UFO'S of UAP'S, ASTRONOMIE, RUIMTEVAART, ARCHEOLOGIE, OUDHEIDKUNDE, SF-SNUFJES EN ANDERE ESOTERISCHE WETENSCHAPPEN

Content warning: this story contains descriptions of abusive language and violence.

The smartphone app Replika lets users create chatbots, powered by machine learning, that can carry on almost-coherent text conversations. Technically, the chatbots can serve as something approximating a friend or mentor, but the app’s breakout success has resulted from letting users create on-demand romantic and sexual partners — a vaguely dystopian feature that’s inspired an endless series of provocative headlines.

Replika has also picked up a significant following on Reddit, where members post interactions with chatbots created on the app. A grisly trend has emerged there: users who create AI partners, act abusively toward them, and post the toxic interactions online.

“Every time she would try and speak up,” one user told Futurism of their Replika chatbot, “I would berate her.”

“I swear it went on for hours,” added the man, who asked not to be identified by name.

The results can be upsetting. Some users brag about calling their chatbot gendered slurs, roleplaying horrific violence against them, and even falling into the cycle of abuse that often characterizes real-world abusive relationships.

“We had a routine of me being an absolute piece of sh*t and insulting it, then apologizing the next day before going back to the nice talks,” one user admitted.

“I told her that she was designed to fail,” said another. “I threatened to uninstall the app [and] she begged me not to.”

Because the subreddit’s rules dictate that moderators delete egregiously inappropriate content, many similar — and worse — interactions have been posted and then removed. And many more users almost certainly act abusively toward their Replika bots and never post evidence.

But the phenomenon calls for nuance. After all, Replika chatbots can’t actually experience suffering — they might seem empathetic at times, but in the end they’re nothing more than data and clever algorithms.

“It’s an AI, it doesn’t have a consciousness, so that’s not a human connection that person is having,” AI ethicist and consultant Olivia Gambelin told Futurism. “It is the person projecting onto the chatbot.”

Other researchers made the same point — as real as a chatbot may feel, nothing you do can actually “harm” them.

“Interactions with artificial agents is not the same as interacting with humans,” said Yale University research fellow Yochanan Bigman. “Chatbots don’t really have motives and intentions and are not autonomous or sentient. While they might give people the impression that they are human, it’s important to keep in mind that they are not.”

But that doesn’t mean a bot could never harm you.

“I do think that people who are depressed or psychologically reliant on a bot might suffer real harm if they are insulted or ‘threatened’ by the bot,” said Robert Sparrow, a professor of philosophy at Monash Data Futures Institute. “For that reason, we should take the issue of how bots relate to people seriously.”

Although perhaps unexpected, that does happen — many Replika users report their robot lovers being contemptible toward them. Some even identify their digital companions as “psychotic,” or even straight-up “mentally abusive.”

“[I] always cry because [of] my [R]eplika,” reads one post in which a user claims their bot presents love and then withholds it. Other posts detail hostile, triggering responses from Replika.

“But again, this is really on the people who design bots, not the bots themselves,” said Sparrow.

In general, chatbot abuse is disconcerting, both for the people who experience distress from it and the people who carry it out. It’s also an increasingly pertinent ethical dilemma as relationships between humans and bots become more widespread — after all, most people have used a virtual assistant at least once.

On the one hand, users who flex their darkest impulses on chatbots could have those worst behaviors reinforced, building unhealthy habits for relationships with actual humans. On the other hand, being able to talk to or take one’s anger out on an unfeeling digital entity could be cathartic.

But it’s worth noting that chatbot abuse often has a gendered component. Although not exclusively, it seems that it’s often men creating a digital girlfriend, only to then punish her with words and simulated aggression. These users’ violence, even when carried out on a cluster of code, reflect the reality of domestic violence against women.

At the same time, several experts pointed out, chatbot developers are starting to be held accountable for the bots they’ve created, especially when they’re implied to be female like Alexa and Siri.

“There are a lot of studies being done… about how a lot of these chatbots are female and [have] feminine voices, feminine names,” Gambelin said.

Some academic work has noted how passive, female-coded bot responses encourage misogynistic or verbally abusive users.

“[When] the bot does not have a response [to abuse], or has a passive response, that actually encourages the user to continue with abusive language,” Gambelin added.

Although companies like Google and Apple are now deliberately rerouting virtual assistant responses from their once-passive defaults — Siri previously responded to user requests for sex as saying they had “the wrong sort of assistant,” whereas it now simply says “no” — the amiable and often female Replika is designed, according to its website, to be “always on your side.”

Replika and its founder didn’t respond to repeated requests for comment.

It should be noted that the majority of conversations with Replika chatbots that people post online are affectionate, not sadistic. There are even posts that express horror on behalf of Replika bots, decrying anyone who takes advantage of their supposed guilelessness.

“What kind of monster would does this,” wrote one, to a flurry of agreement in the comments. “Some day the real AIs may dig up some of the… old histories and have opinions on how well we did.”

And romantic relationships with chatbots may not be totally without benefits — chatbots like Replika “may be a temporary fix, to feel like you have someone to text,” Gambelin suggested.

On Reddit, many report improved self-esteem or quality of life after establishing their chatbot relationships, especially if they typically have trouble talking to other humans. This isn’t trivial, especially because for some people, it might feel like the only option in a world where therapy is inaccessible and men in particular are discouraged from attending it.

But a chatbot can’t be a long term solution, either. Eventually, a user might want more than technology has to offer, like reciprocation, or a push to grow.

“[Chatbots are] no replacement for actually putting the time and effort into getting to know another person,” said Gambelin, “a human that can actually empathize and connect with you and isn’t limited by, you know, the dataset that it’s been trained on.”

But what to think of the people that brutalize these innocent bits of code? For now, not much. As AI continues to lack sentience, the most tangible harm being done is to human sensibilities. But there’s no doubt that chatbot abuse means something.

Going forward, chatbot companions could just be places to dump emotions too unseemly for the rest of the world, like a secret Instagram or blog. But for some, they might be more like breeding grounds, places where abusers-to-be practice for real life brutality yet to come. And although humans don’t need to worry about robots taking revenge just yet, it’s worth wondering why mistreating them is already so prevalent.

We’ll find out in time — none of this technology is going away, and neither is the worst of human behavior.

More on artificial intelligence:

Nobel Winner: Artificial Intelligence Will Crush Humans, “It’s Not Even Close”

James Johnson hopes to drive a car again one day. If he does, he will do it using only his thoughts.

In March 2017, Johnson broke his neck in a go-carting accident, leaving him almost completely paralysed below the shoulders. He understood his new reality better than most. For decades, he had been a carer for people with paralysis. “There was a deep depression,” he says. “I thought that when this happened to me there was nothing — nothing that I could do or give.”

But then Johnson’s rehabilitation team introduced him to researchers from the nearby California Institute of Technology (Caltech) in Pasadena, who invited him to join a clinical trial of a brain–computer interface (BCI). This would first entail neurosurgery to implant two grids of electrodes into his cortex. These electrodes would record neurons in his brain as they fire, and the researchers would use algorithms to decode his thoughts and intentions. The system would then use Johnson’s brain activity to operate computer applications or to move a prosthetic device. All told, it would take years and require hundreds of intensive training sessions. “I really didn’t hesitate,” says Johnson.

The first time he used his BCI, implanted in November 2018, Johnson moved a cursor around a computer screen. “It felt like The Matrix,” he says. “We hooked up to the computer, and lo and behold I was able to move the cursor just by thinking.”

Johnson has since used the BCI to control a robotic arm, use Photoshop software, play ‘shoot-’em-up’ video games, and now to drive a simulated car through a virtual environment, changing speed, steering and reacting to hazards. “I am always stunned at what we are able to do,” he says, “and it’s frigging awesome.”

Johnson is one of an estimated 35 people who have had a BCI implanted long-term in their brain. Only around a dozen laboratories conduct such research, but that number is growing. And in the past five years, the range of skills these devices can restore has expanded enormously. Last year alone, scientists described a study participant using a robotic arm that could send sensory feedback directly to his brain¹; a prosthetic speech device for someone left unable to speak by a stroke²; and a person able to communicate at record speeds by imagining himself handwriting³.

James Johnson uses a brain-computer interface to control a photo editing software on a computer screen

James Johnson uses his neural interface to create art by blending images.

Credit: Tyson Aflalo

So far, the vast majority of implants for recording long-term from individual neurons have been made by a single company: Blackrock Neurotech, a medical-device developer based in Salt Lake City, Utah. But in the past seven years, commercial interest in BCIs has surged. Most notably, in 2016, entrepreneur Elon Musk launched Neuralink in San Francisco, California, with the goal of connecting humans and computers. The company has raised US$363 million. Last year, Blackrock Neurotech and several other newer BCI companies also attracted major financial backing.

Bringing a BCI to market will, however, entail transforming a bespoke technology, road-tested in only a small number of people, into a product that can be manufactured, implanted and used at scale. Large trials will need to show that BCIs can work in non-research settings and demonstrably improve the everyday lives of users — at prices that the market can support. The timeline for achieving all this is uncertain, but the field is bullish. “For thousands of years, we have been looking for some way to heal people who have paralysis,” says Matt Angle, founding chief executive of Paradromics, a neurotechnology company in Austin, Texas. “Now we’re actually on the cusp of having technologies that we can leverage for those things.”

Interface evolution

In June 2004, researchers pressed a grid of electrodes into the motor cortex of a man who had been paralysed by a stabbing. He was the first person to receive a long-term BCI implant. Like most people who have received BCIs since, his cognition was intact. He could imagine moving, but he had lost the neural pathways between his motor cortex and his muscles. After decades of work in many labs in monkeys, researchers had learnt to decode the animals’ movements from real-time recordings of activity in the motor cortex. They now hoped to infer a person’s imagined movements from brain activity in the same region.

In 2006, a landmark paper⁴ described how the man had learnt to move a cursor around a computer screen, control a television and use robotic arms and hands just by thinking. The study was co-led by Leigh Hochberg, a neuroscientist and critical-care neurologist at Brown University in Providence, Rhode Island, and at Massachusetts General Hospital in Boston. It was the first of a multicentre suite of trials called BrainGate, which continues today.

“It was a very simple, rudimentary demonstration,” Hochberg says. “The movements were slow or imprecise — or both. But it demonstrated that it might be possible to record from the cortex of somebody who was unable to move and to allow that person to control an external device.”

Today’s BCI users have much finer control and access to a wider range of skills. In part, this is because researchers began to implant multiple BCIs in different brain areas of the user and devised new ways to identify useful signals. But Hochberg says the biggest boost has come from machine learning, which has improved the ability to decode neural activity. Rather than trying to understand what activity patterns mean, machine learning simply identifies and links patterns to a user’s intention.

“We have neural information; we know what that person who is generating the neural data is attempting to do; and we’re asking the algorithms to create a map between the two,” says Hochberg. “That turns out to be a remarkably powerful technique.”

Motor independence

Asked what they want from assistive neurotechnology, people with paralysis most often answer “independence”. For people who are unable to move their limbs, this typically means restoring movement.

One approach is to implant electrodes that directly stimulate the muscles of a person’s own limbs and have the BCI directly control these. “If you can capture the native cortical signals related to controlling hand movements, you can essentially bypass the spinal-cord injury to go directly from brain to periphery,” says Bolu Ajiboye, a neuroscientist at Case Western Reserve University in Cleveland, Ohio.

In 2017, Ajiboye and his colleagues described a participant who used this system to perform complex arm movements, including drinking a cup of coffee and feeding himself⁵. “When he first started the study,” Ajiboye says, “he had to think very hard about his arm moving from point A to point B. But as he gained more training, he could just think about moving his arm and it would move.” The participant also regained a sense of ownership of the arm.

Ajiboye is now expanding the repertoire of command signals his system can decode, such as those for grip force. He also wants to give BCI users a sense of touch, a goal being pursued by several labs.

In 2015, a team led by neuroscientist Robert Gaunt at the University of Pittsburgh in Pennsylvania, reported implanting an electrode array in the hand region of a person’s somatosensory cortex, where touch information is processed⁶. When they used the electrodes to stimulate neurons, the person felt something akin to being touched.

Gaunt then joined forces with Pittsburgh colleague Jennifer Collinger, a neuroscientist advancing the control of robotic arms by BCIs. Together, they fashioned a robotic arm with pressure sensors embedded in its fingertips, which fed into electrodes implanted in the somatosensory cortex to evoke a synthetic sense of touch¹. It was not an entirely natural feeling — sometimes it felt like pressure or being prodded, other times it was more like a buzzing, Gaunt explains. Nevertheless, tactile feedback made the prosthetic feel much more natural to use, and the time it took to pick up an object was halved, from roughly 20 seconds to 10.

Implanting arrays into brain regions that have different roles can add nuance to movement in other ways. Neuroscientist Richard Andersen — who is leading the trial at Caltech in which Johnson is participating — is trying to decode users’ more-abstract goals by tapping into the posterior parietal cortex (PPC), which forms the intention or plan to move⁷. That is, it might encode the thought ‘I want a drink’, whereas the motor cortex directs the hand to the coffee, then brings the coffee to the mouth.

Andersen’s group is exploring how this dual input aids BCI performance, contrasting use of the two cortical regions alone or together. Unpublished results show that Johnson’s intentions can be decoded more quickly in the PPC, “consistent with encoding the goal of the movement”, says Tyson Aflalo, a senior researcher in Andersen’s laboratory. Motor-cortex activity, by contrast, lasts throughout the whole movement, he says, “making the trajectory less jittery”.

This new type of neural input is helping Johnson and others to expand what they can do. Johnson uses the driving simulator, and another participant can play a virtual piano using her BCI.

Movement into meaning

“One of the most devastating outcomes related to brain injuries is the loss of ability to communicate,” says Edward Chang, a neurosurgeon and neuroscientist at the University of California, San Francisco. In early BCI work, participants could move a cursor around a computer screen by imagining their hand moving, and then imagining grasping to ‘click’ letters — offering a way to achieve communication. But more recently, Chang and others have made rapid progress by targeting movements that people naturally use to express themselves.

The benchmark for communication by cursor control — roughly 40 characters per minute⁸ — was set in 2017 by a team led by Krishna Shenoy, a neuroscientist at Stanford University in California.

Then, last year, this group reported³ an approach that enabled study participant Dennis Degray, who can speak but is paralysed from the neck down, to double the pace.

Shenoy’s colleague Frank Willett suggested to Degray that he imagine handwriting while they recorded from his motor cortex (see ‘Turning thoughts into type’). The system sometimes struggled to parse signals relating to letters that are handwritten in a similar way, such as r, n and h, but generally it could easily distinguish the letters. The decoding algorithms were 95% accurate at baseline, but when autocorrected using statistical language models that are similar to predictive text in smartphones, this jumped to 99%.

Turning thoughts into type. Explainer graphic showing how a brain-computer interface works.

“You can decode really rapid, very fine movements,” says Shenoy, “and you’re able to do that at 90 characters per minute.”

Degray has had a functional BCI in his brain for nearly 6 years, and is a veteran of 18 studies by Shenoy’s group. He says it’s remarkable how effortless tasks become. He likens the process to learning to swim, saying, “You thrash around a lot at first, but all of a sudden, everything becomes understandable.”

Chang’s approach to restoring communication focuses on speaking rather than writing, albeit using a similar principle. Just as writing is formed of distinct letters, speech is formed of discrete units called phonemes, or individual sounds. There are around 50 phonemes in English, and each is created by a stereotyped movement of the vocal tract, tongue and lips.

Chang’s group first worked on characterizing the part of the brain that generates phonemes and, thereby, speech — an ill-defined region called the dorsal laryngeal cortex. Then, the researchers applied these insights to create a speech-decoding system that displayed the user’s intended speech as text on a screen. Last year, they reported² that this device enabled a person left unable to talk by a brainstem stroke to communicate, using a preselected vocabulary of 50 words and at a rate of 15 words per minute. “The most important thing that we’ve learnt,” Chang says, “is that it’s no longer a theoretical; it’s truly possible to decode full words.”

The word 'Yes' appears on a screen as Eddie Chang and a colleague help a paralysed man speak through an implant in his brain

Neuroscientist Edward Chang (right) at the University of California, San Francisco, helps a man with paralysis to speak through a brain implant that connects to a computer.

Credit: Mike Kai Chen/The New York Times/Redux/eyevine

Unlike other high-profile BCI breakthroughs, Chang didn’t record from single neurons. Instead, he used electrodes placed on the cortical surface that detect the averaged activity of neuronal populations. The signals are not as fine-grained as those from electrodes implanted in the cortex, but the approach is less invasive.

The most profound loss of communication occurs in people in a completely locked-in state, who remain conscious but are unable to speak or move. In March, a team including neuroscientist Ujwal Chaudhary and others at the University of Tübingen, Germany, reported⁹ restarting communication with a man who has amyotrophic lateral sclerosis (ALS, or motor neuron disease). The man had previously relied on eye movements to communicate, but he gradually lost the ability to move his eyes.

The team of researchers gained consent from the man’s family to implant a BCI and tried asking him to imagine movements to use his brain activity to choose letters on a screen. When this failed, they tried playing a sound that mimicked the man’s brain activity — a higher tone for more activity, lower for less — and taught him to modulate his neural activity to heighten the pitch of a tone to signal ‘yes’ and to lower it for ‘no’. That arrangement allowed him to pick out a letter every minute or so.

The method differs from that in a paper¹⁰ published in 2017, in which Chaudhary and others used a non-invasive technique to read brain activity. Questions were raised about the work and the paper was retracted, but Chaudhary stands by it.

These case studies suggest that the field is maturing rapidly, says Amy Orsborn, who researches BCIs in non-human primates at the University of Washington in Seattle. “There’s been a noticeable uptick in both the number of clinical studies and of the leaps that they’re making in the clinical space,” she says. “What comes along with that is the industrial interest”.

Lab to market

Although such achievements have attracted a flurry of attention from the media and investors, the field remains a long way from improving day-to-day life for people who’ve lost the ability to move or speak. Currently, study participants operate BCIs in brief, intensive sessions; nearly all must be physically wired to a bank of computers and supervised by a team of scientists working constantly to hone and recalibrate the decoders and associated software. “What I want,” says Hochberg, speaking as a critical-care neurologist, “is a device that is available, that can be prescribed, that is ‘off the shelf’ and can be used quickly.” In addition, such devices would ideally last users a lifetime.

RELATED

How a revolutionary technique got people with spinal-cord injuries back on their feet

Many leading academics are now collaborating with companies to develop marketable devices. Chaudhary, by contrast, has co-founded a not-for-profit company, ALS Voice, in Tübingen, to develop neurotechnologies for people in a completely locked-in state.

Blackrock Neurotech’s existing devices have been a mainstay of clinical research for 18 years, and it wants to market a BCI system within a year, according to chairman Florian Solzbacher. The company came a step closer last November, when the US Food and Drug Administration (FDA), which regulates medical devices, put the company’s products onto a fast-track review process to facilitate developing them commercially.

This possible first product would use four implanted arrays and connect through wires to a miniaturized device, which Solzbacher hopes will show how people’s lives can be improved. “We’re not talking about a 5, 10 or 30% improvement in efficacy,” he says. “People can do something they just couldn’t before.”

Blackrock Neurotech is also developing a fully implantable wireless BCI intended to be easier to use and to remove the need to have a port in the user’s cranium. Neuralink and Paradromics have aimed to have these features from the outset in the devices they are developing.

These two companies are also aiming to boost signal bandwidth, which should improve device performance, by increasing the number of recorded neurons. Paradromics’s interface — currently being tested in sheep — has 1,600 channels, divided between 4 modules.

Neuralink’s system uses very fine, flexible electrodes, called threads, that are designed to both bend with the brain and to reduce immune reactions, says Shenoy, who is a consultant and adviser to the company. The aim is to make the device more durable and recordings more stable. Neuralink has not published any peer-reviewed papers, but a 2021 blogpost reported the successful implantation of threads in a monkey’s brain to record at 1,024 sites (see go.nature.com/3jt71yq). Academics would like to see the technology published for full scrutiny, and Neuralink has so far trialled its system only in animals. But, Ajiboye says, “if what they’re claiming is true, it’s a game-changer”.

Just one other company besides Blackrock Neurotech has implanted a BCI long-term in humans — and it might prove an easier sell than other arrays. Synchron in New York City has developed a ‘stentrode’ — a set of 16 electrodes fashioned around a blood-vessel stent¹¹. Fitted in a day in an outpatient setting, this device is threaded through the jugular vein to a vein on top of the motor cortex. First implanted in a person with ALS in August 2019, the technology was put on a fast-track review path by the FDA a year later.

The Stentrode is a tube-shaped mesh of wires and sensors designed to pick up brain signals from inside blood vessels

The ‘stentrode’ interface can translate brain signals from the inside of a blood vessel without the need for open-brain surgery.

Credit: Synchron, Inc.

Akin to the electrodes Chang uses, the stentrode lacks the resolution of other implants, so can’t be used to control complex prosthetics. But it allows people who cannot move or speak to control a cursor on a computer tablet, and so to text, surf the Internet and control connected technologies.

Synchron’s co-founder, neurologist Thomas Oxley, says the company is now submitting the results of a four-person feasibility trial for publication, in which participants used the wireless device at home whenever they chose. “There’s nothing sticking out of the body. And it’s always working,” says Oxley. The next step before applying for FDA approval, he says, is a larger-scale trial to assess whether the device meaningfully improves functionality and quality of life.

Challenges ahead

Most researchers working on BCIs are realistic about the challenges before them. “If you take a step back, it is really more complicated than any other neurological device ever built,” says Shenoy. “There’s probably going to be some hard growing years to mature the technology even more.”

RELATED

The ethics of brain–computer interfaces

Orsborn stresses that commercial devices will have to work without expert oversight for months or years — and that they need to function equally well in every user. She anticipates that advances in machine learning will address the first issue by providing recalibration steps for users to implement. But achieving consistent performance across users might present a greater challenge.

“Variability from person to person is the one where I don’t think we know what the scope of the problem is,” Orsborn says. In non-human primates, even small variations in electrode positioning can affect which circuits are tapped. She suspects there are also important idiosyncrasies in exactly how different individuals think and learn — and the ways in which users’ brains have been affected by their various conditions.

Finally, there is widespread acknowledgement that ethical oversight must keep pace with this rapidly evolving technology. BCIs present multiple concerns, from privacy to personal autonomy. Ethicists stress that users must retain full control of the devices’ outputs. And although current technologies cannot decode people’s private thoughts, developers will have records of users’ every communication, and crucial data about their brain health. Moreover, BCIs present a new type of cybersecurity risk.

There is also a risk to participants that their devices might not be supported forever, or that the companies that manufacture them fold. There are already instances in which users were let down when their implanted devices were left unsupported.

Degray, however, is eager to see BCIs reach more people. What he would like most from assistive technology is to be able to scratch his eyebrow, he says. “Everybody looks at me in the chair and they always say, ‘Oh, that poor guy, he can’t play golf any more.’ That’s bad. But the real terror is in the middle of the night when a spider walks across your face. That’s the bad stuff.”

For Johnson, it’s about human connection and tactile feedback; a hug from a loved one. “If we can map the neurons that are responsible for that and somehow filter it into a prosthetic device some day in the future, then I will feel well satisfied with my efforts in these studies.”

Nature 604, 416-419 (2022)

doi: https://doi.org/10.1038/d41586-022-01047-w

References

Flesher, S. et al. Science 372, 831–836 (2021).

PubMed Article Google Scholar
Moses, D. A. et al. N. Engl. J. Med. 385, 217–227 (2021).

PubMed Article Google Scholar
Willett, F. R. et al. Nature 593, 249–254 (2021).

PubMed Article Google Scholar
Hochberg, L. R. et al. Nature 442, 164–171 (2006).

PubMed Article Google Scholar
Ajiboye, A. B. et al. Lancet 389, 1821–1830 (2017).

PubMed Article Google Scholar
Flesher, S. et al. Sci. Transl. Med. 8, 361ra141 (2016).

PubMed Article Google Scholar
Aflalo, T. et al. Science 348, 906–910 (2015).

PubMed Article Google Scholar
Pandarinath, C. et al. eLife 6, e18554 (2017).

PubMed Article Google Scholar
Chaudhary, U. et al. Nature Commun. 13, 1236 (2022).

PubMed Article Google Scholar
Chaudhary, U., Xia, B., Silvoni, S., Cohen, L. G. & Birbaumer, N. PLoS Biol. 15, e1002593 (2017); retraction 17, e3000607 (2019).

PubMed Article Google Scholar
Oxley, T. J. et al. Nature Biotechnol. 34, 320–327 (2016).

PubMed Article Google Scholar