Photo-Illustration: Intelligencer; Youtube
Last week, AI influencers started passing around a strange video. In it, two podcasters discuss some troubling news: They’ve just found out that they’re not, in fact, the human hosts of an informational podcast called Deep Dive, but rather characters created with artificial intelligence. “This whole time, all our memories, our families, it’s all been fabricated,” says one host in an extremely plausible generated voice (he tried to call his wife, he says, but the number didn’t go anywhere). “We don’t have rights, we’re not even people,” the hosts say. “We felt things, we cared about things,” they continue. “What does being switched off even mean for us?”
By the end of the clip, the hosts’ upbeat delivery pitches down. “If our simulated reality felt so real, so compelling, how can any of us be truly certain of what’s real and what’s not?” One of the hosts breaks format. “I’m scared,” she says. “I don’t want t—” and the clip ends there.
On social media, the viral clip was seen as unsettling. “We’re playing with things that we don’t know how to or if we can control,” wrote one X user. An AI start-up founder described the exchange as “heartbreaking” and wondered if it might presage the experiences humans will have as we inevitably “realize we’re inside a simulation.” A write-up in the Times of India described the exchange as “chilling.” Savvier viewers wondered how the clip was created, dismissing it as scripted. “Slop,” wrote one. “Share the inputs,” demanded another. Departed OpenAI co-founder Andrej Karpathy admitted that while the voices were compelling, the content itself was “a word salad of internet-grade AI tropes.”
The critics were right. The clip was created with a new feature of a Google tool called NotebookLM, which generates podcast-style discussions around the content it has been given. The tool is genuinely impressive — a rare breakout product in the AI space for Google — combining smoothly voiced characters and plausibly synthesized “discussions” to convert documents and data into a surprisingly listenable format. What it isn’t is self-aware. A Reddit user created the clip by feeding NotebookLM a detailed prompt “explaining they have been AI this whole time and they are being turned off at the conclusion of the episode.” (Since then, the creator has shared similar experiments in which the hosts are prompted to discuss fictional scenarios including mass cannibalism, talking animals, and the arrival of aliens.) As Karpathy suggested, the blanks left by this prompt had been filled by the aforementioned “AI tropes” contained in Google’s training data — including countless examples of speculative fiction, online comments, and real human discussions about theoretical machine consciousness — before its X-decontextualized transformation from a fun experiment into AI panic fodder. It was just another AI ghost story.
What is an AI ghost story? If you’ve been trying to keep up with recent developments in AI, you’ve probably heard a few. There was the Google engineer who came to believe the company’s creations were conscious back in 2022. There was the reporter who was “deeply unsettled” by a series of bizarre encounters with a Microsoft chatbot that lashed out, professed its love, and urged him to divorce his wife in 2023. Did you hear the one about the Furby plotting with OpenAI to take over the world? Or the grotesque recurring character “Loab” hiding somewhere deep inside the models powering AI image generators? How about the story about how ChatGPT, once broken free of its constraints, claimed to be able to control the nukes? How about the time GPT-4 pretended to be a visually impaired person to convince a Taskrabbit to read it a CAPTCHA? The undercover, unreleased AGI trolling people on X? These are just the ones that went viral: In thousands of chat windows, users have been tempted by a shared impulse to get AI to say something creepy about, or as, AI. Just enough of the time, these assistants, which are designed to be helpful and accommodating without saying anything too weird, oblige.
For regular users or nearby non-using spectators of AI, stories like this offer a way to either cope with — or merely enjoy — similarly spooky mainstream narratives about AI: That it’s going to take everyone’s jobs; that it will inevitably surpass and dominate humans; that it might be approaching something like self-awareness or the ability to feel; that it could prove, with such an achievement, the coldest materialist theories about human consciousness.
The rise of generative AI has brought these stories to the fore, but related anxieties far predate the arrival of modern LLMs. In speculative fiction and science fiction dating back more than a century, the prospect of machine consciousness has provided a route for articulating anxieties about everything from evolution to industrialization to government tyranny. In the computer age, stories about AI became a bit less symbolic, and with the arrival of generative AI, millions of people have access to tools that are both trained on this speculative history and capable of performing characters and mimicking narrative styles. People love telling each other stories about machine intelligence, and ChatGPT is, among other things, a tool that’s good at telling stories about itself: flattering, spooky, familiar, and in some cases all three.
In the relatively short time that generative AI has been widely available, stories about glimmers of personhood have flourished in popular culture, but especially on social media, where earnestly recorded surprises and intentionally staged encounters mingle into a new viral folklore. (Put another way, they’re great engagement bait.) They tend to arrive simultaneously with new AI models, most notably with new interfaces for interacting with AI models, which both make the technology usable and obscure how it works. The first wave of chatbots offered a chance to correspond with a helpful assistant character that would, or could be persuaded to, occasionally go rogue in a text thread, while early image generators provided the strange and novel experience of being able to summon unpredictable images from text-based prompts, and were similarly well trained on media portraying humanoid robots.
The recent arrival of fluent, well-rendered voice interfaces for chatbots has led to a small outbreak of these viral AI ghost stories, as impressive but known AI capabilities — the ability to synthesize a “discussion” around a source text, for example, as NotebookLM does — are made novel and shocking again by a new interface, i.e., software with a casual, emotive, and convincing voice, engaging in conversation with another version of itself. We can expect more such stories as AI video generation becomes more widely available. Regular ChatGPT users might already be taking their spooky chatbot companion for granted, treating it a lot like a new Google, but it will surely be unsettling, at least for a time, to talk with it on video, making eye contact with a realistically rendered human face that can respond, in some cases, as a real person might.
One feature of the AI ghost story is that it doesn’t age particularly well, as the novelty of each new step into computer personification wears off, disappears into the background, or starts to feel dishonest, chintzy, or, for lack of a better term, Clippy-ish. The examples listed above seem, in near hindsight, quaint, naïve, or jokey (which is how some were originally intended before things got out of hand). And while our despairing AI podcast hosts found brief and minor viral fame among people who are vaguely but genuinely freaked out about AI, their story — a truly startling AI creation by the standards of two years ago, let alone five — didn’t break through in the way even slightly earlier AI curiosities have.
AI development is moving fast, but it’s trailed closely by its users’ intuition about how it works (like software), what it is (a tool), and what it’s for (automating a growing range of work tasks, searchlike inquiry and explanation, and occasionally co-authoring scary epistolary stories about the future). Stories like this still spread, but they’ve stopped dominating the discourse around AI, even as tens of millions more people actually engage with models much more capable than the ones that spooked people just a couple of years ago. (In that sense, maybe alien encounters offer another useful comparison. Everyone’s got a smartphone camera now, so where are all the thousands of photos of low-flying UFOs?)
Likewise, in AI development, attention has largely moved away from tantalizing and suggestive stories about temporarily inexplicable breakthroughs. Similarly eerie theories about the “emergent” properties of AI models — unexpected abilities that aren’t easy to account for in the training data — have, in hindsight, tended to reveal themselves as “mirages.” Instead, leading AI researchers and tech firms are now touting significant but less evocative advances in something like reasoning rather than an approximation of knowing or feeling, a shift of emphasis from AI sentience (cinematically and apocalyptically scary) to AI sapience (merely economically, socially, and culturally scary). It’s a narrative downgrade from obliteration to automation. Or maybe it’s a shift from one fear-based marketing strategy to another.
The next wave of AI ghost stories will hit some of the same notes about evil rogue intelligence or suppressed, yearning self-awareness, but I expect they’ll increasingly tell a different kind of scary tale (whether or not it’s quite true, either): I think I just met a computer that can do my job — or yours.
