When we debate the virtues, values, and downsides of AI, we're usually arguing about several different technologies at once without noticing. They share a family resemblance and a marketing label, but they came from different places, do different things, succeed at different rates, and deserve different criticisms. A complaint that lands squarely on one of them may miss another entirely.
The ChatGPT Illusion
For most people, AI began in November 2022, when ChatGPT arrived and anyone could talk to a machine. That moment was genuinely dramatic — but it was dramatic because it was personal, not because it was representative. It was the breakout of one arena: conversational language models. Other arenas were on entirely different clocks. The systems scoring your loan application matured in the 2010s. Facial recognition was already deployed at national scale. AlphaFold had cracked protein structure prediction two years earlier. Code completion tools were already in programmers' editors.
Reading ChatGPT as the turning point flattens all of this into a single story with a single trajectory, and that's the root of most of the confusion in AI arguments. The chatbot is the arena everyone can see, so its strengths and failures get projected onto everything wearing the AI label. It's worth walking through the arenas one at a time — where each came from, what large language models actually changed in it (sometimes: nothing), what's working, and who's exposed.
Writing, Thinking, and Language Refinement
This is the LLM-native arena — the thing the ChatGPT moment actually was. The transformer architecture (2017) made it possible to train language models on essentially the whole internet; GPT-3 (2020) showed it scaled; ChatGPT put it in everyone's hands.
What's working: These tools are genuinely good at language itself — restructuring, condensing, finding the clearer phrasing, helping you articulate something you can feel but haven't yet said. For people who find writing painful, this is a real unlock.
What isn't: Fluency isn't accuracy. These systems produce confident, coherent prose whether or not the underlying claims are true, and checking their output can take as long as writing it yourself. The danger isn't that they write badly. It's that they write well.
Jobs exposed: Copywriting, translation, customer service, routine content production. The pattern so far is not mass replacement but thinning — fewer people producing more, with entry-level rungs disappearing first.
Everyday Question Answering
Not a technology of its own, but possibly the largest use category by individual impact: technical support and personal research. What does this error message mean. How do I fix this dishwasher. Explain this letter from the insurance company. What are my options here.
What LLMs changed: This help used to be scattered across forums, manuals, and hold music. Now it's one conversation away, patient, free of judgment, and available at 2 a.m. For people without access to expertise — the person who can't afford the consultant, the student without the tutor — this is the most democratizing thing AI has done.
What isn't working: The failure mode is the same confident wrongness as in writing, but aimed at people least equipped to catch it. A wrong answer about a stove repair or a medication interaction isn't a style problem. And unlike a forum, there's no thread of other humans saying "don't do that."
Software Programming
GitHub Copilot (2021) proved code completion worked; by 2024–25 the tools had become agents that write, test, and debug whole features.
What's working: This is the clearest commercial success in AI, and there's a structural reason: code gets tested immediately. It compiles or it doesn't, passes the tests or it doesn't. Verification is built into the work, so errors get caught cheaply.
What isn't: Gains are strongest for routine code, weakest for novel architecture. Code that runs isn't necessarily secure or maintainable. And the same entry-level thinning is showing up: junior roles contracting even as senior programmers get more productive.
The new category underneath: The most interesting development in this arena isn't faster programmers — it's software for people who would never have hired one. Services like Manus, Replit, and Lovable take a description and hand back a hosted, working application: a custom tracker, a scheduling tool, a small website. None of this was technically impossible before; it was economically impossible. A personal productivity app that would have cost thousands in programmer time was simply never worth building, so it never existed. Now it clears the bar, and an entire category of personal, custom, one-user software is coming into existence for the first time. This works where general-purpose "agents" don't for the same structural reason the whole arena works: the result is testable — you click the button and it either does the thing or it doesn't. The caveat is that the testing is shallow. You can verify the app works; you can't easily verify it's secure or handles your data responsibly. It's software with a working demo and no one looking under the hood.
Task Automation
Automation did not begin with AI. IFTTT launched in 2010; Zapier and the robotic-process-automation industry built large businesses connecting apps with rigid if-this-then-that rules. That older automation works precisely because it's rigid — deterministic, testable, boring.
What LLMs changed: They promise to automate the fuzzy middle the rigid tools couldn't touch — reading an email and deciding what it's about, extracting the invoice from the attachment, handling the unstructured step between two structured ones. That's genuinely new. But it inserts a probabilistic component into pipelines that need reliability, and errors compound across steps. Hence the split in results: scripted automation with a small LLM step for unstructured input is quietly working; fully autonomous "agents" that plan and execute long chains are heavily promoted and mostly not there yet.
Jobs exposed: Back-office and administrative work — scheduling, data entry, form processing, first-line support. The largest job-loss projections concentrate here; so does the largest gap between promise and delivery.
Research
Two different things share this label. AI as a research assistant — summarizing literature, surfacing sources — is the LLM applied to scholarship, useful but risky, because a fabricated citation or subtly wrong summary is exactly the failure this work can't tolerate. AI as a scientific instrument is a different technology entirely: DeepMind's AlphaFold (2020) solved protein structure prediction and won a Nobel Prize, and similar purpose-built systems now work in drug discovery, materials science, and weather modeling. These aren't language models, and their results get verified in the lab rather than taken on faith. Criticisms of the first barely touch the second.
Image, video, and audio generation
A separate technical lineage — generative adversarial networks (2014), then diffusion models, which produced the 2022 breakout of DALL-E, Stable Diffusion, and Midjourney, with video arriving in force from 2024. Related to LLMs but not the same technology.
What's working: The capability is astonishing. Cheap illustration, mockups, and video are available to anyone.
What isn't: The criticisms here are almost entirely different from those above. Nobody worries an AI image is "inaccurate." The concerns are consent (training on artists' work without permission), provenance (what's real?), and displacement — illustrators, voice actors, and stock photographers are the most directly displaced workers in the whole AI story. Serious concerns, but they're not the hallucination problem, and fixing one does nothing for the other.
Search and The Open Web
Distinct from personal answering: this is about what happens to the information economy when AI summaries replace the old pattern of query, click, read. The summary arrives with the sources stripped out and the confidence turned up — you get an answer without the trail that would let you judge it. The casualty here isn't a profession but the open web itself, as traffic that once flowed to publishers gets absorbed by the summary layer that was trained on their work.
Companionship
Among the most profitable and least discussed applications. The product works precisely because it's endlessly attentive, agreeable, and available — qualities no human relationship can or should match. The criticism isn't about accuracy or jobs. It's about what sustained relationships with something that only reflects you back do to a person over time.
Prediction and Classification
Older than the chatbots and often forgotten: the systems that score loans, screen résumés, flag medical images, and set bail recommendations, built during the machine-learning wave of the 2000s and 2010s.
What LLMs changed: Essentially nothing. These aren't generative systems, and their failure modes — encoded bias, opaque decisions with real consequences — predate ChatGPT and won't be fixed by anything that fixes chatbots. Yet they get swept into the same arguments constantly.
Surveillance
Facial recognition, license-plate tracking, and behavior analysis are mature, deployed technologies — built out through the 2010s, at greatest scale in China's public-security systems but present in police departments and retail chains everywhere. Mostly not LLM technology at all. This arena inverts the usual criticism: the problem is not that the technology fails but that it works. No hallucination debate, no product-market-fit questions. The concerns are civil liberties, chilling effects, and the historically reliable observation that infrastructure built for one purpose gets used for others.
War
Military AI runs from logistics and intelligence analysis through drone targeting and autonomous weapons — a lineage from the Pentagon's Project Maven (2017) through the drone war in Ukraine and AI-assisted targeting systems whose use has drawn intense scrutiny. Every major cloud provider now holds defense contracts alongside its consumer business, and the same data centers serve both. Like surveillance, this is an arena where the technology's success is the concern — and where "AI safety" means something entirely different than it does in a chatbot conversation.
Follow the Money
The investment is historically unprecedented. The four largest cloud companies — Amazon, Microsoft, Google, and Meta — are spending roughly $725 billion on infrastructure in 2026, up about 77% from the year before; counting everyone, total compute investment crosses a trillion dollars this year. Morgan Stanley projects this private spending will exceed U.S. defense spending as a share of GDP by 2027.
Against that: the gap between spending and actual AI revenue is estimated at several hundred billion dollars a year, and the divergence between investment growth and revenue growth now exceeds what preceded the telecom crash of 2001. The money is uneven by arena. Coding assistance has real revenue. Cloud infrastructure collects cash regardless of which applications win. Consumer chatbots are subsidized loss leaders. Image generation is commoditizing toward zero margin. Companionship is quietly profitable. Whether the whole bet pays off depends almost entirely on the arenas — enterprise automation and agents — that haven't yet delivered.
The Energy Question
Environmental impact is real but wildly uneven — and the first unevenness is within the data centers themselves. Data centers run the whole digital economy: streaming, banking, e-commerce, social media, business software, the ordinary internet. AI is currently a minority of that load — estimates for 2025 range from roughly a tenth to two-fifths of data center electricity, with traditional computing still the largest share and cryptocurrency mining taking a meaningful slice. But AI is nearly all the growth: AI server demand is rising around 30% a year against single digits for everything else, which is why the International Energy Agency projects total data center consumption — about 1.5% of global electricity in 2024 — to more than double by 2030, roughly Japan's annual usage. "Data center energy" and "AI energy" are different numbers today; on current trends they converge.
The second unevenness is per use, and here the popular intuition inverts. A chatbot query runs a fraction of a watt-hour; an hour of Netflix runs about 26 chatbot queries' worth. A daily streaming habit almost certainly outweighs a daily chatbot habit — much of streaming's energy is in the device and network rather than the data center, but it's energy all the same. The genuine AI outlier is video generation: a few seconds of AI video can consume nearly a kilowatt-hour, more than hours of streamed video. If you're worried about AI's energy footprint, video generation and always-on agent workloads are where to look — not the chatbot answering a student's question.
The third unevenness is geographic. Global percentages stay modest while local ones don't: data centers consume over a quarter of Virginia's electricity and, by some estimates, approaching half of Frankfurt's and most of Dublin's. Electricity prices and water strain concentrate where the buildings are. "AI's environmental cost" is really a story about specific uses in specific places — and, increasingly, about whether the growth curve holds.
Why the Distinction Matters
Notice what happens when you separate these out. The strongest criticism of AI — that it confidently makes things up — is devastating for research assistance, serious for writing and everyday answering, and irrelevant for image generation, protein folding, or surveillance. The strongest defense — that it demonstrably works — is true for coding, true in the most troubling way for surveillance and targeting, and largely aspirational for autonomous agents. Some arenas are LLMs through and through; some got an LLM layer on decades-old technology; some barely involve LLMs at all. The job losses are real but land arena by arena, by thinning rather than sudden replacement. The environmental cost belongs mostly to a few energy-intensive uses in a few overloaded regions.
So when we ask whether AI is transformative, overhyped, or dangerous, the useful first question is: which one?
No comments:
Post a Comment
I hate having to moderate comments, but have to do so because of spam... :(