since my last post on the topic i have continued to receive pitches about "new" AI applications, but most of them have been repeats of (bad) ideas i've heard before. but there are exceptions. below are two that stand out.
last month, when a respected US basketball player died tragically, at least one human being employed by MSN's purported editorial department decided it was the perfect time to whip out the obituary-writing bot. the result, archived here, looks almost identical to most of the soulless word hash regurgitated by large language models under the "news" heading--and they might have gotten away with it, too, if it hadn't been for that headline: brandon hunter useless at 42.
msn, microsoft's news brand, has operated with mostly AI "reporters" since 2020.
in more ambiguous news, there are also now a slew of "ai obituary generators"--online tools that turn details about dead loved ones into complete obituaries. i've never had to write one but i've heard it's miserable work, so maybe this is one use-case that is actually worth the compute power. maybe.
self-driving vehicles have already made the world worse. the tesla autopilot feature was involved in 736 crashes and 17 deaths as of july. they weren't really gadgetbahns (grifty/pointless transportation "innovations") unti alphabet's own waymo fused one dubious idea (self-driving cars) with arguably the first gadgetbahn ever (taxis). in 2017 they released a plague of self-driving taxis on unsuspecting US streets in 2017. waymo currently "drives" in greater phoenix (whose transit authority, SEPTA, will operate at a 240 million dollar deficit in 2024) and san francisco (where extreme de facto segregation means even decent public transit services like BART and muni functionally only serve the wealthy). and while waymo's cars are not teslas, they still aren't as crashproof as their marketing suggests. in june, a waymo vehicle killed a dog.
naturally some tech entrepreneur has been looking for a third hot luxury tech trend to combine with this disaster to balance the whole thing out somehow. e-hang, a guangzhou-based corporation, has decided that the thing this hot mess needs is drones. they want to put human beings in unmanned aerial vehicles and fly them through the airspace over city centers--what chinese government types call its "low-altitude economy". and wouldn't you know it, the shit is actually selling.
the EH216 is technically an eVTOL (electric vertical takeoff and landing) vehicle, designed for "short trips" (its battery maxes out at 19 miles). it reaches a top speed of 62 mph/100kph. in october the chinese aviation authority officially declared the 216--and its two-seater version, the EH216s--skyworthy, though only test voyages have been completed so far. the company has already won contracts to sell 100 EH216 series units to the UAE and 50 to shenzhen for "aerial tourism". they also convinced to the catalonian government to build a vertiport at lleida-alguaire international airport in spain.
the impossibility of safely regulating the airspace over high-density cities aside, i am cautiously optimistic about society's ability to forcibly reject this new autonomous development firmly. waymo's self driving vans and cars are targets of constant vandalism in arizona, in particular because they seem categorically incapable of navigating a standard T-intersection. waymo and cruise (a general electric subsidiary) are among the driverless cars that routinely attempt to pull over at the first sign of fog--in san francisco. three days ago, someone torched an unoccupied waymo car with a firework in downtown sf. in the same city the activist group safe street rebels, have been systematically disabling autonomous vehicles since 2022 by "creating unicorns" (gluing orange cones to their hoods so they freak out and refuse to move).
if drone taxis ever fly in the US, i'd predict people will respond to them the way they already have to waymo vans and cars, and to drones, delivery robots, and other technological symbols of extreme wealth and need inequality--techno-luxury products that wander too close to a public that instantly recognizes what they are. as in, they will attempt to shoot them down by any means available.
a few years ago, "opting out" of participation in big data experiments didn't sound like a joke. the biggest tech companies at least made overtures toward ethical harvesting, signing vague letters to the FTC and even putting together short-lived ethics boards they could conveniently disband once the bad headlines fizzled out. GDPR proved that state authorities can at least mandate that big tech be transparent in its business practices--and that tech sold internationally has to follow the law of the strictest regulator in its market.
recently, though, some of the most powerful players in consumer data have all but dropped the ethical charade. google in particular seems to have abandoned all pretense of giving consumers choices about what they share and how. surrendering your data to ad tracking is now the inescapable price of using google search, built into google chrome as well as chrome-skinned browsers like brave and chromium. i know i'm not the only one seething about this development, however long it has been hailed by tech writers as the inevitable endgame of google's fucked up business practices. but there's good news--at least if you're as petty as i am. if you can't avoid giving your data to companies that want to use it to sell you more shit, there is one other choice you can actively make. you can decide to turn in data that is so bad, so spectacularly piss-poor, that it not only stops the tracker from predicting your consumer preferences--it actively damages the models the company is training for that purpose.
i first learned about data poisoning decades before it was named. feed by m.t. andersen (2002) was the first sci fi horror i ever consumed. it's a cyberpunk dystopia where teens grow up with an implant that gives them unceasing access to the feed--essentially a proto-social-media newsfeed--that is fully integrated into their conscious brains. getting the implant is technically optional, though the decision is usually made by parents on behalf of their infant children as to whether they want their kid to grow up connected to their peers and culture or in effective social isolation. one teen, struggling to maintain her identity after getting the implant much later in life than normal, discovers she can bamboozle the feed's constant appraisals of her needs and desires by buying things she hates and fantasizing about futures she doesn't want--at least temporarily.
data poisoning in this reality looks only slightly different. the types of data companies collect have diversified to include biometric traits like voices and faces, as well as individual styles in visual art, music, and writing. the goals of collection are also multiplying, as applications in surveillance, security, medicine, and basic data research are sold to wider markets. poisoning modern data sets takes more than clicking on ads you don't like or wearing anti-surveillance makeup: it takes dedicated tools that multiply the impact of the poison to match the scale of the invasive harvesting.
one example that would have been hard to explain in 2002 is the ubiquity of image-based AI engines trawling the web for visual data to chew up into free "art". diffusion models like stableDiffusion and midJourney use huge, illegally-scraped datasets to generate "new" art, stealing from human artists' creativity and labor while multiplying the environmental footprint of "art" production a thousandfold. glaze, a tool developed by researchers at U.Chicago, combats this by adding "a new dimension" to the art that is invisible to humans but which confuses AI scrapers so much that it hampers the model's ability to keep scraping. the program suffuses each pixel with nonsense data that misleads the engines trying to parse the image, without changing how it looks on a computer screen. the alteration protects art from copying, but the researchers caution that it's not future-proof: there is always the chance that someone will build an algorithm that can easily undo glaze's masking.
in their 2019 article, "the fantasy of opting out", media researchers helen nissenbaum and finn brunton caution that "There is no simple solution to the problem of privacy, because privacy itself is a solution to societal challenges that are in constant flux." presently, the dominant challenges to average internet users' privacy is companies like google using their market dominance to inflict exploitative data policies on the world.
one of the hallmarks of minority belief systems is the way the same object can have different meanings depending on whether it is viewed from a standpoint within a certain group or from without. for scientologists, who have been using devices they call "e-readers" as part of their strenuous interview process (what they call "auditing"), an e-meter is a sacred artifact. e-meter stands for electropsychometer, a tool that makes exactly as much sense as its beleaguered name implies.
in physical terms, the e-reader does little more than reasure the electrical resistance of human skin; however, due to their small-batch artisanal origins and unique connection to the CoS, e-meters come with hefty price tags as well as the hallmark of well-made proprietary technologies, the EULA. end-user license agreements are legal contracts between technologists and users. these days, they are rarely seen outside of the terms and conditions of a software application. so why do e-meters come with these contracts?
i'm autistic. this is both an old story and a recent and ongoing development. the old story because i've probably been showing obvious signs of my neurotype since before i could speak. recent and ongoing, because i was diagnosed at the ripe old age of 31, and i'm still figuring out what that means for me.
eventually i'll be writing about this stuff too but i also want to share other works that resonate with my experience. today that's gonna be the youtube channel of the wondrous yo samdy sam, a late-diagnosed autistic woman with ADHD who also happens to be a scientist and science communicator of the rawest caliber. in particular, this video from 2023 does an amazing job exploring the woefully under-studied (but much discussed) overlap of autistic people and people with adhd (i like 'AuDHD'). once i learn how to embed inline media in my posts, it's over for you hoes, but for now you can watch it on her channel.
in 2019, popular mechanics asked me to write a 'decade in review' article about how AI evolved during the 2010s. the resulting piece was more of an explainer of AI buzzwords that were relatively new to me at the time than a glimpse into my personal opinions on any of these technologies. over the past four years, while the technologies behind automated bias and mega-dataset processing have only changed a little, the role of AI in human life is coming into sharper focus. and since this is a blog post, i can just come out and say it: AI fuckin' sucks, guys.
when journalists become 'established' on a beat, strangers on the internet start sending story ideas out of the blue. sometimes the idea is tangentially related to something that interested you anyway and it all comes together in glorious harmony, like when a random PR email about IoT-enabled buoys dovetails with a mini-obsession on my part with FCC airwave regulation to illustrate the increasingly crowded radio space in north america. most of the time, though, the pitches tech journalists get are chewed-up corporate drivel. interviews are always proposed with the CEO or founder--a sure sign that the executive, not a media expert, is the one designing the pitch. the very last person you want to ask about why a technology matters is the person who believes the right choice of words will make them the next tech billionaire. when the product in question is even slightly AI-adjacent, all the above gets worse: the executive's expectations for publicity and profit are even higher, while their understanding of how their own product actually works tends to gradually approach zero.
given these circumstances, it may not be surprising that i've never even followed up on an unsolicited AI pitch, let alone turned one into an actual story. in fact, the pitches i have received on the subject regularly remind me that we're still in the phase of AI evolution where anyone with enough money can pay a programmer to slap an algorithm on an ill-gotten dataset and sell subscriptions to whatever it churns out. while i'm probably never going to write an *official* article about all these garbage ideas and why they suck, it occurred to me recently that it might be worth explaining why that will never happen, even if only to myself.
so, as much out of spite as intellectual curiosity, i'm compiling a list of the worst AI-related business ideas I've heard over the past four years. some of them were pitched to me directly; others came from big PR newsletters, credulous tech rags, and tumblr. all of them elicit emotions in me that could not possibly have existed in a time before AI, so thanks for that. please enjoy these horrible, free ideas, and please pray with me to robot satan that not one of them ever successfully disrupts shit.
the EKG-headband-chatbot-therapist. finally, a technology that replaces expensive, human-error-riddled concepts like 'therapy' and 'empathy' with highly monetizable chatGPT. this one is only harmless as long as people see through the scam. but mental health care is expensive and inaccessible, so i worry about a future where garbage chat apps get hooked up to mental health surveillance channels to further pathologize and criminalize addiction and mental illness. large language models (LLMs) like the one that powers chatGPT also come with fun built-in vulnerabilities that are only starting to come to light. attackers can jailbreak the algorithm to mess with the bot's instructions--for example, by overriding safety features that keep it from encouraging self-harm, or by instructing the bot to imitate an abusive person (or to demand credit card information).
AI-generated foraging guides written by nonexistent authors with nonexistent qualifications. AI has been "writing" books for half a decade already, but i personally didn't understand how dangerous that could be until i heard about this particular genre. because if there's one thing you want to trust a large language model to explain to you, it's the difference between toxic and edible mushrooms. while i haven't personally heard of anyone dying because they thought the foraging guide they bought had been fact-checked by a human being, it seems like that's the only possible outcome of selling this mush as a cheaper alternative to genuine expertise. in the words of Alexis Nikole Nelson (@blackforager), "i'm just worried that someone who doesn't know any better is going to poison themselves."
AI business books. books written by supposed AI business experts sell AI itself as an "exponential revenue driver". while many clearly buy into their own hype, my impression is that some of these dorks think they're only exploiting other executives, which is a noble pursuit. the problem is that we live on a planet suffering under the crushing "growth" of unrestrained capitalism. books like these help executives continue to delude themselves that data is immaterial, which means AI is magic that makes money out of theoretical numbers without hurting anyone. just like these authors, the execs and business hopefuls who buy these books see the economy itself as a bottomless font of profit and personal glory. that makes it a lot easier for everyone invovled to cordon off their expertise around the "business side" of AI, relegating human and environmental costs to footnotes in their own journeys to success.
AI assembly lines. an AI-first company is one that relies on an algorithm, but also, inevitably, on anonymous, undercompensated human labor. Mary L. Gray and Siddharth Suri's Ghost Work from 2019 explored the growing ranks of AI laborers--the humans workers who label datasets, flag mistakes, and otherwise provide common-sense checks to the AI's overconfident statistical analysis. they found that the field is much bigger and more diverse than they'd expected--and that worker protections are virtually nonexistent.
bias-multiplying surveillance and security systems. this thinking removes accountability from the human chain of command and creating technological deniability for violent treatment of unfairly feared groups. products like these are typically sold to business owners and cops--more of that good old-fashioned consolidation of power.
minority language control. linguistics has a long history of chewing up Indigenous and other oppressed groups' languages for the sake of intellectual exercise. AI hasn't changed that so much as made the process faster, less labor-intensive, and potentially much more profitable. openAI's translation app, whisper, was trained on thousands of hours of audio footage scraped from the web, including over a thousand hours of Maori language--all without endorsement, let alone input, from actual Maori people. the result of the scraping and analysis--a mediocre translation app--is hardly as important as the reaction of the speaker population. Maori ethicist and academic Karaitaiana Taiuru told Eco-Business, "Data is like our land and natural resources...If Indigenous peoples don't have sovereignty of their own data, they will simply be re-colonised in this information society."
machine learning drones. as if enabling remote murder weren't depressing enough, some drones now have AI-enabled navigation systems. these weapons learn from new environments and even make decisions in the air--where, of course, no one can physically stop them from carrying out unsupervised statistical robot justice. this is the most extreme example i can think of that demonstrates how AI can be used to create artifical boundaries between the human decision-maker and the humans and environment impacted by that decision. when WWII generals wanted bombs deployed, they at least had to hand off immediate culpability to the human pilots who carried out those orders. AI drones, and whatever other autonomous self-teaching weapons get churned out in the future, are the lethal endpoint of bias automation so far.