The discussion of mental privacy and the challenges of inferring internal vs. volitional speech when motor control is impaired is v important, not something i had considered before. Of course what someone says is not necessarily what they are thinking - and thus very high accuracy for semantic decoding without the ability to tell those apart would be extremely bad for agency.
I also love how they discuss the need for decoding prosody, stress patterns, and all the non "series-of-phonemes-or-morphemes" things that make spoken language expressive - i can only imagine the kind of vocal dysphoria that would come from being able to say the literal words you intend in a way that feels almost natural but still just isn't your voice. That would be especially important in error correction contexts, where spoken language is constantly replete with repair, and to signal that you need subtle prosodic and gestural cues in addition to the content of the words. Thinking of @dingemansemark 's work on this
The section on the ethics of incorporating language models that might "morph a person's intended message to conform with the statistical trends found within language corpora" is interesting to me - it's true that language is highly ordered, and a model that reduces a lot of the uncertainty over the possible structure of words would be helpful (and necessary atm, they say, given poor accuracy of speech decoders), and i wonder about the ability for the current-gen LLMs to be distilled in such a way that they can just do "autocorrection" for tuning based on syntactic features without being coercive. IDK enough about model distillation or domain tuning to know. in any case it certainly is not how the current foundational models are being developed. They argue for the need to tune speed/fluency (more language model influence) vs agency, and i wonder what that model architecture would look like if not some sort of weighted ensemble model.
The spectre of proprietary prosthetics going defunct hangs over that in a real way - if your speech is dependent on some cloud LLM platform, that's a nightmare indeed, and some serious open source work would need to be done to make a plausible alternative given the regulatory environment of medical software and hardware. The closing note on integration with other wearables is haunting - we joke about neural surveillance re: neuralink for now, but like most things, it will likely be trialed on people who have few other options.
@jonny@dingemansemark Thanks! I'd stress that this paper is really more from my colleagues than me (3rd author on a 4 author paper) as they're all much more expert than I am on the neuroscience of speech and current technical capabilities. I'm happy to have contributed ideas about agency, privacy, disability, and human-centered design. Narayan and I in particular are continuing to do work on ethical and societal implications of this research, hoping to have more on this later this year.
@jonny@dingemansemark Well I'm always glad when someone reads a paper of ours, but I didn't want to assume credit for some of the really interesting technical distinctions in the article that originated from my co-authors. Our other work is interesting too, promise! (Though, not a guarantee that it's equally interesting to you...)
Also want to re-plug in this thread the hiring announcement that started this conversation, we're looking for an empirical neuroethics postdoc and a decision neuroscience RA: https://decisionlab.ucsf.edu/hiring/
Last night #UCSF was the scene of a joyous celebration in honor of #Dr.BruceMiller's parents, Milton and Harriet Miller. #Music, #art, #poetry, and #science -- especially #neurology -- filled the halls. And lots and lots of love. So honored to be there and to have the #memoir I wrote with Bruce be one of the gifts given to the people who came.
McKenzie Funk's latest story is about one-time drug smuggler Hank Asher and how we in the United States all came to be surveilled in massive integrated databases.
@ct_bergstrom The amazing and troubling thing about this piece is how it links a single person's strange, pathological psychological makeup to structures that we now all live in and are largely resigned to (indeed, that many perceive as normal).
@ben Hi, just found your writing and appreciate these thoughts. I don't think realistically that Mastodon was or would ever be suited to this mass-democratic role, and not just because of design clunkiness. I find this Noah Smith piece useful (https://www.noahpinion.blog/p/the-internet-wants-to-be-fragmented)--the potential role for Mastodon is probably not as a common "town square," but instead a throwback to the old internet of blogs, forums and chatrooms. Those were niche spaces, requiring a lot of effort to find and appealing to tightly-knit groups. Just like most people didn't engage with those spaces, most people don't care about what Mastodon has to offer--e.g., lots of people like algorithmic feeds because they reduce the burden of choice and self-curation. But for people who desire them, smaller and focused communities can play important democratizing and organizing roles.