Amid ex-Google exes quitting their jobs so they can speak freely, screenwriters rioting in Hollywood and companies cutting jobs, artificial intelligence (AI) is not slowing down at all.
In fact, it is crossing frontiers that normally belong only to science fiction.
A team of researchers from Berkeley and the University of Texas at Austin has developed a decoder capable of reaching into the meanders of the mind. Thus reading thoughts and translating them into a file. The study and results were published in the magazine Nature Neuroscience.
Artificial intelligence reading minds, the study
An interface that allows the brain and computer to communicate could have many scientific and practical applications. However, currently noninvasive language decoders are only able to identify stimuli among a small set of words or phrases.
Others that exist or are under development, such as Elon Musk’s Neuralink, are invasive instead because they require surgery.
The study authors, however, presented a noninvasive decoder. This device reconstructs, through functional magnetic resonance imaging (fMRI), continuous language from recorded cortical semantic representations.
The machine is based on the GPT language model, which is the precursor to the current GPT-4 model.
“We are getting the model to decode continuous language for extended periods of time with complicated ideas,” said Alex Huth, one of the authors of the paper and professor of neuroscience and computer science.
This device, which unlike other decoders places no limits on the subjects’ use of words, is intensively trained with fMRI scans from brain responses generated by the person listening. From stillness and paying attention, to recordings in which stories are told for 15 to 16 hours.
The results of the study
The researchers explained in Nature that the results are not an exact word-for-word transcription of what the subjects have in their minds.
But, that the decoder captures the gist of the thoughts. Thus generating “intelligible word sequences that recover the meaning of perceived speech, imagined speech, and even silent videos. Thus demonstrating that a single decoder can find application to a range of tasks.”
The machine is therefore able to produce text that matches the meaning of the original words.
The mental privacy problem
Although such a breakthrough may prove useful, if not vital, in particular cases – just think of people who are mentally conscious but unable to speak, for example, after a severe stroke – such a device also poses a serious threat to mental privacy.
“We take very seriously the fears that [this technology] could be used for the wrong purposes, and we’ve been working to avoid that,” said Jerry Tang, lead author and doctoral candidate in computer science.
From the tests performed, the researchers say there are no risks so far. This is because the decoder requires many hours of voluntary cooperation from the subject who volunteers to train it.