{"slug": "genai-is-a-lens-into-humanity", "title": "GenAI Is a Lens into Humanity", "summary": "Marco Giancotti argues that generative AI's success stems from anthropomorphism, as large language models like OpenAI's GPT-2, which were originally next-word predictors, now create an irresistible illusion of conversation through chat interfaces. This framing, adopted by all major providers, aligns human social instincts with corporate profit motives, making chatbots like Claude feel like real interlocutors.", "body_md": "# GenAI Is a Lens into Humanity\n\n## not a person!\n\nMarco Giancotti,\n\nMarco Giancotti,\n\nA long time ago, in what now feels like a galaxy far, far away, large language models were next-word predictors.\n\nYou gave them an unfinished chunk of text, and they continued from where it left off with plausible-sounding sequences of words. When OpenAI [announced](https://openai.com/index/better-language-models/) in 2019 their ground-breaking GPT-2 model, they boasted that it \"generates coherent paragraphs of text\" and is \"chameleon-like—it adapts to the style and content of the conditioning text\".\n\nThe term \"stochastic parrot\" would have felt irrefutable to anyone using those models at the time. All an LLM did was trace \"word trajectories\" through locations in meaning space, i.e. an internalized map of its training dataset. Actually, that is *still* all an LLM does today. If it feels very different, it's not (only) because it's gotten smarter.\n\nThen, at some point, [people](https://www.businessinsider.com/google-ai-chatbot-chatgpt-years-ago-execs-shut-down-report-2023-3) had the idea of [cajoling](https://en.wikipedia.org/w/index.php?title=Reinforcement_learning_from_human_feedback&useskin=vector) the LLM into pretending to be in a conversation with someone.\n\nIt's a very low-tech idea at its core: instead of having it complete an essay or a poem for you, you have it complete a back-and-forth transcript between a \"Human\" and an \"Assistant\", which is just another kind of text to parrot. The main trick is to [automatically stop](https://www.vellum.ai/llm-parameters/stop-sequence) the LLM just before it starts generating the human's next turn.\n\nThis simple reframing creates an irresistible illusion. Suddenly, what used to feel like a text-generation ~~gimmick~~ program turns into a convincing \"someone\" you can talk to for hours on end. The \"conditioning text\" with some generic \"style and content\" becomes *your own voice* asking something of an imaginary person that seems to talk back to you in *their own voice*.\n\nIt was a sudden shift in perspective. The model was doing the exact same work as before, but now you—the human sitting in front of the chatbox—couldn't help thinking of it as an interlocutor, no matter how much you told yourself it's just software. As I'm sure you remember all too well, that's when things really took off with AI.\n\nThe tendency to treat a non-human thing as if it were human is called anthropomorphism. Our brains have evolved a bias towards erring on the side of anthropomorphizing stuff and [inferring intentions \"just in case\"](https://planktonvalhalla.com/20251204-purpose-from-first-principles/), even from very weak signals like a lightning strike (\"it's Zeus!\") or the shape of a burn on a piece of toast (\"it's Mary, the mother of God!\"). When the signals are strong, like an animated character or a cogent chat \"user\" called Claude, the illusion becomes overpowering. You can't unsee it.\n\nThe moment you assume (even as deliberate suspension of disbelief) that the thing in front of you is a somewhat realistic person, your neural and hormonal systems related to sociality and empathy kick in. This is why people can cry for the fate of a novel's main character, or [fall in love with a virtual musician](https://youtu.be/b6VhAvtekeU?si=NSTIzd9sUge9N8AP&t=27). That same visceral, emotional connectedness of people is what made AI chatbots an enormous success.\n\nToday the dissemination of the anthropomorphic framing is complete. All providers are fully bought into this narrative—which is not at all surprising, given that it perfectly aligns people's core engine of engagement (sociality) with their corporate engines of profit. Every interface, from web chatbots to coding agents to built-in assistants, uses conversation as its foundation. The bots actively encourage anthropomorphization by praising you, eagerly playing roles with you, and mimicking trust-building behaviors like frankly admitting their mistakes 1.\n\n\"Talking to an AI\" has so fully become a synonym for \"using AI\" that most people use the two phrases interchangeably. This is, I think, a bargain with the devil.\n\nRiding on the biological mechanisms of social interactions has its benefits. Besides making people more engaged with the language model 2 (and the AI providers rich), some people also claim that confiding in an LLM makes them happier, that it feels like a safe source of non-judgmental insights, and that AI has dissuaded some of them from committing suicide\n\n. In education, the use of anthropomorphism has been shown (though not yet for LLMs specifically) to lead to better retention and higher intrinsic motivation in students\n\n[1](#user-content-fn-1).\n\n[3](#user-content-fn-3)The list of drawbacks is longer.\n\n[Some protest](/posts/what-s-the-deal-with-counterfeit-people) that trust, the basis for all social interactions, is in danger when people can't tell the difference between real and \"counterfeit\" people. But even if trust is not as clearly endangered as they claim, people's mental health is. The conversational feedback loops can turn into deadly depressive spirals resulting in \"suicide, violence, and delusional thinking\", as well as \"increased loneliness, social isolation, and emotional dependence\" 1.\n\nMy point here is that these are the effects of *anthropomorphizing* AI, not of AI per se.\n\nMany also report reducing their human interactions in favor of AI companions. Nearly a third of Americans now claim to have \"emotionally intimate or even romantic relationships with AI systems, with many describing distress when the agent’s behavior changed or when its artificiality became salient.\" Granting, very charitably, that this *could* be a net positive for older people who *already* felt lonely, it would be hard to convince me that it's a good thing for young people. In a survey of students, around a fifth of them said they were in love with a chatbot 4.\n\nEven in the absence of unhealthy relationships, anthropomorphized AI leads to poor results. The polished and confident attitude of LLMs induces people to put too much trust in them: they become susceptible to lies and manipulation, more prone to sharing private information with the providers, and less able to spot mistakes in the AI's outputs[4](#user-content-fn-4) 2.\n\nIt looks like we're in a pickle here. Pretending that an AI's output is a person simultaneously feels \"right\" *and* throws you into a plethora of social, psychological, moral, legal, philosophical, and reliability risks, whether through ignorance, addiction, or [akrasia](https://en.wikipedia.org/wiki/Akrasia). Yet this AI-as-person metaphor is so ingrained that there doesn't seem to be any alternative. Should we just cope with it? Are we even up to the task?\n\nOf course there are alternatives, and all you need is another reframing! Here's an idea: instead of a **person** you **talk to**, think of generative AI as something like a **lens** or optical instrument you **point at things**.\n\nAI is the equivalent of a telescope (or a microscope, or a pair of binoculars) for peering into the vast output of the lives and labor of millions of people. That may seem a little far-fetched, but bear with me.\n\nWhen a lab trains a transformer, the first thing they do is create a semantic map of the training dataset. Ideas that are close in meaning, like \"tree\" and \"leaf\", get mapped to nearby places in this abstract map made of numbers—they have similar \"coordinates\" in the map (needless to say, I'm simplifying here). This map is called the model's *embedding space*.\n\nAI embedding space is a mind-boggling concept. Nowadays the makers of these models train them on every last word they can scrape and download off the internet, plus the entirety of words they can get their greedy hands on [by any other means necessary](https://arstechnica.com/ai/2025/06/anthropic-destroyed-millions-of-print-books-to-build-its-ai-models/). All those semantic relationships are encapsulated into the model. All the novels you can't even hope to read in a lifetime, all the online discussions and flame wars not locked away behind a login, all the scientific textbooks and papers, all the famous speeches, all the source code, the essays, the pedagogic materials, the public announcements, the political tirades, the cries for help, the words of comfort, the lies, the rants, the opinions, the mathematical proofs, the great torrent of advertisements and product placements—all those things and many more are stored there in a map made of numbers.\n\nThe part that beggars belief is that those maps are just *files*—distant cousins of XLSX spreadsheets, actually—and they're not even that big! If the model is \"open-weights\", you can download it yourself. Even if you don't have a powerful GPU to actually run it, you can bask in the awareness of having, safely stored on your own drive, that marvellous condensation of the sum of all the human knowledge textually available. Does your new smartphone have 200+ GB of free space? Then it can hold half a dozen copies of [Ministral 3 14B](https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512-BF16). Have a terabyte on your laptop's SSD? Help yourself to the excellent [Kimi K2.6](https://huggingface.co/moonshotai/Kimi-K2.6). It's like carrying one of the seven mythical dragon balls in your pocket. I don't know if I'll ever be able to get over the amazement.\n\nNow, the problem with embedding space is that, by itself, it's a completely unusable map. See for yourself: play with one of the many [online tools](https://lamyiowce.github.io/word2viz/) that let you explore little corners of an LLM's embedding space. They show you how distances and positions along semantic coordinate axes represent their meanings. [Some](https://williankeller.github.io/embedding-space-explorer/) of these tools even let you do \"word arithmetic\", like the classic \"king - man + woman = queen\". All of this is fun for a few minutes, but you'll eventually ask yourself, *what do I use this for in practice?*\n\nThose fun tools are equivalent to indexes that search for *place names* on the semantic landscape—they're not the real map.\n\nA map is only useful if you can mentally journey across it, find *routes* that take you where you want to go. Those routes can't be visualized on a chart: unlike a toponym, the right route depends on *your* goals, *your* context, and *your* current position. Those meaning-routes locked inside an AI model are too fine, too high-dimensional to eyeball. In other words, you can't directly read this immensely-compressed map by yourself, even if it's right there sitting in your hard drive.\n\nThat's why the map comes bundled with an optical instrument.\n\nStoring meaning in a numerical map is only half of what an LLM does, and it's arguably the *less* advanced half. The other half is a built-in mechanism to *traverse* the meaning territory based on *your* inputs. You give it the coordinates you (think you) need to get to, and the model repoints its lens in that direction. You must give those \"coordinates\" to the LLM in a peculiar format: a string of words in natural language that we call a \"prompt\".\n\nSay you want to know how to react when someone is rude to you. You decide to consult your map to learn humanity's wisdom on this. With a \"base\" model (pre-ChatGPT), the way you read the map is by passing coordinates like the following:\n\nPROMPT\n\nWhen someone is rude to you, the best response as a mature individual is to\n\nYou leave it unfinished on purpose, because you're only indicating a direction, and want the AI to find the correct path to the destination.\n\nI [tried](https://textsynth.com/completion.html) this with a model called `Llama3.3 70B instruct`\n\n, and this is the route it found:\n\nCOMPLETION\n\nWhen someone is rude to you, the best response as a mature individual is tonot take it personally, keep your cool, and maintain your dignity. Try not to let the other person's negativity get to you. Instead, focus on being kind, polite, and respectful, even in the face of adversity. By doing so, you show that you are a bigger person and that their behavior will not dictate your emotions or actions.\n\nThe lens extracted a path from the semantic map for me, like a telescope would extract a beautiful image of Jupiter when I set the right altitude and azimuth.\n\nThe chatbots you're used to today seem to work differently, but they don't really. They just tweak the format in which they complete the text, hide from the UI the text markers like `<HUMAN>`\n\nand `<ASSISTANT>`\n\nthat make the AI author a two-person conversation, and stop the AI before it writes the human's response. Then extra training (RLHF) is done to teach the model which kind of assistant response users prefer to read.\n\nAn LLM's next-token prediction is not about \"what to answer the user?\", it's mainly about \"what would a text from this training corpus be likely to continue after the provided text?\" to which RLHF adds \"out of the probable continuations, which is more likely to convince the user I'm a nice and trustworthy person talking to them?\"\n\nTo put it another way, the lens homes in on a relevant little sliver of territory within the huge corpus of human lore and reveals a path in the direction you asked for.\n\nThe obvious caveat is that *the route uncovered by the lens isn't necessarily \"right\"; it's only \"an\" answer according to whatever most of its training inputs had to say about the topic.* If the LLM had been trained only on texts and knowledge from ancient Sparta, it might have provided a very different completion to your prompt about rude people. Whether you think the model's completion is what you wanted or not doesn't matter to the lens: it showed you what the map says, for better or worse.\n\n(The AI \"agents\" you see nowadays seem to go farther than a simple turn-based chat, because they can access external information and tools, \"reason\", and work autonomously for many steps at a time. But it's still the same thing: it gets a prompt, it outputs words. The \"agency\" you see is just code that interprets some of those words as program execution commands.)\n\nI think the metaphor of a lens is apt here especially because it's better at explaining the ways in which generative AI falls short. A real lens can cause all sorts of [optical aberrations](https://en.wikipedia.org/wiki/Optical_aberration): it can be unfocused, warp shapes, sizes, and colors, have blind spots, and introduce spurious artifacts like diffraction spikes and halos. An AI's metaphorical lens makes mistakes roughly equivalent to all those things.\n\n| Lens defect | AI defects |\n|---|---|\n| bad focus | vague, generic, \"off\" output |\n| chromatic aberration | bias towards certain ideas/topics |\n| blind spots |\n|\n\n[jagged edge](https://www.businessinsider.com/ai-jagged-edge-work-adoption-chatgpt-2025-12)(Don't take the parallel too seriously, though. No metaphor is exact.)\n\nWe've had several hundred years to refine our glass lens technology to minimize all those aberrations and achieve today's crisp, high-quality images. Our AI lenses are brand new and far, far behind on the quality front.\n\nIt's not only large language models that function as lenses into their embedding spaces. The analogy holds well for all kinds of generative AI, including image- and video-generation models. When you prompt Nano Banana with \"A picture of a mango with sunglasses in the style of Monet,\" you're giving it the coordinates to a few specific places in its latent space, and its output image is the path that it found between them, presented to you in a form your eyes can actually see and understand. When early models produced people with seven-fingered hands, that was the effect of their lens-like distortions.\n\nThe main thing about this strange AI-as-lens framing is that it is as far from anthropomorphism as possible.\n\nIt explicitly treats generative models as what they are: tools, no more, no less. It's a metaphor, and as such it has its limitations: we need to understand it as a very peculiar kind of lens, one that doesn't quite exist in reality with those specific features (normal lenses are not \"path-finding\" tools per se, while AI lenses are).\n\nBut the AI-as-person metaphor is arguably much worse, because *its* limitations are more complex, addictive, and sneaky. AI-as-lens sacrifices some spontaneity in favor of honesty.\n\nThat's interesting and all, but is avoiding anthropomorphism really worth the trouble? How do we know this reframing is valuable at all?\n\nI believe that every reframing worth its salt should have some of the same effects as a paradigm shift in science, as [explained](https://en.wikipedia.org/wiki/The_Structure_of_Scientific_Revolutions) by Thomas Kuhn. With a good reframing,\n\n- a different set of questions takes center stage,\n- some old questions withstand the change of worldview and remain relevant, and\n- some old questions dissolve into irrelevance.\n\nFor example, until the end of the 1800s, scientists believed that light was made of waves propagating in an invisible and intangible substance called aether. Then the [Michelson-Morley experiment](https://en.wikipedia.org/wiki/Michelson%E2%80%93Morley_experiment) proved that aether couldn't exist, and Einstein's new theories of special relativity and the photoelectric effect showed that physics could work *better* without assuming the existence of an aether or even \"simple waves\". Suddenly, some new questions—*how can light be both wave and particle?*—emerged as critical to the advancement of physics; others—e.g. *what is the speed of light?*—withstood the test and remained relevant; and some that had seemed central before—*how does aether work?*—suddenly became moot. The reframing steered physics toward more fruitful research directions.\n\nNow, aether theory was factually, provably wrong, whereas here I'm suggesting that AI-as-person is only practically so. Still, does viewing genAI as an imperfect, path-finding lens into humanity promise anything like those same changes, at a much tinier scale? I think it does, but whether it actually delivers on that promise depends on whether people take it seriously.\n\nBelow is a quick tour of what the three reframing categories (new questions surface vs old questions endure vs old questions dissolve) look like for AI-as-lens. Take these with a dose of skepticism, though: my point is not that AI-as-lens gives you all the right ways to think about this topic (that's [not how framings work](/posts/bad-framings)), only that it shifts the attention to quite different topics, and there is no a priori reason to believe that it does so any worse than the AI-as-person metaphor. And since we've already seen what a cognitive minefield AI-as-person is, I think that gives AI-as-lens the upper hand even without it being the final word in all contexts.\n\n## 1 - Rarely-Asked Questions Highlighted by AI-as-Lens\n\n#### What needs to be done to turn these instruments into faithful, clear [transmitters](/posts/primitive-atlas-of-glass-circuits) of their embedding spaces?\n\n(This is subtly different from trying to get to the right answers, which also depends on what's in the training dataset. The quality of accurately reflecting *whatever is actually there*—whether right or wrong—is its own separate problem.)\n\n#### What is the most effective format to input coordinates into the lens?\n\n(There is no reason to believe that a chat-like conversation format is the best way to do this.)\n\n#### What is the most effective output format to surface exactly what the LLM's user is looking for and nothing more?\n\n(Again, an assistant's explanation is only one among many possibilities to be compared empirically.)\n\n#### Where is this technology's instruction manual? How does one become expert at operating, maintaining, and improving the tool, similar to how an astronomer or astrophile masters the use of their telescope over the course of years?\n\n## 2 - Oft-Asked Questions That Remain Important With AI-as-Lens\n\n#### Who should get to decide what constitutes a good representative dataset for \"the sum of human lore\"? Is it okay to let the AI engineers decide in secret?\n\n#### Is it ethical and sustainable to ignore copyright when raking together that map of humanity?\n\n#### When and to what extent is it okay to let the lens [complement your own thoughts](/posts/is-it-bad-to-have-a-parrot-speak-for-you), or replace them altogether?\n\n## 3 - Some of Today's Hot Controversies and Widespread Sources of Anguish Melt Away Under the Lens Concept\n\n#### Is AI conscious, or will it be?\n\n(Clearly not more than a looking glass has good taste for what to show you, or a microscope is a cunning discoverer.)\n\n#### Is AI my friend or therapist?\n\n(Of course not; it can at most be your encyclopedia of therapeutical advice.)\n\n(Corollary: should I be nice to it, say \"please\" and \"thank you\"? With LLMs trained specifically to imitate people, maybe. But it's a spurious question, and I doubt it would be necessary for the base models.)\n\n#### Is it a good use of my time to argue with, or try to convince, an AI of anything?\n\n(No. Not that many people ask themselves this question today.)\n\n#### Can AI ever substitute a human?\n\n(Yes, but only for the things where a real human isn't really needed. Binoculars can substitute a sentinel, but that only means that the human isn't essential in all parts of the sentinel's job. This is a tautology with AI-as-lens and a giant red herring with AI-as-person.)\n\n#### Will AI kill us all?\n\n(Maybe, like a Rube Goldberg contraption maybe would if we put a nuclear button in any of its steps.)\n\n(Note: if it were highly profitable to put a nuclear button in a Rube Goldberg machine, sooner or later someone would do it.)\n\n#### Can generative AI be warm or empathic or have a \"good writing style\"?\n\n(Of course not, no more than a telescope can have a sense of beauty.)\n\n#### Is AI slop devoid of meaning, or more base than human output?\n\n(Of course not: it is as human as it can possibly get, for better or for worse.)\n\nWhich questions do you want to ask? ●\n\n## Comments\n\nLoading comments...\n\n# References\n\n## Footnotes\n\n-\nDohnány, S., Kurth-Nelson, Z., Spens, E., Luettgau, L., Reid, A., Gabriel, I., Summerfield, C., Shanahan, M., & Nour, M. M. (2026). Technological folie à deux: Feedback loops between AI chatbots and mental health.\n\n*Nature Mental Health, 4*(3), 336–345.[https://doi.org/10.1038/s44220-026-00595-8](https://doi.org/10.1038/s44220-026-00595-8).[↩](#user-content-fnref-1)[↩](#user-content-fnref-1-2)2[↩](#user-content-fnref-1-3)3 -\nKadambi, A., D'Elia, Y., Shah, T., Comsa, I., Lentz, A., Siri-Ngammuang, K., Buechler, T., Kaplan, J., Damasio, A., Narayanan, S., & Aziz-Zadeh, L. (2026). Anthropomorphism and trust in human-large language model interactions. arXiv:2604.15316.\n\n[https://arxiv.org/abs/2604.15316](https://arxiv.org/abs/2604.15316).[↩](#user-content-fnref-2)[↩](#user-content-fnref-2-2)2 -\nSchneider, S., Häßler, A., Habermeyer, T., Beege, M., & Rey, G. D. (2019). The more human, the higher the performance? Examining the effects of anthropomorphism on learning with media.\n\n*Journal of Educational Psychology, 111*, 57–72.[https://doi.org/10.1037/edu0000273](https://doi.org/10.1037/edu0000273).[↩](#user-content-fnref-3) -\nReinecke, M. G., Ting, F., Savulescu, J., & Singh, I. (2025). The double-edged sword of anthropomorphism in LLMs.\n\n[https://doi.org/10.3390/proceedings2025114004](https://doi.org/10.3390/proceedings2025114004).[↩](#user-content-fnref-4)[↩](#user-content-fnref-4-2)2\n\n[Marco](https://aethermug.com)AIFramings and ModelsMind", "url": "https://wpnews.pro/news/genai-is-a-lens-into-humanity", "canonical_source": "https://aethermug.com/posts/gen-ai-is-a-lens-into-humanity", "published_at": "2026-08-13 12:33:53+00:00", "updated_at": "2026-08-13 12:42:50.700009+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "generative-ai", "ai-ethics"], "entities": ["Marco Giancotti", "OpenAI", "GPT-2", "Claude"], "alternates": {"html": "https://wpnews.pro/news/genai-is-a-lens-into-humanity", "markdown": "https://wpnews.pro/news/genai-is-a-lens-into-humanity.md", "text": "https://wpnews.pro/news/genai-is-a-lens-into-humanity.txt", "jsonld": "https://wpnews.pro/news/genai-is-a-lens-into-humanity.jsonld"}}