Are we ready for brain-computer implants?

Note: Transcripts are automatically generated from episode audio, and are not fully corrected for spelling, grammar, and formatting.

MATT ANGLE: You might trust people, you might trust Matt Angle, or you might trust Elon Musk, but who knows a year or two from now who’s going to be at the helm? One thing that’s more durable is trusting a business model.

NITA FARAHANY: We want to empower people to make choices, and that’s both a right to use technology, access their own brain when they want to, and a right to not have it used against them. Ultimately, it’s about empowering people to make choices for themselves.

RANA EL KALIOUBY: Hi, everyone. We are fresh off Masters of Scale Summit. It’s an annual two-day event in San Francisco where we hear from some of the world’s most visionary voices, including co-host Reid Hoffman, Google’s Eric Schmidt, and Signal’s Meredith Whittaker. I had a fantastic time taking it all in, and I hosted several conversations of my own and will be bringing those to the show, starting with today’s episode.

Live on the Summit stage, I spoke with two top experts in the area of brain-computer interfaces. Matt Angle is the CEO of Paradromics, building tiny chip implants for the brain. These devices turn neural signals into words without the need to speak or type. And AI makes these devices much more powerful, processing and interpreting brain data at unprecedented speeds. Our second guest is Nita Farahany. She teaches law and philosophy at Duke University and is the author of the book, The Battle for Your Brain: Defending the Right to Think Freely in the Age of Neurotechnology. She’s pioneered this idea of thought privacy, the right to own your own thoughts in an era where computers can actually read minds. I’m so excited for you to hear this one, so let’s get into it. I’m Rana el Kaliouby, and this is Pioneers of AI, a podcast taking you behind the scenes of the AI revolution.

Copy LinkHow brain implants turn thoughts into speech

We are going to be talking about computer implants in your brain, but before we do that, we wanted to show a short video of your work, Matt, at Paradromics. Tell us, who are we about to see, and what are they doing?

ANGLE: We’re going to have a conversation, and pretty quickly I think it’s going to jump to mind reading. A lot of people wonder, how soon is that? What would that look like? What you’re going to see in this next clip is a woman in our first clinical trial. She’s a participant with motor neuron disease. She has severe dysarthria, which means that she has lost her ability to speak clearly. She enrolled in the trial to help us understand how brain-computer interfaces can help people who have lost the ability to speak communicate again. What you’ll be seeing is her thinking words in her inner monologue, and as she thinks those words, they will appear on the screen directly from her brain.

EL KALIOUBY: And to clarify, this is quite different from eye tracking, where you point at a trackpad and, with your eye movement, it types.

ANGLE: Yeah. Many of us have seen Stephen Hawking give speeches, but those were prerecorded speeches where he was communicating through motor movements, kind of typing things out, at something like one or two words per minute. You’ll see this comes a lot faster because it’s reading directly from the mind.

CLINICAL TRIAL: You guys have been married almost 45 years, so what advice do you have for us for marriage?

PATIENT: Don’t sweat the small stuff.

CLINICAL TRIAL: What would be your message to your coworkers as you’re retiring?

PATIENT: It has been an absolute privilege to be a part of the team.

EL KALIOUBY: Very powerful work. Can you clarify what part of the brain this device is capturing?

ANGLE: This device is implanted in an area called 55B, and it’s kind of the Grand Central Station for language in the brain. It represents a lot of different things. When we’re trying to speak, the language that we’re trying to say with our muscles is there. It also represents when we think, in the quietness of our own mind, a series of words. Even right now, in your 55B, my words are in your brain.

EL KALIOUBY: Is that like the Debbie Downer voice in my head? Okay.

ANGLE: Yeah. If you’re ruminating and thinking about things, it would be there. It’s very interesting because it’s received, expressed, and internalized language.

EL KALIOUBY: Nita, why does it matter where this device is implanted?

FARAHANY: First, incredible, right? It’s amazing to see that somebody who is unable to speak is able to speak again, and I’m deeply excited about that. Matt and I have been in conversation for a while, since my book came out in 2023, The Battle for Your Brain: Defending the Right to Think Freely in the Age of Neurotechnology. What I said was, we’re pretty quickly moving backward from this kind of intentionally communicated speech, which is what you were doing at that point. That is, people painstakingly, in their motor cortex, forming the words to try to speak and communicate them into the world, and getting to the place where the things that you’re passively thinking in your mind, like your actual inner monologue — the mental space where you are you, and all of those things that you’re not saying out loud and might never say out loud, can suddenly be accessed. That’s amazing for people who need it. It’s terrifying that that space, that space of mental privacy, that thing where you don’t say, “No, I really love your couch,” or “No, honey, that dress looks amazing on you,” could suddenly be accessed. My book really looked at how that might happen not for a person like this, who we absolutely want to make it possible for, but for the rest of us who might opt into a similar kind of technology, a wearable technology, where it might overcome the last frontier of privacy we have, our last bastion of privacy.

Copy LinkHow close mind reading technology really is

EL KALIOUBY: So is this just around the corner? We’ve been talking about brain-computer interfaces for years. How close is this technology?

ANGLE: It’s already around the corner. You just saw it. We’ve been having conversations about this, and a lot of the early BCI, just like Nita was describing, involved people trying to say something. The physical act of trying to say something, even if you’re paralyzed, produces motor signatures that allow you to move an arm or create the sounds of speech. But already we’re going up one level in the cognitive stack and detecting the words that you are thinking. I can tell you that BCI companies will continue to go up that stack because the number of people who can get a communications prosthesis increases as you move up the stack. There are people who have stroke, and the area of the brain that we’re recording from right now might be damaged, so we might have to go a level up. That’s a really good thing, right? Someone who has aphasia, who can’t express themself anymore, being able to go directly into the area of the brain where they’re thinking, where they have the intention and the semantic concept of what they want to say, and record that—that’s really exciting.

FARAHANY: Right. It’s also incredible. Again, let me just say, even just for this patient, you want to go back as early as possible. Me having this conversation right now with you, I’m not painstakingly choosing every single word. It just comes naturally, right? My thoughts flow from me naturally. How I’m feeling flows naturally. To give people that capability makes a lot of sense. In fact, that’s why a lot of consumer-based companies are starting to look at how to develop wearable neurotechnology that allows you to interact with your devices. So instead of your smart glasses requiring that you actually move around, you have a neural band from Meta that allows motor neuron activity to be picked up at your wrist, from your brain down your arm to your wrist. It picks up your intention to type or swipe or move, and increasingly more companies are trying to get to that point. They’re not going to be as good, obviously, at getting an inner monologue as what yours does. But we want to go backward. It makes it more seamless. It makes it more seamless for her. And the further back you go, there are at least two problems, right? The first one is there is no safe space. There’s your implicit automatic reaction. Think about the bias that you have. You’ve worked on it for a lifetime. You know that you have particular biases that are internal. You automatically react to somebody or something, but you filter it at many later steps of consciousness. That’s picked up. And who are you? Are you the person who automatically reacted to something? Or are you the person who expressed it into the world?

EL KALIOUBY: Yeah.

FARAHANY: Does the device decide which one you are?

ANGLE: As an engineering problem, you can already see the difference between attempted speech and imagined speech. So we can choose to build a product that separates those streams and lets users decide how it’s used. But I think the point is, the company that records the data will have access to it, and you have to trust whoever you’re allowing to put something in your brain.

EL KALIOUBY: Yeah.

FARAHANY: Right.

EL KALIOUBY: Trust is key.

FARAHANY: Yeah.

EL KALIOUBY: Yeah.

FARAHANY: Well, I trust you, Matt, but some of the other players…

EL KALIOUBY: Well, let’s talk about that.

FARAHANY: Right?

Copy LinkWhy business models shape trust in brain tech

EL KALIOUBY: Because you’re very focused on this idea of a prosthetic and restoring speech and communication to people who really need it. There is a consumer version of this, which is kind of the human 2.0, where we’re all getting these implants in our brains. It’s giving us entirely new abilities. Elon Musk’s Neuralink is the best example of that. Is that the vision? Is that the endgame for Paradromics?

ANGLE: Well, something that I would say is, you might trust people. You might trust Matt Angle, or you might trust Elon Musk. We may be there. Who knows a year or two from now who’s going to be at the helm? But one thing that’s more durable is trusting a business model.

One of the places where we’ve seen tech go in the wrong direction is when the users stop being the customers. A lot of us have experienced social media apps and a number of different platforms where we’re getting a service, but someone else is paying, and so the services end up being steered toward them and not us. That’s one place where I think we have good alignment right now for implantable brain-computer interfaces. We’re building medical devices. These are devices that are prescribed by a doctor for a person. They’re paid for by a hospital, reimbursed by a payer, and the business model is channeled toward us delivering things for that patient in a regulatory setting.

FARAHANY: You’re not going to be selling my thoughts. But Matt is right for two reasons that this is a problem. One is, already there are fights happening where people are trying to acquire the neural data. They’re trying to acquire the neural data from the companies. They’re trying to acquire the neural data from the research institutions. There are patients who want access to their neural data who are being told, “It’s not your data, it’s our data because we collected it.” And there are offers on the table to buy the neural data, and if those offers come through, then suddenly it becomes the commodification of neural data. The trust collapses. And it’s your brain, right? It’s literally what you’re thinking and what you’re feeling. I mean, that’s not hyperbole anymore. It’s literally what you’re thinking and what you’re feeling.

The second thing we haven’t talked about yet is the convergence with AI. This part’s really important, because in order for all of this to be interpreted, it’s interpreted using AI models. More and more of the AI companies have major investments in neurotech companies. Elon, right? OpenAI. They’re increasingly on-device. They’re integrated, right? Let’s unpack that with neurotech, because I can understand you need machine learning and predictive AI to translate the signals, but what role does AI play?

ANGLE: Here’s a great practical intersection.

EL KALIOUBY: Okay.

ANGLE: So we’re building these tools for people who are paralyzed. Why wouldn’t we want to give them excellent Claude integration? Why wouldn’t we want them to be able to prompt AI agents, even maybe humanoid robots, directly with their thoughts? We have visual language action models now that could control a robot by just thinking about what you want it to do.

FARAHANY: Can I ask you, as you’re building this, to unpack something? Up on the screen, your patient had “pleasure” that then switched to “privilege.”

Is that an AI model predicting what her brain signals actually mean, and then she’s overriding what the predicted next word is and correcting it?

ANGLE: That was what’s called an n-gram model. We hear about large language models. That’s a small language model, and that’s where, if you know the phonemes, like the first five phonemes, then you can predict the sixth phoneme. What you saw was a great example: the sentence was, “The excitement of the race,” and then it switched to, “The excitement of the work.” That was an example of, early on, the language model thinking it knew better, and then when it got to the end of the sentence, it went back and looked and said, “No, actually, based on the brain data, it was this sentence.”

FARAHANY: Right.

ANGLE: So you could see it flipping back and forth in real time.

FARAHANY: Right. So when you have the model that lives on the devices, and we’re moving further and further back into your thought processes, and the model is predicting what you think and feel, and then there’s output, how do we make sure that there’s still a human who is thinking those things rather than experiencing what a model has predicted based on early signals that then have feedback into the person and change how they’re thinking? Do you then think, “No, no, no, this was not the word I thought,” and it corrects it?

ANGLE: Well, there are a few things. One is, we all want…

FARAHANY: But you might not even experience it that way, right?

ANGLE: Yeah.

FARAHANY: You would experience it as if this is my thought, because—

ANGLE: Well, not necessarily. We all went through this with voice-to-text. Voice-to-text was terrible for a while.

FARAHANY: Yep.

ANGLE: And in a noisy environment, it still sometimes isn’t that good.

FARAHANY: But that’s later. You’re speaking it. You’re talking about picking up thought processes before you’ve ever articulated them.

ANGLE: Not what we’re doing right now, but yes, that’s a thing that will exist. Then you have a shared-control architecture where some of that signal is coming from you and some of it is coming from the AI. I’m actually glad you mentioned it. That’s one of the reasons why we’re really long on high data rate. Take speech, for instance. If each of you were sitting there writing down what I was saying right now, which I hope you all are, it takes about 40 bits per second to transcribe human speech. So if you want to build a speech prosthesis based on a wearable device with two bits per second, that means the AI is filling in 38 bits per second of the language you’re communicating. If you’re trying to control a robot, for example, we could probably have a humanoid robot that would play golf for us if we said, “Hey, go out there and play golf for me,” and it would come back and say, “Yeah, 18 holes in one. Congratulations.” But if you wanted embodied control of that robot — if you’re paralyzed and you want to play golf and feel what it’s like — you would need a lot more data to communicate back and forth. So there’s this trade-off with agency: the more data you have, the more you get to embody that interaction.

Copy LinkHow AI fills gaps in neural decoding

EL KALIOUBY: Can we talk about use cases? We’re talking about language for now, but is it language-agnostic? If I’m thinking in Arabic, would it capture that or not?

ANGLE: Right now, the models are trained on a specific language. But if you think about it, automated translation can happen in milliseconds.

EL KALIOUBY: Okay. And is language the only use case, or are there other use cases?

FARAHANY: That’s what I was about to ask. You mentioned Claude integration, and I think we kind of breezed past that for a moment. Is it multimodal? Image? What you were really hinting at is something different, which is not just Claude integration the way we’re thinking about it — like, let’s go type in our output — but the superpowered human who is now merged with Claude, and then maybe a bionic body, and then maybe flying swarms of drones.

ANGLE: Okay. “Merged” is a little much. That’s a loaded word, but I’ll tell you what I think will happen in the next five years.

FARAHANY: There is a company called Merge, right?

EL KALIOUBY: Okay.

FARAHANY: So that’s different.

ANGLE: Someone who is in a clinical trial, who has a brain-computer interface in a higher cognitive area, is going to be able to communicate intentions directly with an AI system. Right now, the way that we interact with Claude or ChatGPT is that you type a prompt, realize it wasn’t exactly what you wanted, then do a new prompt, and there’s an iterative process. Or, if you’re a prompt engineer, you write a huge prompt at the beginning. But a lot of the time, that context, that intention, exists in your mind from the outset. So it will be possible for people who have certain kinds of BCIs to put more context into that initial prompt.

FARAHANY: So instead of vibe coding, you’re vibe living, right?

EL KALIOUBY: Yeah.

FARAHANY: It’s just sort of like that, right?

ANGLE: I mean, you have to think about it. Right now, we’re looking at cognitive outsourcing. Part of when you go to write something, the act of trying to sequence it and put it into words makes you realize your initial intention wasn’t fully baked out. So some of that work you do is important. If we start working in a different flow, we will have to adapt.

Copy LinkWhy thought privacy needs stronger protections

EL KALIOUBY: Okay. So you talk about thought privacy.

FARAHANY: Yeah.

EL KALIOUBY: Tell us, what does that mean, and what is it going to look like?

FARAHANY: Yeah.

EL KALIOUBY: Because we need it, clearly.

FARAHANY: Well, my next book, which comes out this March, is called Cognitive Extinction: Saving Human Thought While We Still Can. I get a little darker from the battle. There was the battle, and then —

ANGLE: We lost the battle.

FARAHANY: We’ve lost—

ANGLE: Cognitive extinction.

FARAHANY: No, there’s still hope.

EL KALIOUBY: Need hope.

FARAHANY: Saving human thought while we still can.

But here’s why: I care deeply about a space of mental privacy, right? There’s this space that is you, the space where you can cultivate your own thoughts, the space where you can try on things without anybody else judging you, and having that space that’s your own. I’m deeply excited for the patients you’re helping and think that it’s incredibly important. But I worry about what you were just talking about with offloading. Already, a lot of the early work shows that when you use AI systems to bypass metacognition, the part of you that’s actually present and working through it, the you, the agency that we’re talking about in this conference, kind of fades into oblivion, and the machine takes over, right? You sit in the backseat. You’re like an automaton. We’ve all done that. We’re in our car, we’re driving somewhere, we don’t even realize we’re doing it, and we get there. And I don’t mean our Tesla. I mean literally, we’re meant to be driving ourselves. So there’s a world in which, the way you describe it, it’s the person, through metacognition, prompting and changing. And then there’s the world in which you just sort of sit back and you’re like, “Well—”

EL KALIOUBY: The AI agent is basically running the show.

FARAHANY: “The AI agent can sort of sense what I want.”

ANGLE: But driving is a great example, because there are some processes where we want to be more present and more engaged, and then there’s driving to work, when we want to think about something else, or doing our taxes, where we actually want minimal engagement.

FARAHANY: But will we even know the difference, right? The world in which AI is literally inside our brain, especially because we’re already seeing AI is very good at doing things like manipulating. Or you describe the prompt, and you think, “That’s not quite what I want.” Usually it works the opposite way, which is you put some incoherent prompt into the system, it comes back and says, “You’re so smart. That is exactly what you just said.” It’s not at all, but you’re like, “Yeah, that is what I said,” and you ratify it. Suddenly it becomes the thing you put out into the world. How is that, when you have merger between— I know you don’t like the word merge. Joining? These things happening together.

EL KALIOUBY: Well, can you give us an example of how you safeguard against some of the ways this can go wrong? What are some of the protections we can put in place?

ANGLE: Well, it is a product design exercise.

So one example is that if, in this neural state space, you can separate things that people want to say from things that people want to think privately, just allowing those things to be treated differently according to the way the user wants. Now, one question would be: Are there constraints that you even want to put in place against the user’s intentions? So there’s a world where you say, “Someone should be free to decide how present they want to be when filling out their tax returns,” or there’s a world where we might, as a society, say, “I don’t know, actually. I think you should still have to pay attention to this,” or, “You should still have to go through the thought process of writing out exactly what you want your robot to do.”

FARAHANY: Okay, now you’re starting to make me nervous on the other side, because who gets to decide those kinds of choices.

ANGLE: Is it you?

FARAHANY: Probably. Yeah, it’d probably be me.

Copy LinkWhat agency means when brains meet AI

EL KALIOUBY: Wait, this is a heated debate. I love it. Final question to both of you: You’re taking human in the loop to a whole new level.

FARAHANY: Okay.

EL KALIOUBY: The theme of this Summit is agency.

FARAHANY: Yeah.

EL KALIOUBY: Right? So in this world of convergence of AI and neurotech, what does it mean to be human, and what does it mean to be in control? I’m going to give you the last word, so Matt, you go first.

ANGLE: From my perspective, we want to give people the most optionality, and you have the most optionality if you have a high data rate between the system you’re controlling and yourself. If you have to defer to the AI to do a bunch of things for you, and you don’t have enough input into steering that process, then you do start outsourcing things significantly. If you have enough data to, let’s say, embody the robot or make sure that the speech that’s coming out is coming from you and not from a language model, then that gives you the most options. You still have to make a lot of hard decisions, but I think, from the engineering perspective, data gives options.

EL KALIOUBY: Amazing. Nita, how do we stay in control?

FARAHANY: I mean, how do I say no to that? I have been advocating for a basic right, which is the right to self-determination over your brain and mental experiences. I call it cognitive liberty.

We are making great headway in laws and policies, but also in design, and working with companies like Matt’s and others to develop standards as to what that means, where the technology is designed in ways that enhance user agency, autonomy, and self-determination, but also safeguard against the misuse and commodification of neural data that ultimately puts products at odds with where people are. It’s about giving people user-level controls. It’s the reason why I also resist this idea of putting limits on what people can do with their own devices. What we want to do is empower people. We want to empower people to make choices, and that’s both a right to use technology, access their own brain when they want to, and a right to not have it used against them, or used in ways that oppress them. Ultimately, it’s about empowering people to make choices for themselves.

EL KALIOUBY: Yeah. Well, thank you both so much. Thank you for joining us on the show.

ANGLE: Thanks for having us.

FARAHANY: Thank you.

EL KALIOUBY: Thank you.

FARAHANY: Thanks, everybody.