Meta Researchers Develop System Translating Brain Activity to Text in Real Time Without Implants

Researchers from Meta's Artificial Intelligence (AI) department have announced a significant breakthrough in brain-computer interfaces (BCI). The new system, named Brain2QWERTY v2, is capable of converting brain activity directly into textual phrases in real time, completely eliminating the need for surgical implants.

The technology utilizes external magnetic sensors to record neural signals, achieving a level of accuracy previously only accessible through invasive procedures. The main goal of this project is to restore communication ability to millions of people suffering from brain injuries, strokes, or neurodegenerative conditions affecting speech.

Although intracranial electrode surgeries have already demonstrated effectiveness in neuroprosthetics, the complexity and biological risks of such interventions limit their widespread application. This new non-invasive approach offers a safe and scalable alternative to bridge this gap.

Distinction Between Mind Reading and Intention

Despite the appeal of this technology, there is a technical and philosophical distinction regarding what the system actually does. In an interview with Olhar Digital News, in the column Olhar do Amanhã, neurobiologist, futurist, and professor at the Federal University of São Paulo (Unifesp), Álvaro Machado Dias, explained to host Marisa Silva that the machine does not read thoughts, but rather records the motor intention to perform an action.

Machado Dias explained: 'They do not capture the thought; they only capture this motor intention. What happens is that when we type something, there is a 'thought capsule,' a fragment that serves for motor planning.' He added that the prefrontal cortex contains an area called the supplementary motor area, which participates in planning and remains synchronized during the text input process.

To capture this dynamic, Meta used magnetoencephalography (MEG), large equipment that combines the temporal precision of electroencephalography (EEG) with the spatial resolution of functional magnetic resonance imaging (fMRI). Essentially, the machine can display where and when electrical activity occurs in the brain while a person is planning to type.

Unlike previous methods that relied on manual identification of neural patterns, Brain2QWERTY v2 employs advanced end-to-end deep learning. The architecture processes raw signals and fine-tunes large language models (LLMs) based on neural data.

As the neurobiologist notes, the key to the system's success lies in the collaboration between the sensor and the language model. 'The machine determines slightly what is being read, and the LLM tries to translate it into a phrase context. The combination of these two things provides a reading that is not the thought itself. No one extracts a thought from people's minds, but rather from the behavior underlying it, guiding what will be written, i.e., the intention.'

To train the model, the Meta team collected data from nine volunteers over 10 hours of recording per person, totaling about 22 thousand recorded phrases while participants typed under the MEG device.

The results impressed the scientific community: Brain2QWERTY v2 achieved an average word recognition accuracy of 61%, significantly surpassing the 8% recorded by other non-invasive methods. For the participant with the best results, accuracy reached 78%, with more than half of their phrases translated with no more than one minor error.

Furthermore, researchers observed that accuracy improves logarithmically linearly as the volume of data increases. This indicates that the performance gap separating non-surgical methods from invasive implants can only be gradually reduced by increasing the scale of model training.

Risks of Manipulation and Neurorights

The expert emphasized that mind reading would pose a risk of 'unauthorized' intrusion and manipulation, whereas intention preserves human autonomy and consciousness. He believes that progress in this area opens doors for urgent discussions regarding neurorights and cognitive privacy. If a device can smoothly predict what you are about to write, the commercial use of this information by large technology corporations raises concerns.

The Unifesp professor warned: 'The thing starts working. Soon the machine, through this intention, will identify what you are about to type, and therefore what you indirectly aim for or think, and offer you advertising.' He noted: 'That is the risk. It is the risk of old-fashioned manipulation through context to convince you of something—be it a product or an ideological message, which could become much more problematic.'

The discussion of neurorights also extends to the management of biological and genetic data that may predict neurodegenerative diseases or mental disorders, risking discriminatory use by insurance companies and corporations in countries lacking strict legal protection.

In conclusion, despite rapid AI progress in neuroscience, scientists agree that the core of the human mind remains intact. Transforming neural impulses into conscious and abstract experience remains unexplored territory. Machado Dias concluded: 'These current mind readings are always proxies [approximations]. They are not actually in the symbolic domain. When you imagine something, you think in words, and all of this is related to what remains a mystery to us: how this neural activity, resembling Morse code, transforms into the mental experience we have. Nobody knows, and perhaps this is the biggest mystery we will have to solve forever.'

olhardigital.com.brAugust 13, 13:32

Meta uses artificial intelligence to block underage accounts on social media in Australia

Meta removed over 756,000 profiles suspected of belonging to Australian citizens under the age of 16 between December and June. This blocking was implemented following the enactment of legislation that imposes restrictions on adolescent accounts on social media platforms.

According to Reuters, the total number of deactivated accounts amounted to 462,000 on Instagram and 294,000 on Facebook. Meta itself stated that this work will continue as pressure from Australian authorities on technology companies increases.

This new volume of removals exceeds the data Meta disclosed in January, when the company reported the deletion of 331,000 Instagram accounts and 173,000 Facebook accounts.

The company assures its commitment to complying with Australian law, despite having expressed opposition to the measure along with other platforms. The government, in turn, accuses social networks of being insufficient in combating the continuous use of services by minors.

The legislation came into effect on December 10, created in response to authorities' concerns about the impact of social media on the physical and mental health of children and adolescents.

Details of oversight and future disputes

Data provided by Meta indicates that the legal dispute may intensify in the coming months. The Australian government suggested increasing the maximum fine for non-compliance with the law to A$99 million (approximately R$362.7 million) and granting greater power to the regulatory body to request documents from companies.

Furthermore, Australia alleges that the platforms have created circumstances that could hinder the enforcement of the ban. This issue also draws the attention of other governments that are evaluating the adoption of similar restrictions for younger users.

A parliamentary hearing has been scheduled for Friday, where representatives from Meta, TikTok, Google (owner of YouTube), and Snapchat will testify before regulatory and governmental authorities.

How Meta identifies minors

Meta employs artificial intelligence to identify profiles that may belong to individuals under the age of 16. Instead of relying solely on the age information provided by the user, the system analyzes various indicators present in the profile.

Among the analyzed elements are references to birthdays and mentions of school series, as well as reports regarding potential minor accounts. The company also uses age estimation tools based on photographs. Previously, the method consisted of trying to deduce the age based on the user's behavior and activity on the platform.

Another change implemented prevents anyone from attempting to create a new account if a previous profile was removed on suspicion of belonging to a minor.

Meta communicated that monitoring will continue, implying that the number of accounts removed in Australia is expected to keep growing.

tecnoblog.netAugust 7, 01:19

Meta launches Muse Code, a new AI focused on programming and game development

Meta has introduced Muse Code, a new artificial intelligence agent designed for programming and game creation. This model is based on the recently launched Muse Spark 1.2 and has the capability to execute large-scale coding tasks.

With this launch, the company aims to compete with giants in the generative AI sector, such as Anthropic and OpenAI. On X, Mark Zuckerberg shared results from internal tests, comparing the new agent with various Large Language Models (LLMs) on the market, including Claude Opus 5 and GPT-5.6 (in Sol and Terra versions).

Preliminary results indicate that Muse Code's performance is still at an intermediate level when compared to its direct competitors. A notable feature of Muse Code is its ability to handle extensive projects, which it achieves by dividing tasks among multiple agents.

Two games developed with this new AI were presented on Meta's blog, displaying simplified mechanics and graphics. Muse Code is already available in beta version. Muse Spark 1.2 is specifically aimed at coding activities, ranging from bug fixing to data interpretation, while maintaining the qualities of Muse Spark 1.1 for general AI agents.

Although related, no information was provided regarding the model's improvements in cybersecurity or scientific research. The figures presented by Meta suggest that the AI falls slightly short of current Anthropic models, although it manages to surpass GPT-5.6 in certain situations. In analyses conducted by Terminal Bench 2.1, Muse Spark 1.2 ranks second, with performance close to Claude Opus 5, contrary to what Meta's internal benchmark shows.

Among the points highlighted by Zuckerberg's company are the ability to manage prolonged tasks and maintain context understanding throughout the entire process. Additionally, the AI can map and improve its own training, surpassing Muse Spark 1.1 in this aspect. Focused on programming, the new agent promises good performance in game development, which often requires multiple layers of code and various logical conditions.

In the release of Muse Spark 1.2 on its blog, Meta also demonstrated the AI's potential with three examples, two of which were simple games. The first, named Embervault, features an isometric perspective and a board format where the player controls a robot that must eliminate constantly pursuing enemies; the gameplay demands fluidity, but the website showed freezes during the test. The second game, Avo Lawn, is a 2D tower defense type, where zombies advance slowly towards the player. In this game, the user collects seeds and acquires fruits to strengthen their defenses, lasting five attack waves, presenting a slow pace but with potential to be a good mobile game.