plausible scene even when it has guessed the wrong details.
How accurate are these systems?
There is no single accuracy number for “AI reading a brain scan.” The number depends on the task, candidate set, participant, scan protocol, training duration and evaluation method.
- Candidate identification: success means selecting the correct item from the alternatives used in the experiment. Increasing or changing the candidate set changes the difficulty.
- Category classification: success means assigning the right label, which can be easier than recovering an individual image.
- Feature prediction: success is measured against a representation of visual properties, not necessarily against pixels.
- Image reconstruction: scores may assess semantic or perceptual similarity, and different metrics can favor different kinds of output.
- Caption retrieval: similarity to a reference description measures shared meaning; it does not require the exact wording and should not be reported as word-for-word recovery.
Do not compare results across papers unless the stimuli, number of choices, participants, brain regions, training data and metric are aligned. A study that identifies one image from a fixed set and a study that generates a new scene are answering different questions.
Can the same decoder work on anyone?
Usually not without additional calibration. Functional anatomy, scanner characteristics and response patterns vary between people. A decoder trained on many hours from one participant is tuned to that person’s neural code.
Cross-person studies can learn a conversion between corresponding responses, but that conversion is itself a trained component. It does not turn a decoder into a plug-and-play tool for an unscanned person. Claims that a model works universally should specify how many participants supplied training data, what alignment procedure was used and whether the test person was independent of training.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Can an MRI read your thoughts?
That wording overstates the evidence. The visual studies described here ask participants to look at controlled stimuli and then infer properties of those stimuli. They do not provide unrestricted access to memories, beliefs, intentions or every spontaneous thought.
Even adjacent language-decoding work illustrates the distinction. An NIH summary of a 2023 fMRI language study reported that the decoder more often produced text that did not exactly match the original wording but captured its gist. That finding is useful context for understanding semantic versus verbatim decoding; it is not a visual-reconstruction accuracy result.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteA person’s subjective experience also contains information that a stimulus label cannot express. Two people can look at the same photograph while attending to different details, recalling different memories or feeling different emotions. A model that predicts the photograph’s category has not thereby recovered those private reactions.
Rank #4
- Learn about your brainwaves, train your meditation, and develop your own applications with the mindwave mobile wireless headset.
- Bt/ble Dual mode module and support iOS, Android, PC, and Mac platform. Detects raw-brainwaves, eeg power spectrums (Alpha, beta, etc.), esense meters for attention, meditation, and future algorithms.
- More than 100 brain training games and educational apps available from the NeuroSky online store. Uses a single AAA battery (not included) for 8-hour battery run time
What a credible claim should disclose
When reading a paper, demonstration or news report, check the following details before accepting a claim that AI can see through a scan:
- Modality: Is the evidence from fMRI, and does it use a particular field strength such as 7T? Do not transfer the result to EEG, CT or an unspecified “brain scan.”
- Stimulus control: Were images shown during scanning, or was the model asked to decode an internally generated image or thought?
- Task definition: Is the system identifying one candidate, classifying a category, predicting features, generating an image or retrieving a caption?
- Candidate restrictions: Was the correct answer selected from a known set, or was the output unconstrained?
- Training data: How much participant-specific data was collected, and were the test stimuli held out?
- Evaluation metric: Is performance measured by exact match, category accuracy, perceptual similarity or semantic similarity?
- Human comparison: If people judged the reconstructions, were they asked about identity, meaning or visual detail?
- Generalization: Were new participants, scanners or stimulus types tested, or only the conditions used for development?
Practical and scientific limits
fMRI is indirect and slow
The blood-oxygenation signal reflects neural activity through vascular changes and has limited temporal resolution. It is informative about distributed visual responses but does not provide a frame-by-frame electrical recording of every neuron.
Scanning is demanding
Participants must remain in a scanner, follow the viewing protocol and tolerate a noisy, restrictive environment. Motion, changes in attention and differences in how a person views the stimulus can reduce the quality of the data. These requirements are far removed from a consumer camera or a routine casual scan.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Models can exploit shortcuts
A decoder may rely on regularities in the stimulus set, the participant’s viewing behavior or the scanner session. Strong performance on one curated dataset does not prove robustness to different images, lighting, cultures, visual styles or mental states.
Meaning and detail can diverge
A model may correctly infer that a scene contains a person outdoors while missing identity, expression, text, small objects or the participant’s focus. Conversely, a generated image may include detailed elements that are visually plausible but unsupported by the signal.
Privacy and consent implications
These systems are research methods, not established consumer mind-reading products. The work cited here requires laboratory fMRI, controlled stimuli and trained models; it does not substantiate an app or device that lets an ordinary user decode another person’s vision.
Because the data are participant-specific and costly to collect, consent and governance are central. A person should know what stimuli are shown, what models are trained, how raw scans and derived representations are stored, and whether data could be reused for a different decoding task. A model that can infer a viewed category under consented conditions should not be treated as authorization to infer private mental content in unrelated settings.
What the evidence supports today
As of the latest cited visual-caption work in 2025, the defensible conclusion is narrow but significant: AI can extract useful information about viewed visual content from fMRI in controlled experiments. Researchers have demonstrated candidate-image identification, category decoding, attention-sensitive reconstruction, generative scene reconstruction and caption retrieval.
Those demonstrations depend on the scan modality, the participant, the training procedure and the evaluation task. They show inference from a biological signal, not a universal mind-reading capability. Until a claim specifies those conditions—and separates semantic similarity from exact recovery—the safest interpretation is that the model is making a constrained prediction, with the generative system filling in what the brain data do not uniquely determine.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
