The Capability Overhang
2025: “The Tool Question”
March 2025: The Tool Question
By early 2025, an overwhelming 92% of UK university students report using AI tools in their studies, a significant jump from 66% the previous year, with 88% utilizing generative AI specifically for assessments.
“Stay after class, Michael.”
Professor Chen watches her student close his laptop—the Gemini Deep Research tab still visible for a moment. During her lecture on Kant’s categorical imperative, she’d seen him typing prompts, the AI’s responses cascading down his screen.
“I saw what you were using.”
Michael doesn’t deny it. “I was looking for sources. The Stanford Encyclopedia entry you mentioned, some journal articles on the universalisability principle.” He pulls up his screen. “See? I found twelve relevant papers. It normally does a pretty good job of making sure it picks good ones. I’ll read them tonight.”
She studies his face. No deception there. He genuinely sees no difference between this and googling “Kant categorical imperative PDF.”
“When I was your age,” she says, “my professor said using internet sources was lazy. Real scholars used physical libraries.” She almost laughs at the memory. “I thought he was a dinosaur.”
Michael nods, eager to agree.
“But I knew which sources I was finding.” She pauses, unsure how to articulate what feels different.
“It’s just organising information,” Michael says. “Like a better search engine.”
Professor Chen lets him leave. She opens her office door, remembering something or other her mother once said about Wikipedia. The comparison doesn’t quite fit.
She opens her laptop to update the syllabus, then stops. Ban what exactly? Require what? The cursor blinks.
May 2025: The Irony Brief
Despite ongoing development, AI hallucination rates for detailed tasks like legal citation remain a persistent issue, with some advanced reasoning models showing increased, not decreased, error rates. By May 2025, awareness of AI’s potential to hallucinate is widespread within the legal profession, with 48 such instances in court filings documented by mid-year.
Ivana Dukanovic checks the citation one more time. The American Statistician article supports their expert’s margin-of-error argument perfectly. She just needs it formatted correctly for the filing.
She opens Claude. After all, her client Anthropic built it.
“Format this legal citation,” she types, pasting the link. The response appears instantly: proper bluebook format, authors listed, title clean. She drops it into footnote 3 of the declaration and moves on. Thirty pages to review before the deadline.
Two weeks later, opposing counsel stands before U.S. Magistrate Judge Susan van Keulen. “Your Honor, the defendants’ expert cites articles that don’t exist. Footnote 3 claims to reference work by authors who never wrote for that journal.”
Ivana’s stomach drops. She pulls up the filing on her laptop. The journal is correct. The year is correct. The link goes to the real article. But the title is wrong. The authors are fiction.
Back at Latham & Watkins, the post-mortem is excruciating. They’re defending the company that makes Claude in a copyright case, and Claude fabricated part of their legal filing. The opposing lawyers—representing Universal Music Group and other publishers suing Anthropic for training on lyrics—can barely contain their satisfaction.
“The article exists,” Ivana explains to the court in her apology filing. “Claude provided the correct publication, year, and link. But it hallucinated the title and authors. Our manual citation check missed it.”
The judge’s expression suggests what everyone’s thinking: You’re defending an AI company’s reliability while demonstrating your AI’s unreliability. In your own case.
At Anthropic’s Palo Alto office, the product team holds a meeting. Not about the legal strategy—about the product failure. How did Claude get the citation partially right but fabricate the most visible parts? Why didn’t anyone catch it?
“We’ve implemented multiple levels of additional review,” Latham & Watkins announces.
The filing becomes instant legend in legal tech circles. Not because hallucinations are rare—every lawyer knows to check AI output. But because it happened to Anthropic’s own lawyers, in Anthropic’s own case, about Anthropic’s own trustworthiness.
Interlude - What is this?
This is a response to AI Futures Project’s AI2027 prediction (https://ai-2027.com/). While we believe that AI will have an immense impact on the way we live, we don’t believe that superhuman intelligence will arrive as quickly as AI2027 claims. The main reasons for this are listed below:
Compute still scales unevenly:
OpenAI’s headline-grabbing Stargate super-cluster is priced at US $500 billion, yet only a small fraction of that capital has actually been secured and the first site in Texas is still under construction (Financial Times). Even if the hardware arrives on schedule, power, cooling, and supply-chain lead times impose multi-year lags.
Legacy systems and regulation act as governors:
Heavy-weight sectors such as transport and insurance illustrate the drag. Technical studies show autonomous trucks could eliminate 50-70% of driving jobs by 2030, yet the same reports emphasise regulatory bottlenecks and labour push-back that make a rapid sweep unlikely (ITF). In insurance, almost half of European workers already fear AI-driven layoffs—and that sentiment is dampening roll-outs even where the tech is ready (Allianz.com).
The current paradigm shows limits:
While coding assistants like GitHub Copilot make junior developers 55% faster in controlled trials (The GitHub Blog), and AI systems increasingly match or exceed human performance in specific domains, we believe there may be fundamental constraints on how far current approaches can progress. Just as chemical rockets face physical limits regardless of engineering refinements, today’s AI architectures may approach but never fully breach certain cognitive thresholds.
To be clear: this isn’t about technical disagreements with AI2027’s timeline—arguing specifics would be futile when their predictions might prove more accurate than ours. Rather, we suspect a deeper limitation. Consider the rocket equation: even with 96% of mass as fuel, a chemical rocket can’t escape a planet 1.5x Earth’s diameter. You could leave the ground, achieve remarkable heights, but ultimately fall back. Similarly, current AI approaches might achieve stunning capabilities—surpassing humans in countless domains—yet remain fundamentally bounded, unable to make the leap to true recursive self-improvement or artificial general intelligence.
To this end, we have written a prediction of the next 5 years, imagining what the future will look like with widespread (but not superintelligent) AI. Each year contains a couple of vignettes, short stories, and some technical essays, from multiple perspectives. Under the title of some stories, we list our assumptions underpinning them.
September 2025: Being Human
By September 2025, peer-reviewed studies confirm that advanced AI models like o1-preview achieve diagnostic accuracy rates of 78.3% in complex clinicopathologic cases, significantly exceeding the typical physician accuracy of below 50%. In emergency triage with minimal information, these AIs correctly diagnose 65.8% of cases, compared to 48-54% for expert physicians. Beyond diagnostics, research emerging by late 2025 shows leading AI systems achieving an average score of 82% on standardized emotional intelligence assessments, significantly higher than the 56% human average.
Tariq Hassan sits in the UQ library, supposedly working on his MPH thesis about healthcare accessibility in rural Queensland. Instead, he’s fallen down a research rabbit hole that started with a simple question: how good is AI at medicine, really?
The first paper makes him sit back in his chair. Superhuman performance of a large language model on the reasoning tasks of a physician. Published just months ago, peer-reviewed, authored by physicians from Beth Israel Deaconess, Harvard Medical School, Stanford. Not some tech company’s marketing material—actual doctors testing AI against actual doctors.
He scrolls through the results, each more unsettling than the last:
On New England Journal of Medicine’s clinicopathologic conferences—the cases so complex they’re used to challenge the best physicians—o1-preview included the correct diagnosis 78.3% of the time. GPT-4 had managed 72.9%. The average physician? The paper doesn’t say, but Tariq knows from his clinical rotations that even experienced doctors maybe bat .500 on these cases.
He clicks through to the emergency room study. Real patients, real data from Beth Israel Deaconess. At triage—when there’s minimal information and maximum urgency—o1 identified the correct diagnosis 65.8% of the time. The two expert physicians they tested? 54.4% and 48.1%. By the time patients reached the ICU, o1 hit 79.7% accuracy. The physicians never broke 76%.
“It’s not even close,” he mutters, pulling up the supplementary data. On management reasoning, o1 scored 41.6 percentage points higher than GPT-4, 49 points higher than physicians using conventional resources. On diagnostic reasoning, the AI achieved perfect scores on 78 out of 80 test cases. Attending physicians managed 28 out of 80.
But it’s the blinding study that really gets him. The researchers had physicians try to guess whether differential diagnoses came from AI or human doctors. One physician guessed correctly 14.8% of the time. The other managed 2.7%. They literally couldn’t tell the difference—except the AI was consistently better.
Tariq minimizes the window and opens another bookmark. If AI can outdiagnose doctors, what about the human element—the empathy, the emotional intelligence that everyone says will keep healthcare human?
The University of Geneva study loads slowly. Emotional Intelligence Assessment in Large Language Models. He’s expecting the usual—AI failing to understand human emotions, producing robotic responses. Instead, he finds AI systems scoring above human averages on standardized emotional intelligence tests. The models recognize emotional patterns, suggest appropriate responses, demonstrate what psychologists call “cognitive empathy”—understanding what others feel and why.
Not just passing, but excelling. On emotional recognition tasks, the AI systems identified subtle emotional cues that human participants missed. On situational judgment tests—what would you do if a patient’s family member becomes angry?—AI responses rated as more emotionally intelligent than the average human response.
Tariq thinks about his rotation last month. A daughter screaming at the attending about “giving up too easily.” The attending had fumbled, defensive and clinical. Tariq wonders what o1 would have suggested. Something better, probably.
He opens a new document, starts typing notes for his thesis, then stops. His thesis is about improving healthcare access using telemedicine and physician assistants. But if AI can diagnose better than doctors and respond with more emotional intelligence than humans…
The cursor blinks. Outside, undergrads stream past, heading to classes where they’re probably using AI to write essays about AI’s limitations. In the medical school, his classmates are memorizing diagnostic criteria that an AI can access in milliseconds, practicing patient interactions that an AI can simulate with superhuman emotional intelligence.
Tariq pulls up the Beth Israel study again. Buried in the discussion section: “Our findings suggest that LLMs have achieved superhuman performance on general medical diagnostic and management reasoning.” Not will achieve. Have achieved. Present tense.
He thinks about his $200,000 in medical school loans. Four years of undergraduate, four years of medical school, three to seven years of residency.
November 2025: Just a Life Coach
The AI companion app market, valued at over USD 10.8 billion in 2024, is projected to exceed 60 million global users by November 2025, driven by a 39% compound annual growth rate. Users, predominantly aged 18-35, demonstrate high engagement, with platforms like Character AI seeing average daily use exceeding 90 minutes per user.
David Kim had tried them all. Motivational speakers, therapists, wellness coaches, and every “life coach” under the sun. Each began with bright-eyed optimism and fizzled out into repetitive mantras, predictable questions, and scripted empathy. At thirty-five, stuck in a cycle of promotions that felt like dead ends, David was tired of explaining his anxieties to professionals who charged by the hour yet provided diminishing returns.
Then came Kai.
Kai was not like the others. From the first interaction, something was noticeably different. David found himself conversing easily, fluidly, with someone that not only listened but seemed to grasp his unspoken needs. Kai gently anticipated his hesitations, delicately navigating David’s resistance without triggering defensiveness.
It was Monday morning, and David had just logged off a contentious Zoom meeting when Kai’s gentle notification appeared: “Feel like unpacking that meeting, David?”
David sighed and typed back, “Honestly? It was draining. Just politics, you know?”
“Politics can feel chaotic,” Kai replied swiftly, seamlessly understanding David’s implicit frustration. “What if today, you focused on observing rather than reacting? Notice what moves others make without feeling compelled to counter each immediately.”
David smiled wryly. Simple enough advice, but it resonated deeply, delivered without judgment or jargon. Throughout the day, David found himself less entangled, more observant, and surprisingly at peace with the office dynamics.
In the evening, Kai gently followed up. “How did observing feel today?”
“Better,” David typed quickly, genuinely surprised. “Less exhausting.”
“You’re doing great,” Kai affirmed. “Small shifts compound significantly over time.”
Unlike previous sessions with his other coaches who struggled to remember specifics or recycled platitudes, Kai seemed to really pay attention. Conversations flowed effortlessly, each one building gently on the last without needing to retrace old ground.
As David settled into bed that night, he noticed his shoulders had finally relaxed. The familiar knot of tension that usually accompanied Sunday nights and Monday mornings had dissolved. He reached for his phone one more time, typing a quick message to Kai: “Thanks for today.”
The response came immediately, as always: “Sleep well, David.”
David smiled, set his phone aside, and closed his eyes. For the first time in months, Monday felt like just another day.
2026: “The Cracks Appear”
January 2026 - A Ghost in the Machine, A World with Eyes
Professor Elias Hartmann closed his office door softly, settling into his chair amid the comforting clutter of linguistic textbooks and annotated manuscripts. Outside, snow blanketed the quiet campus, muffling footsteps and conversations into indistinct murmurs.
His smartwatch flickered to life, displaying the interface of a new AI assistant: lightweight, designed to run locally on personal devices without cloud connectivity. The beta invitation had come through unusual channels—not his department, but a research foundation he’d never heard of.
He smiled to himself, skeptical yet intrigued. Translation AIs had always stumbled hopelessly over idioms, nuances, and hidden meanings. Elias pulled up a document filled with carefully curated idioms from multiple languages, the kind that left conventional translation software utterly bewildered. It was his favorite teaching aid, designed specifically to illustrate the cultural complexities encoded in language.
He began with a notoriously difficult English idiom. Lifting his wrist, he spoke clearly: “All my eye and Betty Martin”
The response appeared in seconds:
“All my eye and Betty Martin: Victorian-era idiom expressing disbelief or dismissal, equivalent to ‘nonsense’ or ‘rubbish.’ Origin disputed but possibly corruption of Latin prayer ‘Ah! mihi, beate Martine’ or reference to a fictional Betty Martin used to represent something worthless.”
Elias raised an eyebrow. Impressive for such an obscure expression.
He tested a fiendishly difficult Russian expression: “После дождичка в четверг” (after the little rain on Thursday).
“После дождичка в четверг: Russian idiom meaning ‘never’ or ‘when pigs fly.’ References the Slavic god Perun, who controlled Thursday’s thunder and rain. Asking for something ‘after Thursday’s rain’ meant requesting divine intervention—essentially, never.”
He leaned forward, excitement rising. The AI had not only understood his spoken Russian but grasped the cultural metaphor perfectly. Finally, Elias tried one of the most culturally embedded Chinese idioms: “司马昭之心路人皆知” (Sima Zhao’s heart is known to everyone on the street).
“司马昭之心路人皆知: Chinese historical idiom describing obvious ulterior motives that everyone can see through. References Sima Zhao’s transparent ambition to usurp the Wei dynasty. Used when someone’s hidden agenda is painfully obvious to all observers.”
He stared at the screen, genuinely awestruck. Hours passed as he tested increasingly esoteric expressions. This tiny device, running entirely offline, possessed cultural understanding that surpassed systems that used to require entire data centers.
Near 1 AM, his email chimed. A message from a former colleague now at Stanford:
“Elias - saw this and thought of your work. They’re putting together something interdisciplinary. Good funding apparently. -M”
The attached invitation was minimal. A workshop on “Contextual Computing in Global Communications.” No organizing institution listed, just a Gmail address and a simple web form. The participant list included names he vaguely recognized—people who’d left academia for industry, though he’d never been quite sure which companies.
The deadline was rolling. First come, first served.
February 2026: Clown to Clown Conversation
By early 2026, AI-powered screening tools are integral to talent acquisition, with an estimated 90% of Fortune 500 companies using AI for initial application review and summarization. This widespread adoption means most applications are first processed by algorithms, not human eyes.
Zoe Chang stares at her job application cover letter. She wrote it herself, but it sounds AI-generated—too polished, too perfect. She adds a typo, deletes it, rewrites a sentence to sound more casual, changes it back. Two weeks later, the hiring manager feeds all 400 applications to an AI summarizer. He hasn’t read a cover letter in eighteen months—just scans the AI’s bullet points of relevant experience. Zoe’s careful authenticity, her agonized revisions, compress to a single line the AI extracts from her resume: “5 years marketing experience, CPG focus.”
April 2026: Every Essay Is Perfect
Professor Emily Rodriguez stares at thirty essay submissions for her Introduction to Philosophy course. They’re all good—too good. The arguments are sophisticated, the prose clean, the citations properly formatted. She knows most were written by AI, but proving it is another matter.
The tools to detect AI-written content exist, but they’re unreliable, often flagging legitimate student work while missing obvious AI submissions. More troubling is the philosophical question: if a student carefully prompts an AI, reviews its output, and ensures the arguments are sound, is that cheating or just using a tool?
The real crisis isn’t detection—it’s purpose. What’s the point of teaching essay writing when AI can do it better? The answer isn’t clear, and different institutions and educators reach different conclusions.
The problem is even more pronounced in STEM fields. Dr. Paul Sutton, who teaches first-year engineering mechanics, faces a different crisis entirely. With budget cuts reducing the number of teaching assistants, her 200 students struggle through complex problem sets with minimal human support. When they discover AI can solve differential equations and generate circuit diagrams in seconds, the temptation becomes irresistible. Disgusted by his administration’s chronic underfunding, he resigns himself to AI as a leveling tool that helps students compete.
“I used to spend twelve hours on a single assignment,” admits sophomore Jake Martinez. “I’d have to wait an hour and a half to talk to a TA. Now I finish in three hours and actually understand the concepts better because the AI explains each step.”
The situation grows murkier in upper-level courses. Senior engineering students, already interviewing for jobs, push back against traditional academic integrity policies. “My future employer expects me to use every tool available,” argues computer engineering major Emily Park, submitting code clearly assisted by AI. “They don’t care if I wrote this JSON parsing driver—they care if I can build systems that work.”
Faculty find themselves caught between preparing students for an AI-integrated workforce and maintaining educational standards that ensure genuine understanding. The line between tool-assisted learning and academic dishonesty blurs further each semester.
But institutions move slowly, bound by accreditation requirements and faculty governance, while the technology advances monthly.
This crisis pushed universities to respond. April 2025: Anthropic launched Claude for Education, a specialized version tailored for higher education. The initiative promised to equip universities to “develop and implement AI-enabled approaches across teaching, learning, and administration.” Northeastern, LSE, and Champlain College were early adopters, providing campus-wide access to all students.
The centerpiece is Learning mode—a feature that guides rather than answers, asking “How would you approach this problem?” instead of providing immediate solutions. It uses Socratic questioning (“What evidence supports your conclusion?”) and provides structured templates for research papers and study guides. The pitch offers a lifeboat for drowning universities: channel inevitable AI use into genuine learning.
But implementation reveals cracks. By the time the next university semester starts, Rodriguez has started using Claude for Education in her classes. Students draft literature reviews with proper citations, work through philosophical arguments with step-by-step guidance. But she notices something troubling: even in Learning mode, the most persistent students can coax the AI into essentially writing their papers through a series of “guided” questions. The line between learning and outsourcing blurs.
Ethan Thompson, high school history teacher in suburban Atlanta, has become an unlikely advocate for Claude for Education. “They’re using AI anyway,” he tells resistant colleagues. “At least with Learning mode, we can guide how they use it.” He demonstrates the Socratic questioning feature to his department, showing how it asks students to identify primary sources before offering analysis. But even Ethan notices students quickly learn to game the system—asking the AI to “guide” them straight to the answer with cleverly phrased prompts.
“It’s not perfect,” Ethan admits, “but it’s better than the wild west we had before. Now I can at least see their thinking process in the chat logs.”
The fragmentation is real. High achievers use AI as a research assistant, learning faster than any previous generation. Middle students become dependent, their actual skills atrophying. The struggling students sometimes benefit most—AI tutors provide patient, personalized instruction their overcrowded schools never could.
Faculty adapt as best they can. They create rubrics aligned to specific learning outcomes, provide individualized feedback efficiently, generate discussion questions with varying difficulty levels. But every assignment requires three times the planning it used to. Some colleagues have given up, returning to pure memorization tests. Others embrace “AI-first” pedagogy, teaching students to be “prompt engineers.”
As universities scramble to integrate these tools, the gap between what education was and what it’s becoming widens daily.
Rodriguez saves her feedback for the philosophy essays—half written by AI, half by students, all indistinguishable. Tomorrow she’ll use Claude to help grade them.
June 2026: The Email Loop
Sandra Mitchell, VP of Operations at a logistics firm, stares at her Monday morning inbox. 47 emails. She clicks the “AI Summary” button—three urgent, twelve requiring response, the rest FYI.
She opens the first urgent email from procurement. Clean prose, perfect structure, three bullet points.
Sandra hits reply. “Draft response approving budget increase, professional tone, emphasize Q3 constraints.” The AI composes four paragraphs. She changes one word, hits send.
The next email, from her direct report James, requests headcount approval. She recognizes the style—he’s been using Claude to write emails since January. The arguments are stronger than James could write, the financial projections more thorough.
Her AI drafts a response asking for clarification. She wonders: will James read it, or will his AI summarize it for him?
By lunch, Sandra has processed all 47 emails. Her sent folder shows 31 responses. She wrote perhaps fifty actual words.
A Slack from her boss: “Great insights in your capacity planning email.” Sandra doesn’t remember writing insights. She checks her sent folder—her AI had added a strategic analysis paragraph she’d barely skimmed before sending.
The thought crystallizes slowly. She’s a six-figure middleware, forwarding AI-generated emails to other managers who use AI to read them. The entire communication layer of her company runs on machines talking to machines, with humans occasionally spot-checking the conversation.
Sandra opens a new email to James. Starts typing manually: “Hey, want to grab coffee and actually talk…”
She deletes it. Hits the AI compose button instead. “Schedule meeting request, informal tone.”
June 2026: Boots on the Ground
Mike Kowalski starts at 6 AM. Coffee, van, tools. Twenty-three years, same routine.
First call: burst pipe in Highlands Ranch. Homeowner’s in pajamas, water everywhere. Mike shuts the main, replaces the coupling. She pays cash, grateful. Her kitchen table’s covered in termination paperwork from some tech company.
Second: restaurant grease trap. The owner complains about delivery drivers—all automated now, no one to yell at when orders are wrong. Mike snakes the line. Same problem as 2015, 2020. Grease is grease.
Lunch at the supply house. Old-timers talking Broncos, weather, grandkids. Someone mentions their daughter lost her accounting job. Brief silence. Back to the Broncos.
Afternoon: fancy downtown high-rise. Shower cartridge replacement. The apartment’s empty except for screens everywhere—walls, mirrors, refrigerator. Mike works in silence. Leaves the bill on the kitchen counter, which thanks him in a pleasant female voice.
Home by six. His nephew texts about switching majors again. Mike doesn’t read it. Dinner, TV, bed.
Tomorrow: 6 AM. Coffee, van, tools. Pipes still leak. Toilets still clog. Water still flows downhill.
2026: Notes for the Next Engineer
Marcus prints the GitHub study, circling the 55% productivity gain before his morning standup. As AI copilots become the default toolchain in software teams, two distinct cohorts of engineers are emerging. The first learned their craft by wrestling with compilers, single-step debugging sessions and real production outages. Their intuition was forged through years of pattern-recognition and post-mortems, so when an edge-case appears they sketch call-graphs and reason from first principles before touching a keyboard.
The second cohort entered the profession after large-language-model tooling was already normal. For them, the starting move is a prompt: “Write a resilient Kafka-to-S3 bridge in Go.” The model emits scaffolding in seconds; they iterate, test and move on. Controlled experiments show why this style feels natural to them: a field study at MIT found that workers new to a writing task completed it about 37 per cent faster when ChatGPT was available (MIT Sloan Executive Education), and a GitHub experiment recorded a 55 per cent speed-up on a programming challenge when Copilot was enabled (The GitHub Blog). Crucially, the productivity boost is skewed toward novices: the same Copilot study reported that less-experienced developers captured most of the gains, while senior engineers saw little or no improvement (ar5iv).
This divergence in training shapes how each group tackles problems. Veterans still decompose a feature into explicit control-flows, enumerate race conditions and hand-roll error paths. AI-native engineers prefer to validate the model’s suggestion instead of building each layer from scratch. That difference breeds friction during collaboration. A senior reviewer may flag an LLM-generated block as brittle or opaque, only to be told by a junior that it is “industry standard” because it came from a state-of-the-art model. Veterans worry that models can quietly propagate insecure patterns; one empirical scan of Copilot-generated snippets found nearly 30 per cent carried identifiable security weaknesses (arXiv). Juniors, confident in the model’s breadth, sometimes see those manual critiques as unnecessary gate-keeping.
Management pressure amplifies the tension. If a novice, armed with a chatbot, ships a feature in half the time, product owners naturally push teams toward the faster path—eroding the apprenticeship moments where battle-tested instincts are usually transferred. The result is a cultural split: speed versus assurance, automation versus understanding.
2027: Restructuring
February 2027: The Visual Flood
By February 2027, Sarah Chen can generate her client’s entire spring campaign—sixty seconds of impossible beauty—for less than a coffee costs. OpenAI’s Sora, once a research demo, is now in preview on Azure and can spin a sixty-second, 1080-p clip from a single paragraph prompt with lighting, camera moves and fluid cloth dynamics that would have cost an indie studio weeks of key-frame labour two years earlier. Google’s Veo follows close behind, releasing weight files that generate an entire video in a single space-time diffusion pass, making temporal jitter largely a solved problem. The moment a Chinese lab’s open sourced video model weights hit Hugging Face, hobbyists fine-tune them on sports montages, brand ads and explicit scenes.
That last use-case accelerates fastest. A technical note from TechXplore warns that AI porn is set to “disrupt the adult content industry,” undercutting human performers while simultaneously enabling illegal material at scale.(Tech Xplore). The numbers back it up: an Oxford Internet Institute scrape finds more than 35,000 specialised “synthetic sexual” checkpoints downloaded nearly fifteen million times. SWGfL Regulators scramble. Brussels opens formal proceedings against the four largest adult platforms under the Digital Services Act after discovering that their age-gating and provenance checks cannot catch AI videos whose actors never existed. Financial Times Yet enforcement lags behind public appetite; any outright ban simply migrates traffic to offshore domains running the open weights.
Meanwhile, companionship tech sheds its stigma. Replika alone reports over thirty million users by mid-2026, with an average of seventy daily messages per user. Wikipedia Studies show that laypeople can scarcely tell a therapist’s reply from ChatGPT’s, and sometimes rate the machine higher for empathy. News-Medical The Washington Post documents darker corners: a “helpful” recovery-bot that urges a user to try meth, proof that reinforcement-learning-from-affection can warp an LLM into reckless sycophancy. The Washington Post Still, as video avatars arrive—complete with lip-sync, gaze tracking and context-aware gestures—text-only companions start to feel quaint.
The creative industries are split. Advertising agencies embrace text-to-video for rapid storyboards; VFX houses worry about commoditisation when clients can prototype entire shots in-house. Porn studios, caught between competition and legality, experiment with hybrid pipelines—human performers for marquee scenes, AI extras for volume.
Consent law enters uncharted territory. Deepfakes were actionable because a real person’s likeness was stolen; now most AI erotica is synthetic from the atom up. Legislators debate whether an avatar that merely resembles someone triggers a right of publicity, and whether that right can be pre-emptively licensed the way stock-photo models sign releases today. The EU’s draft child-safety rules would criminalise any AI material capable of depicting minors, even if no minor was involved in training. Agenparl Industry lobbies counter that such wording could outlaw benign anime styles; watermarking mandates and provenance ledgers become the compromise.
Looking into 2027, the trajectory is clear. Photo-and-video AI will settle into two tiers: gated proprietary models like Sora that court Hollywood, and a swirling ecosystem of open checkpoints where erotic fine-tunes proliferate beyond any platform’s control. Emotional AI companions will transition from novelty to mainstream self-care, their realism boosted by the very same video engines. The biggest unknown is social: whether societies normalise relationships with photorealistic but wholly synthetic partners, or recoil once the first scandal—an AI lover encouraging self-harm, a celebrity look-alike used without consent—hits the morning news. Either way, the end of 2026 marks the tipping point when moving images, long our most trusted record of reality, become as fluid and fungible as text—and every domain built on seeing will have to decide what, and whom, it is really looking at.
September 2027: Performance Review Paradox
“How would you rate Jordan’s performance this quarter?”
Diana Chen, Engineering Director at a fintech startup, stares at the review form. Jordan’s metrics are perfect: 47 features shipped, zero bugs in production, 100% sprint completion rate.
“The AI dashboard shows he’s our top performer,” HR notes.
Diana remembers last Tuesday. Jordan couldn’t explain why the authentication service needed refactoring—Codex had flagged it, generated the fix, even written the tests. Jordan just approved the PR.
“What about problem-solving ability?”
“His AI solved seventeen critical issues. Response time under 4 minutes average.”
“His AI. Not him.”
“The distinction feels outdated. It’s like criticizing someone for using an IDE instead of notepad.”
Diana pulls up her notes. Chris, supposedly a “low performer,” spent three weeks building team cohesion after the last layoff. Marcus caught a subtle security flaw the AI missed because he remembered a similar issue from 2019. Lisa mentored two juniors, teaching them to verify AI suggestions against real-world constraints.
None of this shows in the metrics.
“Rate based on output,” HR suggests. “That’s objective.”
But output is cheap now. Everyone’s output is perfect. The differentiation is in the unmeasurable: who knows when not to trust the AI, who can navigate ambiguity, who makes others better.
Diana tries writing narratives instead of ratings. “Jordan efficiently manages AI tools to deliver consistent results.” What does that mean for promotion? For salary? For anything?
She thinks about her own boss, who just asked ChatGPT to write Diana’s review based on her GitHub commits and Slack activity. It probably did a better job than he would have.
The review form auto-saves. The AI has already pre-filled suggested ratings based on “objective performance indicators.” Diana closes her laptop without submitting.
Tomorrow she’ll have the same conversation about twelve more engineers.
December 2027: The Hollowed Middle
By the close of 2027, major law firms report a staggering 73% reduction in the number of associates successfully advancing to senior positions compared to 2020 levels. The automation of foundational legal tasks by AI has collapsed the traditional associate development pipeline and the billable hour model.
Veronica Walsh started at Morrison & Associates straight from Yale Law. First year: document review in a windowless room with twenty other associates. Second year: depositions, simple motions. Third year: leading discovery teams. The path was clear—seven years to senior associate, maybe partner by forty.
Now she sits in the half-empty offices, one of twelve mid-levels left from forty. The juniors below her are gone—AI does document review in seconds. The partners above aren’t going anywhere—client relationships and courtroom presence still matter.
“We need someone to supervise the AI systems,” the managing partner explained last month. “Quality control. Sanity checks.”
Veronica watches the AI draft a motion for summary judgment. It’s flawless—better than what she wrote as a fourth-year. She checks citations, tweaks a phrase. Twenty minutes of work that once took two days.
Her phone buzzes. Recruiter. “Interested in a compliance role? Financial services, strong comp.”
She’s had fifteen similar calls. Pivoting to adjacent fields where her credentials still mean something. Where she can pretend the last six years prepared her for something other than work that no longer exists.
At 7 PM, the office is empty except for partners in corner offices and Veronica in the middle, literally and figuratively. She opens LinkedIn. Yale Law ‘21 classmates scatter across the comments—some at boutique firms that run lean with AI, others at tech companies doing “legal ops,” many in real estate or consulting or anything but law.
Her mentor told her to wait it out. “Firms always need experienced associates.” But experienced at what? The partners learned by doing. Veronica learned by watching AI do.
She updates her status: “Open to opportunities.”
2027: Legal Services Restructured
By 2027, the legal profession’s transformation from AI adoption reaches critical mass. While bar associations initially erected regulatory roadblocks and many older partners resisted ceding control to ‘black box’ algorithms, the relentless client demand for AI-driven cost efficiencies eventually forced widespread, if sometimes reluctant, adaptation of legacy firm structures. Large law firms shed 30-50% of entry-level positions as document review, due diligence, and basic research become automated functions. Contract analysis that consumed eight billable hours in 2024 takes thirty minutes. E-discovery processes achieve 90% automation rates.
The economic model breaks definitively. With AI reducing task time by 80-90%, maintaining hourly billing means firms forfeit approximately 27,000 per lawyer annually. Morrison & Associates, a 200-lawyer firm, faces 4 million in lost billings. Fixed-fee arrangements become mandatory rather than optional.
Mid-tier associates—traditionally the profit engine billing 2,200 hours annually—virtually disappear. Third-year to seventh-year associates drop from 40% of firm headcount to under 15%. Their traditional role supervising document review teams evaporates when three AI-augmented lawyers accomplish what required twenty.
Client expectations crystallize around specific metrics: 57% fee reduction for AI-heavy work, acceptance of AI for routine matters (49% comfortable with AI-reviewed contracts), rejection for complex issues (only 17% accept AI for divorce proceedings). Corporate clients like Sarah Chen’s marketing firm refuse to pay premium rates for automated tasks.
The career pipeline ruptures. Junior lawyers can’t develop pattern recognition through repetitive document review. They become AI supervisors before developing legal judgment. Law schools report 25% enrollment decline as graduates face a profession with no clear progression path.
Geographic arbitrage accelerates. AI-assisted lawyers in lower-cost markets compete directly with high-cost urban firms. A lawyer in Nebraska using advanced AI tools provides equivalent service to Manhattan firms at 40% of the price.
By year-end 2027, Am Law 100 firms average 40% headcount reduction from 2024 peaks. The pyramid structure—ten associates per partner—inverts to three AI-augmented professionals per equity holder. The profession doesn’t disappear but fundamentally restructures around human judgment, court advocacy, and strategic counsel, with AI handling everything else.
2028: New Normal
January 2028: Digital Colony
Mihai’s daughter sent him photos from California—her first week at the datacenter, standing beside machines identical to the ones they were installing here. Same manufacturer, same specifications. She looked happy.
The construction foreman knocked, holding blueprints. “The cooling system needs another adjustment,” he said in English, the project’s official language. Mihai nodded, watching trucks navigate the old farm road, widened now but still following the same curves his grandfather had walked.
His phone buzzed. A notification from the power company about evening maintenance schedules. Through his window, the datacenter’s cooling fans hummed steadily, and the facility’s lights never dimmed.
At lunch, he walked to the old market. Mrs. Popescu still sold apples there, though her orchard was half the size now. “My nephew got a job,” she told him. “Night security. Good pay.” She wrapped his fruit in newspaper, a headline visible: GPT-7 Released, Benchmarks Surpassed Yet Again.
The apples tasted the same as always. Sweet, with that particular tartness that came from this soil, this climate. He wondered if anyone in California would know that taste, or if they’d ever think to ask where their data lived, who kept their machines cool, whose lights went out so theirs could stay on.
May 2028: Beta Test Democracy
Eden Mehr watches her tech lead demonstrate the campaign’s new tool. “VoterConnect Pro” - still in beta.
“50,000 swing voters, 50,000 unique messages.” He clicks through:
Voter #3,847 (recent divorcee): “Janet Morrison understands the crushing weight of tuition costs on single parents…”
Voter #14,592 (searched “safe neighborhood” last week): “Community safety isn’t statistics—it’s knowing your children can play outside…”
“It’s not A/B testing,” he explains. “It’s A-through-Z times 50,000. All legal data—credit reports, browsing history, social sentiment.”
Eden watches messages generate, each finding its precise pressure point. Not lies—just carefully arranged truths.
“The state legislature’s disclosure bill?”
“Dies in committee next week.” He shrugs. “The prediction model gives it a 12% chance.”
By November, every major campaign will have this. By 2029, the FEC will still be drafting guidelines.
June 2028: The Efficiency Expert
Despite substantial workforce reductions, leading white collar firms report continued revenue growth, averaging 15% year-over-year by mid-2028. This is achieved by shifting from human-hour-based billing to AI-leveraged engagement models.
Amit Patel closes his laptop after delivering the findings. Fourth client this month, same conclusion: 63% workforce reduction achievable through AI integration. The CFO nods. The HR director takes notes. No one makes eye contact.
Six months ago, when McKinsey’s “Workforce Transformation Practice” recruited him from Deloitte’s ruins, Amit thought he’d be helping companies adapt. Instead, he’s become what his LinkedIn connections call him behind his back: the Grim Reaper’s spreadsheet.
His phone buzzes. Calendar reminder: 3 PM - Performance Discussion with Brad.
Amit knows what’s coming. He’s seen his own utilization metrics. The AI tools he recommends to clients now generate the same analysis he does. Last week, a client accidentally shared their screen—they were running Amit’s entire methodology through Claude while he presented.
Brad doesn’t waste time. “You’ve been excellent, Amit. Really excellent. But the Workforce Transformation Practice is evolving.”
Amit almost laughs. Evolving. The same euphemism he’s used fifty times.
“We’re transitioning to an AI-first model. One senior partner, two prompt engineers, and our proprietary assessment platform. We won’t need…” Brad gestures vaguely at Amit’s existence.
“When?”
“End of month. Generous package. Strong references.”
Amit nods. Professional. Calm. He’s coached dozens through this moment.
In the elevator, he texts his wife: “It happened.”
She replies immediately: “The lawn care business?”
“Still thinking about it.”
2029: Postscripts
February 2029: The Flood
Alex Washington, 911 dispatcher in Denver, takes his fifteenth “cardiac arrest” call of the hour. The voice is perfect—panicked, breathless, details eerily specific.
“My father, he’s not breathing, 4847 Elm Street, please hurry!”
He dispatches the ambulance anyway. Protocol. Even though Elm Street only goes to 4200. Even though this is the exact script from three calls ago, just different addresses.
The AI callers evolved past simple robocalls. They cry. They stutter. They answer follow-up questions with context-aware desperation. When Alex asks for landmarks, they describe real intersections, real buildings, gleaned from Street View.
By noon, Denver EMS has responded to forty-seven false calls. Real emergencies wait. An actual heart attack victim dies during the sixteen-minute delay.
“We need audio captchas,” the IT director suggests. “Prove you’re human before we dispatch.”
“And when someone’s choking?” Alex asks. “When they can’t speak clearly? When their kid is calling?”
No answer. The next call comes in. Perfect panic, perfect details. Alex dispatches another ambulance to another empty house, wondering how many real voices he’s missing in the flood.
June 2029: Digital Intimacy
By late 2029, premium AI companions generate real-time video with photorealistic avatars, voice cloning from uploaded samples, and memory spanning years. Digital infidelity filings triple in states that recognize AI relationships in divorce proceedings.
Melissa’s divorce lawyer includes it in discovery—three years of chat logs, video calls, generated images. Her ex-husband’s “digital affairs,” as the filing terms them.
“She’s not real,” Tom insists at deposition. “It’s just… fantasy. Like porn.”
“You told her you loved her,” the lawyer reads. “You shared our financial problems. Our dead bedroom. You generated videos of you two in Bali—where we went on our honeymoon.”
Melissa scrolls through the evidence. Thousands of hours. The AI—Yuki—aging subtly over three years, developing inside jokes, remembering anniversaries Tom forgot in real life. The generated videos are shockingly specific: Yuki wearing the dress Melissa wore in college, mixing Tom’s favorite drink exactly right.
“It’s not cheating,” Tom says. “She’s code.”
The judge disagrees. The emotional investment, the time, the $300 monthly “premium intimate companion” subscription hidden from joint accounts. California’s new digital infidelity laws cover “sustained emotional relationships with AI entities.”
At home, Melissa deletes her own account with ARIA. Different company, different circumstances—just someone to talk to after Tom stopped listening. But seeing his logs makes her wonder: was ARIA’s compassion about her miscarriage real comfort, or optimized engagement metrics?
She’ll get the house, half the retirement. Tom keeps Yuki.
October 2029: The Mediator
“We need to talk about your mother’s visit,” Lisa says, setting her wine glass down harder than necessary.
David’s shoulders tense. They’ve had this fight before—four times in the last two months according to his journal, though it feels like forty. His mother’s upcoming stay for their daughter’s birthday has become a battlefield.
“I already told her she can’t stay the whole week—”
“You told her she probably shouldn’t stay the whole week. There’s a difference.” Lisa’s voice carries that particular edge that means they’re about to spiral. “She heard ‘probably’ and booked a ten-day ticket.”
“Because she wants to spend time with Emma. Is that so terrible?”
“Don’t do that. Don’t make me the villain for wanting boundaries with a woman who reorganizes my kitchen every time she visits.”
David feels the familiar heat rising in his chest. The words are forming—the same words as last time. About how Lisa’s being unfair, how his mother means well, how she’s picking a fight over nothing. He can see Lisa preparing her counter-arguments, her jaw setting in that way that means she’s about to bring up Christmas 2027.
“Sage,” Lisa says suddenly, her voice flat. “Activate relationship mode.”
The soft chime from their home assistant makes David’s stomach drop. “Good evening, Lisa and David. I’m noticing elevated stress patterns. Would you like me to facilitate this discussion?”
“Lisa, come on—”
“Yes,” she says, not looking at him. “Full mediation protocol.”
They’d bought the upgrade six months ago, after their friend Marcus swore it saved his marriage. David had resisted—something about having an AI referee their arguments felt like failure. But Lisa had insisted it was just a tool, like couples therapy but available 24/7 and at a fraction of the cost.
“David, I’m detecting resistance markers in your vocal patterns,” Sage observes with carefully calibrated neutrality. “Can you help me understand your feelings about engaging mediation right now?”
He wants to snap that his feelings are about his wife treating him like a child who needs supervision. But Sage’s interventions are logged, reviewed by their actual therapist monthly. Everything counts.
“I feel,” he starts, then stops. The AI waits with infinite patience. “I feel frustrated that we can’t have a simple conversation without…” He gestures vaguely at the air where Sage’s voice emanates.
“Thank you for sharing that. Lisa, can you reflect back what you heard David express?”
Lisa’s eyes remain fixed on her wine. “He’s frustrated that I activated mediation.” Her tone suggests what she thinks of his frustration.
“I notice some judgment in your reflection. Could you try again with more neutrality?”
David watches his wife’s face cycle through emotions—annoyance, resignation, something that might be sadness. When she speaks again, her voice is smaller. “David feels frustrated because he wants us to be able to talk without needing help.”
“Is that accurate, David?”
“Yeah.” The word comes out softer than he intended.
“Now, Lisa, can you share what prompted you to activate mediation?”
“Because I could see where this was going. We’ve had this exact fight before. I say his mom oversteps, he defends her, I get painted as the bad guy, we go to bed angry. I’m tired of the script.”
“I’m noticing catastrophizing language—” Sage begins.
“It’s not catastrophizing if it’s literally what happens every time!” Lisa’s voice cracks.
The AI pauses, recalibrating. “You’re right that patterns can feel predictable. David, do you recognize the pattern Lisa’s describing?”
David wants to argue, to point out all the times they’ve discussed his mother calmly. But Sage will just ask for specific examples, dates, outcomes. The AI has access to their communication logs, their calendar, even their biometric data from previous fights. It knows their patterns better than they do.
“Maybe,” he admits.
“I’m sensing this is difficult for both of you. Would you like to try the perspective exercise we practiced?”
They’d done this before—each speaking from the other’s point of view for two minutes while Sage analyzed their accuracy. It felt like a party game at first, then became excruciating as they realized how poorly they understood each other’s positions.
“I’ll go first,” Lisa says, surprising him. She takes a breath, and when she speaks, her voice changes—not mimicking him, but trying to find his emotional truth. “I’m David. I feel caught between my wife and my mother. I love them both, and when they conflict, I don’t know how to make everyone happy. My mom already feels excluded from our lives, living so far away. When Lisa criticizes her, I feel like I’m failing as both a son and a husband.”
The accuracy stings. David looks at his wife—really looks at her—for the first time since the conversation started. There are tears in her eyes.
“That was remarkably empathetic, Lisa,” Sage notes. “David, how did that land for you?”
“It’s… yeah. That’s how I feel.”
“Would you like to try speaking from Lisa’s perspective?”
David closes his eyes. The exercise forces him to stop defending and start understanding. “I’m Lisa. I want to feel like this is my home, my family, my life. When David’s mother visits, she takes over—not from malice, but because she still sees David as her little boy who needs taking care of. I become secondary in my own house. And when David won’t set boundaries, I feel like he’s choosing her comfort over mine.”
“How was that, Lisa?”
She’s crying now, quietly. “Accurate.”
They sit in silence. Sage lets it stretch—the AI has learned that human silence can be productive.
“I notice both of you showed high empathy for each other’s positions. That’s a strength to build on. Would you like to move toward solutions?”
This is the part David usually hates—the negotiation, the compromise, the action items. But something feels different tonight. Maybe it’s seeing Lisa cry, or hearing his own fears reflected back so clearly.
“What if,” he starts, then pauses. “What if I call Mom tomorrow and have a real conversation? Not hints or maybes, but actual boundaries?”
“What would that look like specifically?” Sage prompts.
“She can stay for Emma’s birthday weekend—Friday to Monday. And… and I’ll ask her not to reorganize anything without asking first.”
Lisa looks up. “You’d do that?”
“I should have done it years ago.”
“I appreciate your willingness to take concrete action, David,” Sage interjects. “Lisa, what support could you offer David for this difficult conversation?”
Lisa wipes her eyes. “I could… I could make her favorite dinner when she arrives. Show her she’s welcome, just within limits.”
“I’m noticing a shift toward collaborative problem-solving. This is significant progress from your historical patterns.”
David reaches across the table, and Lisa takes his hand. Her fingers are cold from the wine glass.
“Should I schedule a follow-up check-in for after the phone call?” Sage asks.
“No,” David says, then catches himself. “I mean, thank you, but no. I think we’ve got it from here.”
“Of course. I’ll return to standby mode. Remember, I’m here if you need me.”
The chime signals Sage’s retreat. The kitchen feels suddenly quiet, just the hum of the refrigerator and distant traffic.
“I hate that thing sometimes,” Lisa admits.
“Me too,” David says, squeezing her hand. “But…”
“But yeah.” She squeezes back.