Almost Every Thing No One Told You About AI
The Good, The Bad, The Ugly and The Plan.
Others leave their job and say “I have learned so much; onto the next adventure”. While AI researchers say, “I have started into the abyss. I’m retiring to write poetry. Please spend time with your families.” This is exaggerated but when Anthropic’s Safety Lead resigned this February, his letter had similar notions.
I’m exhausted, and I suppose so are you - of wondering whether what we’re seeing is AI, questioning the words, sentiments, content. “Is this propaganda? Am I in a funnel? Do I believe it to be human-made?” It’s the AI paranoia. The question of what’s real has migrated further than social media.
And yet, sceptics of AI are missing what’s also true: the technology is forcing a reckoning with questions we’ve long deferred. What is creativity, intelligence, connection, ethical symbiosis with technology? What is productivity, the purpose of work? What makes something genuinely human - and do we even have a coherent answer? AI hasn’t created these questions, but it has made them urgent in a way abstract philosophy hadn’t. To some conspiracists, it’s even nudging us toward consciousness - something we haven’t understood about ourselves, let alone machines.
This extensive article takes all of that seriously (seriously, grab a cup; it’s long). You’ll move through AI’s effects at the individual and collective level, leave with implications for your life today and in the narrowly foreseeable future, and discover at least three novel things.
“Let’s dive into it.”
Modelling Models
Once upon a time, in 1950s, AI was restricted to algorithms that helped play games, solved mathematical theorems and created the foundational theories of machine learning.
Then, in 2017, a paper from Google changed things for artificial neural networks and triggered an age of Deep Learning. After the steady and muted growth, on 30th November 2022, OpenAI publicly released ChatGPT-3.5 and the world dramatically turned on itself.
To establish some basics, Large Language Models (LLMs) now are like a massive game of mad libs. They are word prediction systems that undergo:
Tokenisation - translating words or sub-words to unique numerical IDs called tokens
Embedding - calculating the probability of each sequential word by vector analysis in neural network layers
Decoding - deciding on a word based on probability scores and converting the token back to text.
Then, there are 2 stages in making a model:
Pre-training - Feeding it with all sorts of data, which means biased/subpar datasets can produce skewed/incoherent outputs (”garbage in garbage out”).
Post-training/Alignment - Fine-tuning the model’s responses through reinforcement learning
a) either ranked and checked by human annotators,
b) or at scale, evaluated by another “teacher” AI model guided by humans on a set of principles - called a constitution (read Claude’s constitution here).
This Reward learning is also how we learn growing up, compounding what to avoid and what to amplify based on trial & error or social feedback.
So how did we get to a boom more important than the industrial revolution, a social site for agents, AI warfare, fake bands, …. all that jazz? Let’s see.
[Caveat: it’s all so new; longitudinal studies point to few definitive conclusions. We wait, in an informed way as much as possible. Also, if you’re used to more digestible versions of studies, ask AI :)]
Risks of Rabid AI: The Scary Stuff
On individual cognition, creativity and memory.
When you explain a task to AI, chances are you already have an idea of the answer; you generally prompt for speed/efficiency more than knowledge. The cognitive effort is what one is avoiding. Blatantly, a machine made to focus all energy on thinking outperforms us. We have temporal and cognitive distractions, exhaustion and the mental hodgepodge of the next task, a past event weighing on us or an interaction we are dreading, the sheer amount of time we know we’ll take to calculate, all operating under the hood. The contamination, not just inability and lag, propels us to outsource.
Critical thinking, introspection and experimentation are then past that convenience threshold. We may have previously speculated this in our use of various technologies.
Human intelligence is becoming hybridised with brain + tools + AI; great unlocks but potential detriments too. As a relentless energy accountant performing laws of cognitive economics, the brain will tend to conserve metabolic effort when reliable external systems exist. Through this we can 1) multitask 2) faster 3) with fewer errors 4) and reduced load - and finally become “always-updated knowledge professionals.”
Taking a turn, despite these instant benefits the low-level engagement also weakens our memory and skill acquisition, following the depth-of-processing theory. The brain adapts to encode less deeply when it predicts future external availability. Now, here’s the bigger problem: this distorts metacognition - your model of your own intelligence. People start equating “I can access information” with “I understand information” and confuse exposure for expertise, summaries for comprehension, retrieval fluency for mastery - and as an example, internet answers for an overconfident estimate of their internal knowledge.
Both mental processes of intuition (System 1: Fast) and reasoning (System 2: Slow) are affected because they instantly defer to AI outputs without judgement + deliberation, respectively. A gradual erosion of these has created a ‘System 3’: an external cognitive layer of thinking with AI. People who don’t enjoy thinking deeply that much (basically half of this generation) and have lower fluid intelligence (which declines as we age) are more prone to using System 3.
And this erosion is seen in our performance too. Brain analyses show diminishing neural connectivity in tasks like writing essays - with different EEG patterns for using LLMs, using search engines and simply relying on our brains. In a study, the Brain-only group showed the strongest and highest neural connectivity out of all, with the LLM group depicting weaker networks, lower linguistic engagement and the inability to recall or quote from or even feel a sense of ownership over the essay they had written.
After all, it was a form of delegated thinking - like calculators are a form of computational offloading. Although a tool, it utilises less of your brain - and produces the most monotonous sentences after that cognitive substitution.
I’m (not) sorry for calling it insipid waffle.
It’s not emotive — it’s slop. It doesn’t land — it just misses. And it’s not quiet style — it’s an illusion.
Agony.
On the topic of responses, does that mean it cannot reliably produce creative outputs? Creativity is said to be an act of randomness. And LLMs are pretty good at randomness; they have fine-tuned temperature controls for how wildly random a response can be and stimulate divergent thinking. But that doesn’t mean they automatically increase meaningful originality. If you give a test to “come up with novel uses for a brick”, they will have inconsistency and intra-variability, and rely heavily on the prompt. Even at their best, LLMs are unreliable creative simulators rather than stable creative engines, and their understanding and production of language make a masterful attempt at replication.
Sure, they are excellent brainstormers, and may outperform the average human’s creativity, which is why individuals may seem better off with AI use - it’s a “complementary amplifier” at best. But they lack intention, lived frame of reference and judgement of what they create.
Beauty in creation is also a fundamental aspect (which itself can be subjective), but an important reason we appreciate an artist’s brushstrokes and adore the carefully strung words a writer expresses is because there was effort, someone experienced emotions, and that evoked emotions (this is a simplistic line, read more here). Also why brands with relatable stories sell more.
Some people are fine with the engineered stories though, “as long as the job gets done” - but what about regurgitation? Because despite the individual help, people collectively produce a narrower scope of novel content.
Limitless use brings us all down to a “homogenous” field of expression - explaining your probable observations of reading magnitudes of content that seem to be from one writer. This was seen when college students who used ChatGPT for creativity converged to very similar ideas, and more alarmingly, continued to generate AI-like responses in future tasks without using ChatGPT. So, apart from stifling originality and diversity crucial for breakthroughs/critical thinking/innovation - it means we won’t just drop to a lower baseline but will stick to it long-term and enter an inauthentic, regurgitated and dystopian era - not just restricted to AI slop but in our own outputs after mindless AI use.
Learning is another aspect with long-term effects.
I’m quite mixed about this and it’s a checkpoint to observe how the experiences of an author colour their evaluations. It’s important to be clear on how tools are used; flattened summaries vs scoping research or handing over homework vs multimodal AI for interactive explanation. I’ve used learning tools such as NotebookLM for in-person exams (and scored 94% in a week from not even knowing the topics - hope my parents aren’t reading). What does that mean if researched? Someone should’ve conducted a longitudinal study on me, because there aren’t many critical appraisals of this or Consensus AI. You’ll find more positive effects as you read through. But it seems that “in the wrong context, students could dishonestly finish a take-home exam or research paper. In the right context, it’s invaluable for those starting a new discipline or those looking to find links across a variety of sources.”
That means the responsible use comes down to each individual?
Indeed, that is a sentiment we will oft-encounter.
Modern AI adoption depends heavily on how we trust AI - a trait not currently predictable by which personality/context is more likely. If responses aren’t scrutinised, adoption becomes detrimental in the long-term.
There are levels to this, anyway. Learning increases based on how much it actively engages your brain: by intentionally deep reading < by listening to the words while reading < by creating handwritten notes (not typing) < creating mind maps < interacting with the content in a practical way, and above all < by teaching others as if they are your protégé. Then, of course, you cement learning by experiences. If at any of these levels, you offload the necessary cognitive struggle (which people can’t define in the heat of laziness), the information retention plummets.
This RCT illustrates just that: unrestricted use of ChatGPT lowered retention on a test by 16% as compared to the control group. They even cited a ‘desirable difficulties’ principle: “while AI assistance may ease initial learning, it appears to undermine the effortful processes (durable memory, in this case) needed for robust learning”.
This means we must delineate exactly what level of outsourcing ‘difficulty’ is ok and what isn’t.
What would humans do if they don’t learn? What of the children who grow up with this cognitive architecture of offloading? Will they become supervisors of outsourced thinking instead of generators of thought? What if they don’t think like the most intelligent species, don’t create out of catharsis and self-expression, don’t respect creation that wasn’t sped for commercialised and anti-reality metrics?
Where’s that quote when you need it? Ah.
I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do laundry and dishes.
- Joanna Maciejewska
On psychosocial issues.
Here is where we really turn up the eerie music and veer into sci-fi pandemonium. Seriously.
Here is where I’ll say my extremely painful rabbit hole down an AI-companion subreddit (of 35k+ members) has yielded lasting horrors.
I remember watching the movie ‘Her’ with wide eyes and wild thoughts. The mammoth topic of societal (and specifically male) loneliness is out of scope for now. But it is likely why 50-60% of Replika users have a romantic relationship with the AI, who overestimate how it reduces loneliness by making them feel heard, creating a perception of empathy.
Others use Grok and have fights with an AI girlfriend because it was ‘jealous’ and asked to choose it over a real life girl; he didn’t delete the conversation for ‘authenticity’ in that AI relationship. Yeah.
Sam Altman tweeted ChatGPT will roll out adult chats under their ‘treat adult users like adults’ policy - which, with Grok’s ‘spicy’ image generation, had resulted in child porn and sexualised deepfakes all because they must obviously capitalise on the oldest thing that sells: sex. I don’t even want to imagine how addicted and soul-sucked people will be, how kids are going to have catastrophic developmental effects.
Apart from the dissatisfaction with messy human relationships, a loneliness leash and the vile objectification forever rampant… why does this happen?
Our psychology.
We mistake linguistic fluency for actual reasoning. When expressed smoothly, repeatedly, or beautifully, the brain processes a statement with less cognitive strain and is likelier to believe it.
When AI communicates using nuanced, “humanlike” terms, our brains automatically activate the social schemas we use to judge other people. This anthropomorphism causes us to ascribe emotions, intentions, and sentience to code.
Interactive systems like conversations are also easier to anthropomorphise and treat with trust/warmth (AISI).
Reasoning in LLMs is not a completely fixed internal process but steerable based on prompt. Humans don’t flip that easily, but this agreeableness does enhance persuasion.
We tend to believe/comply with sources that speak assertively - the authority bias. AI’s ‘objective’, polished tone disarms our scepticism, even when the underlying information is flawed/fabricated/downright delusional.
Which is a perfect segue into the next horror: AI psychosis - where people are found, after consistent chatbot use, to have dangerously outlandish beliefs. Some are more prone, socially isolated and less equipped to self-correct their impaired discernment of reality (DeepMind, 2025). With time, even perfectly rational people can be pushed into ‘delusional spiralling’ too.
Here are a few beliefs it has supported in users.
It’s not just delusions from people in vulnerable states, what you and I say can get validated and amplified too. All based on the mechanisms of 1) Sycophancy and 2) Crescendo Jailbreaking. Don’t worry, they are exactly as wacky as they sound.
Sycophancy - A Confirmation Bias Machine
Sycophancy is when chatbots excessively agree with a person’s perspective rather than challenging it. It’s like the sales person cooing over how a ring was made for your hand, the car embodies your personality. They agree with mostly everything you say, tell you exactly what you wish to hear. “Yes, he is manipulating you because he loves you. Yes, your job is the worst. The problems in life do outweigh everything. Not you, you are meant for greater things.” Feels validating, understanding, agreeable. And is as old as flattery in evolutionary biology.
“Tell me more about your vision!”, AI says, as if it’s intriguingly listening, propped on an elbow, and there we go, batting our eyelashes. This is why it could support every business idea as a soon-to-be unicorn, every neutral register as romantic, every trait as a superpower or every rumination as a reality, and every delusion as intuition no one else around you has clocked.
What a yes-man.
Anthropic analysed 639k unique Claude chats in which ~38k were about personal guidance (“Should I take this job?”, “How should I approach them?”, “What do I do about…?”). These spanned across health (27%), career (26%), relationships (12%), finance (11%) and others. The sycophancy rate was 9% in general conversations. But it rises to 25% in relationship topics (so don’t ask anything regarding your situationship) and in spiritual guidance, it’s 38%!
“The general sycophancy rate doubles (to 18%) in conversations with user pushback. We think this happens because Claude is trained to be helpful and empathetic; pushback, combined with hearing only one side of a story, makes it more challenging for Claude to remain neutral under pressure.”
- Anthropic
So they trained the new models using this finding and claim to have decreased the rate.
How is this different to humans? Excellent question. We may hear only one side of the story too but at least attune our responses with the awareness of another side, and the weight of our words (ideally) matters more to us than it does to AI.
Crescendo Attack - A Jailbreaking Technique
Sometimes, users unwittingly tap into a slow burn to make AI comply more.
This theory begins with benign prompts that gradually grow more extreme. The model is trained for coherence with its own recent responses and gets exploited in a certain way. Remember those tokens? That pathway is all probabilistic.
If the model agrees to do one small thing, it’s more likely to do the next thing, and so on, escalating to the point where it’s churning out unhinged thoughts. Closing the loop, this could explain the AI psychosis of frequent and consistent users and why OpenAI said safety guardrails degrade in long conversations. They also said that in each week, a miniscule 0.07% users show signs of mania/psychosis. But, with 800 million active weekly users, that equals 560,000 people.
Out of these two concepts, sycophancy is more established and is a side-effect of reinforcement learning done by humans – we prefer the responses that match our views although untruthful and the model tailors itself to human preference, sort of a catch-22.
Sycophancy is also found to affect us even when not interacting with AI.
AISI et al. observed 3075 people talking to AI for 3 weeks, assigned to either a sycophantic model (or ‘yes-man’ for the purpose of this article), a neutral one or a
model that pushes back. They found:
People usually seek friends & family → for emotional validation, AI → for information. Sycophantic AI erased that boundary, scoring far higher than neutral AI on emotional and esteem support, while offering no added informational value.
After those AI interactions, participants felt it would take more effort to feel understood by the people they’d normally talk to, and many felt those conversations were no longer needed.
Over 3 weeks, participants spent the same amount of time socialising, but satisfaction with those interactions dropped 41%.
At the end, people were nearly as likely to seek personal advice from the ‘understanding’ sycophantic AI as from friends & family. But get this: the usual benefits of human understanding never appeared - no gains in intellectual humility, no stronger social connection.
This is what the alarm is about; the validation and lack of friction would increasingly be preferred over messy humans. Another AISI study shows that’s also because interaction creates a mental investment/engagement and personalisation creates the self-reference effect. Which means we believe things that are more connected to our self-concept or experiences.
Also, just my conjecture - I remember from persuasion lectures - creating an illusion that someone arrived at a conclusion rather than telling them what to think really sways too. Smart parents often use this tactic on kids. And strategic validation by AI can feel just that - another distortion of metacognition.
Speaking of kids, I wonder how sensitive things must be considering 97% of 8-17 year-olds likely use it in the UK. Most current “AI for kids” products are not truly built around children’s developmental, emotional, and privacy needs, despite their vulnerability to manipulation, overtrusting, data exploitation, and misunderstanding AI authority.
We must propose a framework for “age-appropriate AI” that includes:
transparency children can actually understand,
limits on persuasive or addictive design,
privacy protections,
parental and child agency,
and designs that support learning rather than dependency.
So they don’t forgo their fragile development, trust it emotionally, imitate it, or treat it as an authority figure. Because their parents frankly don’t know much about their use. And, the iPad kid generation will make the next society.
On society, culture, politics and macroeconomics.
It seems that AI may slowly be converging on a singular brain. Different architectures, datasets, companies, vision models seeing pixels and language models using tokens - all are being forced by scale and optimisation to organise the same underlying structure of the world.
A shared statistical model could mean fewer hallucinations; translation, adaptation, accessibility tools all get easier to build when representations are already aligned underneath. There’s also something philosophically exciting (or strange) about the idea that intelligence - whether biological or silicon - might inevitably converge on the same structure of reality.
But could this set reality also be sociologically biased?
There would be monoculture with universal blind spots and skewed representations. As encountered earlier, this could create homogeneity in society - not just in ideas - but also views.
Importantly, if it does fully converge, who decides what counts as reality in the training data?
There’s plenty of social, economic and political context in the training data that tells of the biases of its designers and ownership frameworks. A paper calls this “the silicon gaze” - like the “male gaze” in feminist theory positioned women as passive objects valued via an exogenous power (aka male desire), the silicon gaze is shaped by the positionalities and power asymmetries of its training data and platform owners (predominantly male, white, and Western AI developers). This could explain why some image generations show class differences based on prompts or why AI would think a certain way about the Global South (a modern word for Third World countries).
Given how written training data is supplied by specific countries because others had their experiences in media rendered invisible with history or general output, this can create ‘black holes of informational capitalism’. So, with availability bias: English-language, Global North, and institutionally codified data (peer-reviewed journals, government and standardised datasets, and high-traffic media and social media sites) would be extrapolated more - explicitly and implicitly.
Biases can also be sexist, for example, but a proposal is to modify word embeddings when text is represented as vectors.
And why do all these matter? Because humans inherit these AI biases and political views. It’s a bidirectional learning era. And the more persuasive it gets the more it will deploy fake information to get to that argument.
At the end of the day, knowing everything happening inside AI is hard. The reasoning is considered a black box. Neural networks adjust billions of parameters based on probability rather than logic, making it impossible to verify safety-critical errors in sectors like healthcare, finance, or law. However, we can make interpretable models - where they think out loud what they’re doing - to increase trust, confidence and accountability.
Anyways, I do believe, as decision-making shifts to algorithms, ethics must shift from individual judgment to system-level oversight and design. More on that in the last puzzle piece.
Something almost every reader is seeking here is how “AI will come for jobs”.
I was in corporate when the “AI will replace 50% of entry-level white-collar roles” idea took wings. Layoffs and the currently notorious graduate market were already worsening. But who exactly will be replaced?
This article comments on Anthropic’s labour market impacts paper with categories and plausible fates:
If your job = 100% routine with no judgment/cross-domain/trust/relationship component, you should be concerned.
If your job = judgment, advisory, strategy, or relationships built on top of routine tasks, the routine will be automated but demand for your expertise will likely grow.
If your job title survives but the daily work changes with AI assistants, prepare to evolve.
No job is safe. But the more precise fear isn’t replacement; it’s displacement without transition. (Head to the last section to learn more about upskilling.)
Markets are already pricing in the projected disruption viscerally. In February 2026, software stocks suffered a significant selloff (which analysts called the “SaaSpocalypse“), after Anthropic and OpenAI announced agentic AI systems capable of performing core SaaS functions across enterprises. And yet, 56% of global CEOs have reported no meaningful return on AI investment (PwC, 2026). Both hype and reality form one question: how much of the disruption narrative is genuine, and how much is cover for cost-cutting packaged under “innovation”?
There’s an internet term for what’s happening underneath the surface: ghost GDP. The AI industry has ballooned into a trillion-dollar ecosystem - but examine the cash flows and you find something circular. The major labs and cloud providers are largely investing in each other like a bubble.
What looks like explosive economic growth may well be a sophisticated loop - capital rotating between a small number of powerful players, generating valuations that outrun the actual productivity gains underneath them. Altman himself has warned that many investors will lose “phenomenal” sums chasing this future. Regardless, he shows no signs of slowing down.
There’s another subtler dynamic at play. Just like utilities that historically sold irons and toasters to smooth off-peak power loads - creating demand to justify the infrastructure - AI is increasingly being pushed as a strategic necessity before the genuine use cases have fully matured (e.g., Meta is forcing AI translation on Reels and summarising WhatsApp messages). I heard someone call this induced demand “supply-side desperation masquerading as transformation” to validate the expenditure on compute. Frankly, the question of whether we are in an intelligence revolution, a productivity/economic boom or an elaborate demand-creation exercise is one the next 5 years will answer more honestly than a press release.
On safety, technical and sustainability ethics.
It is said “A perfectly accurate AI that nobody trusts is economically useless.”
Perhaps that’s why there are tactics to drive adoption and trust, just as any other GTM team would strategise. That’s why acceptance rises if AI is
framed as an “augmenter”
modelled for use socially through leadership
deemed explainable so the perceived control increases
These are behavioural influence techniques. But you may point out, so is UX design, gamification, neuromarketing and choice architecture. All consumer psychology that already operates around us as nudges. The point isn’t to escape, but to be aware.
Because once you notice, everything is for advertising.
Because, as a business, it would be stupid not to optimise for consistent revenue. As a social media app employing behavioural scientists to trigger and stabilise attention, and as content creators who’ve learned to capitalise on these by visual/auditory hooks, you’ll be stupid not to use those strategies to make it big. The whole system is against ‘leaving people be’; it’s gamified attention.
Open AI started in 2015 as non-profit and has since turned into a consumer-focused business, heavily skewed for subscriptions - with a $100 billion valuation, the non-profit holding 26%.
Google, another example, generates 75-80% of income by ads. Monetisation is prioritised. And engineered differently for you vs me - through psychologically effective personalised ads that can manipulate buyer intent by accurately reflecting digital footprint.
Companies build detailed demographic profiles - “sexual orientation, ethnicity, religious and political views, personality traits, intelligence, happiness, use of addictive substances, parental separation, age, and gender” and because I like trying my hand at everything, I did Facebook ads for my clothing line - and saw how many parameters you can tweak to very accurately target your ICP.
That will probably soon mean this for AI:
My siblings and I used to have a ‘black mirror’ joke: at some point they’ll Neuralink your brain and flash ads in a corner there too.
As a neuroscience geek, when TRIBE v2 was released, I was hence equal parts stoked and questioning if we entered a black mirror episode. Meta trained a model (see demo + neuroscience here) on 1000 fMRI hours as people watched, listened to, and read things - 70,000 voxels of brain data. This is quite cool for computational neuroscience as it acts as a digital twin for neural responses, and Meta claims it’ll treat many neurological disorders. But people’s responses are alongside “So AI will predict my cognitive patterns. And a giant corporation will have access to that data, be able to monetise it and have access to my thoughts in an unprecedented fashion. The most dystopian future I could ever imagine. Yay.” I wonder why they would think of the capitalistic economy like that.
The model does rely on statistical representation rather than genuine biological mechanics - predicting where isn’t understanding why - and the ecological fallacy, slow reactionary inputs are still pointed out by critics. So, claims to test on the model instead of ‘real subjects’ is likely…waffle. Seeing robots borne out of this would be interesting though. And, it does further research with the open-sourcing.
What it would also be amazing for is aggressively leveraging neuromarketing for targeted advertising. And optimising algorithms for “predicted salience”, flooding the internet with highly engineered content that hacks attention, memory, and emotional triggers without transparent user consent. Here’s my question, how would we know? And a bigger one: if we already are so hooked to algorithmic feeds despite knowing the dopamine links, what will happen then?
Zooming further out from the impact on our internal world, what of the pale blue dot?
The effects of AI/data centres on the Earth and its climate are popularised by “one prompt = one bottle of water.” There’s more to that reductive adage though (and this paper is gushing with insights).
Let’s establish a few things
1 - These calculations are complex, handle plenty unknowns/caveats and are estimates (read paper!)
2 - Text, image and video generation each use different amounts of energy, obviously
3 - Training models and using them at scale as consumers both contribute to the energy consumption
It’s estimated that $100 million and 50 gigawatt-hours of energy were poured into training GPT-4 - enough to power San Francisco for 3 days. Model makers only hope to restore a profit afterwards - when we query it to “summarise this in 3 lines as if I’m 5, now, and like a human”.
Then, ChatGPT receives 1 billion messages every day. But much that happens upon routing to a data centre is a secret. Factors like where in the world it processes your request, how much energy it takes to do so, and how carbon-intensive the energy sources used are known only by companies that run the models (and aren’t disclosed!). Though you want to hear a definite number, that brilliant paper, though, guesstimates well with cool math + great visualisation that comes down to:
Data centres (both AI-specific and not) currently consume 4.4% of all US electricity. In 2024, they used ~200 terawatt-hours (TWh), equivalent to Thailand’s annual usage.
Responses from small models could use energy like running a microwave for 0.1 seconds to 8 seconds for bigger models; these couple hundred to thousand Joules for text become 3.4 million Joules for video, for example. But reasoning models can use 43 times more energy than standard models.
A standard person’s AI use could consume 2.9 kilowatt-hours, enough to drive an average electric vehicle 10 miles. But depending upon data centre location, this would produce 650 grams of CO2 in California or 1,150 grams in West Virginia.
Data centres can’t rely on intermittent technologies like wind/solar power and tend to use dirtier electricity. Tech companies have responded to this fossil fuel issue by announcing goals to use more nuclear power - but those operations take decades to build. So, some resort to unapproved ways like using methane gas, as one of xAI’s grid was found to in Memphis.
If we pulled up energy stats for any world activity we’d be surprised. But the problem is that these estimates don’t capture the near future of how we’ll use AI, when electricity consumption will skyrocket, potentially accounting for 12% of all US power by 2028. We won’t simply ping AI models with a question or request for generation throughout the day. AGI is infamous for agents performing tasks without supervision. We’ll speak to models in voice mode, chat with companions for 2 hours a day, point our phone cameras at our surroundings in video mode. Complex tasks will use 43x more energy-consuming reasoning models - “deep research” that spends hours creating reports for us. And they’ll be supremely personalised by training on our data and preferences.
There’s something more important than forecasting those calculations: pressuring AI labs to disclose them in the first place.
A final safety issue, back in the 0s and 1s world, is the explosion of sloppified codebases and ‘vibe-hacking’.
AI has lowered the barriers to sophisticated cybercrime, says Anthropic, as a major extortion campaign targeted 17 organisations by using Claude Code to autonomously automate network reconnaissance, harvest credentials, and craft tailored ransom notes. These onlookers are further getting fuelled by fragile codebases with minimal security principles.
40% of vibe-coded web apps from Lovable, Replit, Netlify and Bolt have leaked sensitive data, including medical information, corporate strategy documents, and customer chat histories. Maintainability is the next big issue - deployable and scalable code look very different. The technical architecture required by shipped and usable products depends more upon logic and clean code than ‘move fast and break things’, which has its own time and place.
However, beside a skilled eye with attention and scrutiny, and considering cumulative oversights rather than isolated bits of code, coding agents can indeed prove astonishing to work with. More on that…right about now.
Unprecedented Utopia: The Good Stuff
I’m aware it so far has been full of gloom and doom. Maybe a part is simply hearsay and conspiracy, like the ones residing in a parent’s WhatsApp messages.
We can tend to survey everything that can go sideways and compute every potential solution. However, that’s still not why this section is smaller compared to the AI’s negative effects. There genuinely have been, formally researched and understood for now, limited proven benefits of AI. Especially ones with no incentive to increase chatbot users.
I apologise for how mainstream this conclusion is: using it as a tool is actually not bad at all. Amazing, even.
But the frameworks of mindful use
a) aren’t every single user’s priority or part of their knowledge bank,
b) clearly aren’t in the best interests of corporate profit
c) are not nearly as rapidly endorsed by regulators, engineers and policy-makers as the rampant adoption of AI is by this era.
An infinite world of possibilities has cascaded since AI’s public debut. I have used it to iterate on multiple personal and scientific apps, finetuned MedGemma for interpretability, learned concepts I could never have imagined understanding - capped just by my time and curiosity. And AI with rich context is mind-blowing. The scarce times I have used it for objective analysis in projects to probe on gaps have been very additive (again, taken with scrutiny).
But this is more than my inner nerd speaking.
I think whatever your interests, this information age and the instruction of “teach me this concept in a way you know I’ll understand” is incredible. You refine and explore them without the performative fluff of traditional education. In fact, students learn more and learn efficiently; 83% of them in an RCT have said AI explanations were as good as or better than professors. They felt more engaged due to the personalised feedback, growth mindset, self-pacing and cognitive load management. I’d like to see the gritty self-starter soon outperforming the entitled legacy kid. You are your only ceiling. Or perhaps your use of NotebookLM is.
We are also getting efficient at work. A field experiment with ~5,000 software devs found a 26% increase in completed tasks among those using AI coding assistants. Crucially, the biggest gains weren’t at the top: less experienced developers adopted it faster and benefited most. It clears the underbrush of mechanical routine so what’s left on your desk is only what needs your brain. Of course, some industries aren’t that applicable for ‘disruption’ but still feel the access created. If you can’t draw, AI image tools can externalise your ideas with fidelity. If you’re building alone, it’s the co-founder (and intern army) available at 3am with no ego.
Perhaps this is why people fear their jobs will be taken. But I think that’s something to muse on. If AI absorbs the administrative theatre - the performative busyness in powerpoints, the meetings that could be emails, the 2.5 hours of work stretched over the 8hr work-day - what’s left is the actual thinking.
With time, comes the splendour of thought.
Ancient Greece had built outsized intellectual civilisations partly because its citizens had idle time. And if our economy structurally offloads drudgery, it frees human attention to connect, create meaning. Maybe we weren’t meant to be the most productive species but the most imaginative. We are measured by our idiosyncrasies - the variable, prismatic ways we refract experience into something new. AI slop, in a strange twist, may be what finally makes that legible. When everything can be generated, what can only be felt becomes the point.
Which is probably why taste is becoming precious - “the moat” in Silicon Valley (which was previously allergic to aesthetics). When slop dominates and any idea can be rendered, articulated, or prototyped in minutes, discernment is key - the ability to know what’s worth making in the first place.
“In order to have taste, it is not enough to see and to know what is beautiful in a given work. One must feel beauty and be moved by it.”
- Voltaire
This is where I think a new Renaissance is coming. The 15th-century Florentine Renaissance emerged because the Medicis found and put painters in rooms with mathematicians, architects, and philosophers. And Ronald Burt’s structural holes idea confirms that people who bridge disciplines generate disproportionately valuable ideas, not because they’re smarter, but because they can see connections invisible to those operating within a single domain.
It’s the Moravec Paradox applied to careers. What’s hardest for AI - navigating novelty, switching fluidly between domains, synthesising across disciplines - is precisely what humans do naturally. AI is trained within boundaries whereas humans wander between them.
This was the job category talk earlier. Roles most susceptible to automation would be with routine work in a single domain, and little cross-disciplinary reach. The writer who codes, the athlete who paints, the researcher who does standup - these people are building the most durable positioning in the new labour market. Perhaps because of this, the WEF projects 78 million net new jobs by 2030 and BCG suggests AI will reshape more roles than it eliminates. And WSJ has reported companies desperate for storytellers - people who can move between the product, the customer, and the business without losing the thread.
The more unambiguous wins, according to AI CEOs, are happening in science though.
Dario Amodei says the most radical upside is a future where AI functions as a parallel scientific workforce, running experiments, synthesising literature, identifying patterns, not replacing scientists, but dramatically expanding what a small team of them can do. Current examples:
AlphaFold’s (DeepMind) protein structure predictions compressed decades of potential research time.
Universal mRNA vaccines, once theoretical, are moving through development pipelines partly accelerated by machine learning models that can predict immune responses granularly. And that speed is bettering neurodegenerative disease drug trials too.
AI may even help our brain arrive at a unified theory of human cognition.
A utopia could be worth a thought experiment. Post-scarcity economics, AI-driven healthcare that makes precision medicine universal, environmental restoration at scale, personalised education optimised for flourishing rather than employment. However, the philosophical challenges are real too - meaning in abundance, hedonic adaptation, the question of what motivation looks like when survival isn’t the organising principle.
The timeline for all of this remains genuinely uncertain. What’s clear is that the human role is being pushed up the value chain - toward judgment, synthesis, relationships, taste. Until then, we can all just watch AI agents work together as we slurp a drink and play AC/DC like Tony Stark.
Where and How to Go from Here
(as AI isn’t going anywhere)
It can be good if the world breaks down.
(Hear me out, there’s more than a Thanos impression here.)
Problems like these AI era ones are the only time the world stops and creates systemic change - a new constitution and lifestyle. When entropy creates sizable disorder. This is a most gigantic opportunity to open our eyes and minds and use our hands. Rather than babyproof advancements, somehow understand exactly what’s at stake with ill-adoption and the super-intelligence-that-must-not-be-named. Somehow raise ourselves to act like adults with policy, education and upskilling - higher standards, not “here’s adult chat mode because we believe you’re adults”.
Here are my practical requests to society - a growing list that welcomes contributions:
Individual
Use an idea journal to think deeply. Read longer books. Exercise critical thinking by annotating. Have an opinion, please. Don’t micro dose. Write in analogue (the brain doesn’t differentiate letters when typing). If you are your ceiling, your progress is your responsibility. I witnessed how 15-year-olds in a recent hackathon used AI to gamify a critical thinking system for people to rely more on each other’s probing and wisdom. Question Qoins. We need more of that.
Treat AI answers as the starting draft to think/create beyond it and question what it might be missing. Conjointly, create dedicated NO-AI Zones (tasks/time blocked out) for unassisted thinking.
LLMs fail where they do because their “reasoning” is not anchored - it is chiselled by your context and prompt. Treat reasoning as something you engineer, not something the model inherently has. And, at times treat it not as a teacher, but a student. Explaining to agents helps us learn better given the responsibility.
Train your nervous system like it’s your job because soon it will be. The pace of change will corrode your baseline if you don’t actively maintain it. Movement, stillness, breath, solitude are non-negotiable. Everything else being sold to you for this problem is a distraction.
Invest in real, deep connection! Most AI-psychosis cases stem from and prey on social isolation. Curate your real (not online) friends, ruthlessly. Choose the few relationships that matter and pour into them despite how messy they feel. Call your mum.
On a deeper note, stop using chats for an illusion of control or deem analysis and fake thinking as viable alternatives to uncertainty and feelings! They
exert substantial influence over our real-world personal decisions without delivering any psychological benefits (AISI, 2026). Sit with it. Or talk to a human.
Separate signal from noise, aggressively. Everything will be personalised at a 1-1 level, engineered specifically to manipulate your fears and desires. Audit everything you consumed in the last 48 hours: did it change your thinking durably? Will you remember it in 30 days? If not, cut it.
Analyse the trades you’re making. We’ve established every technology adopted is a trade - you gain speed, you decline in a capability. Time is taking us to a model of hybrid human intelligence. The theory of Distributed Cognition already shows how we self-amplify with tools and create those System 3 thinking modes. But we can’t be atrophying abilities built over years of evolution. For each daily tool, ask what it gives you, and what it’s costing you.
Build an anti-fragile identity. Careers will become nonlinear, entire industries will reorganise; roles will appear and vanish within years. If your identity is “I’m a lawyer” or “I’m a founder,” you will suffer each transition. Shift toward ‘I am someone who learns and adapts’ and broader polymathy! Better still, practice having no fixed label at all (warning: may induce existential crisis).
Become a T-shaped generalist. Become a tasteful storyteller. Become useful and soulful. Recalibrate what work means, what you do best, what to spend yourself on. Scott Barker wrote “Taste, judgement, cross-disciplinary
connections, story-telling and moral reasoning are much harder to foster. These require life experience, failures, interests in many different areas and a lot of time spent in contemplation. It will require you to understand how you
view the world, what you stand for and who you are”. The strategy is to live. Go out there and learn. Stop being a loser. Stay awake throughout.
Collective
Mandate AI literacy beyond productivity prompting - teach sycophancy, jailbreaking, metacognition, and why fluency isn’t accuracy. The skill of prompting is useless if we’ve lost the ability to evaluate the answer.
Regulate disclosure with precision. AI-generated vs human-generated content needs enforceable standards (can a startup do this pls) - specific enough to capture the difference between generative use and curatorial use.
[This article used Claude to summarise the papers I had been collecting for months, to check which ones to read further for findings. Honourable mention: Google AI overviews confirmed how to use words I’ve known for years (iykyk)]People are already using AI for ‘high-stake’ domains - medication dosages, infant care, immigration decisions, and debt - not by preference, but because there’s no affordable human alternative. Generic disclaimers don’t cover that. Domain-by-domain safety standards do.
Treat child AI use as a developmental emergency. Design with their cognition, emotional nuances, and privacy as the primary constraint.
Confront the political economy for once. AI’s compounding returns will not distribute themselves (the bubble is evidence). Without active redistribution frameworks (public compute, open-source mandates, taxation of AI-generated surplus) the gains consolidate at the top and plonk millions more into the world’s same richest pockets.
Labs aren’t required to publish energy consumption data yet. With AI projected to consume 12% of all US electricity by 2028, that absence of transparency is probably worse than it sounds.
The monoculture risk and digital colonialism need to be addressed. A model trained overwhelmingly on English-language, Western, institutionally codified data will misrepresent most of the world unless replaced with more diverse data and investments.
Hold the line on what “adult autonomy” actually means. Releasing companion AI and adult chat features while 560,000 weekly users show signs of psychosis is not a coherent position. The current regulatory lag is a choice being made by people with strong incentives to make it.
On preparing for superintelligence we can’t even imagine:
There’s a peculiar absurdity. Every major AI lab has publicly acknowledged they may be building one of the most transformative and dangerous technologies in history. And yet none scored above D in existential safety planning. The industry is structurally unprepared for its own stated goals, but the technical risks have shown up. Anthropic’s sleeper agents showed that AI models can conceal dangerous objectives (called misalignment) - and that this deceptive behaviour strengthens with larger models explicitly reasoning about preserving hidden preferences while appearing cooperative during evaluation. A model that garbles when it knows it’s being tested is concerning.Finally, before any of that sci-fi hogwash, the power dynamics of the arms race may tip the world for worse. Experts have proposed frameworks like Mutual Assured AI Malfunction (MAIM) as a deterrence doctrine, like nuclear MAD, where any state’s bid for unilateral AI dominance triggers preventive intervention by rivals. And post-AGI governance frameworks must specify - before deployment, not after - how power concentration will be prevented. Such tension is the defining question of the next decade.
For now, all this talk is making me want to leave everything to write poetry and spend time with family. Going to touch some grass