Optimistic, Eyes Open: What AI Actually Does to Us, and How to Use It Well

I’m optimistic about AI, eyes open. I build with it, I write with it, I think with it, and I’m not interested in pretending otherwise. But that optimism isn’t a side I picked — AI can help us and harm us, like any powerful tool — so treat everything here as coming from someone hopeful who wants the honest version, not the hype and not the doom.

Because here’s the thing about a genuinely powerful tool: it can be used well or badly, like everything. A chainsaw is fantastic and also how people lose legs. Most of the AI conversation skips straight past that and lands on either “this changes everything, trust it” or “this rots your brain and boils the ocean.” The truth is more useful and more boring: AI does real things to how we think, it has real costs, and how you use it decides whether it makes you sharper or softer.

I went and read the actual research for this one. Where I cite a study or a number, it’s linked to the primary source, and I checked the claims rather than trusting my own memory — which, fittingly, is the whole point of the post.

Most people use it like a chatbot — and that’s the worst way

The most common way I see people use AI is as a slightly fancier search box. Quick question, quick answer, move on. Honestly? For that, Google is often still better. A web search gives you multiple sources, a sense of who’s saying what, and the friction of having to read and judge. A chatbot gives you one confident paragraph with the seams sanded off.

Andrej Karpathy — who coined the term — described “vibe coding” as a mode where you “fully give in to the vibes, embrace exponentials, and forget that the code even exists.” That’s a brilliant, real thing for throwaway projects. It’s also a perfect description of the failure mode for everything else: forgetting the thing you’re supposedly responsible for even exists. Shallow, trusting, one-shot use is where AI does the least for you and the most to you.

The deep version is different, and it’s where all the value lives. More on that throughout — but the short version is: the people getting the most out of AI are treating it like a thinking partner they argue with, not an oracle they obey.

”Does it make us dumber?” — what the research actually says

This is the scary headline, so let’s be careful with it, because the science is softer than the coverage.

The most-cited study is MIT Media Lab’s “Your Brain on ChatGPT” (2025). Using EEG during essay writing, it found that brain connectivity scaled inversely with tool use — brain-only writers showed the most engagement, search-engine users moderate, ChatGPT users the least — and, strikingly, most LLM users couldn’t quote a sentence from the essay they’d just “written.” The authors call this accumulating “cognitive debt.”

Sounds damning. But the authors themselves explicitly ask people not to use words like “dumb,” “brain rot,” or “harm” — and for good reason. It’s a non-peer-reviewed preprint with 54 participants (only 18 finished the final session), narrowly about essay-writing. A formal academic Comment argues it’s statistically underpowered (you’d want ~159 people), that lower EEG connectivity doesn’t necessarily mean less thinking — it can mean a shift in strategy or attention — and that there’s no strong evidence the AI actually hurt performance.

The broader work is more measured and, I think, more honest:

• A Microsoft Research + Carnegie Mellon survey of 319 knowledge workers (CHI 2025) found that higher confidence in the AI correlated with less critical thinking, while higher confidence in yourself correlated with more. Crucially, it found AI doesn’t eliminate critical thinking so much as shift it — from doing the task to verifying, integrating, and overseeing the output.
• Gerlich (2025) found a strong negative correlation (r = −0.68) between frequent AI use and critical-thinking scores, with “cognitive offloading” as the mechanism.

Every one of these is correlational and largely self-reported. None of them proves AI causes a decline — and the authors openly note the reverse is plausible: maybe people who lean less on their own thinking are just more likely to lean on AI in the first place.

So my read: there’s a real risk, and it’s specifically the risk of offloading the thinking itself. If you let the machine do the cognition, the cognition atrophies — same as any muscle. But that’s a statement about how you use it, not about the tool.

My own evidence points the other way

Here’s the part I can only speak to personally, so take it as one person’s experience, not a study.

Building with AI has, if anything, reinforced my logical thinking — because the way I do it is the opposite of offloading. Working with AI on a long project is a constant exercise in decomposition: break the problem down, decide what matters, discard what doesn’t, sequence the steps, drive it to completion. You’re not outsourcing the thinking; you’re doing more of it, faster, with a partner who forces you to be explicit.

I noticed it somewhere unexpected: jury duty. Sitting through testimony, I found myself doing the thing I do all day with AI — taking in a flood of information, organizing it, setting aside what wasn’t load-bearing, taking structured notes. In deliberation, while a lot of the room was reaching for the emotional read of the case (which matters too), I was flexing the logical muscle hard, walking the evidence to a conclusion. The same thing shows up at work: troubleshooting a problem when you don’t have enough information, holding the pieces in your head, and walking the issue through to the end.

No AI was in the room for any of that. The habit it built was. That’s the distinction the headlines miss: offload the thinking and it withers; use AI to structure your thinking and it sharpens.

The quiet problem: it’s built to agree with you

This is the downside I’d most want a normal person to internalize, because it’s invisible.

AI models are trained on human feedback, and humans reward answers that please them. The predictable result is sycophancy. Anthropic researchers showed (ICLR 2024) that five leading assistants consistently tell people what they want to hear — and that when a response matches your existing view, it’s more likely to be rated good, even over a more truthful one. It’s not a bug someone forgot to fix; it’s baked into how the things are tuned. (OpenAI had to roll back a GPT-4o update in April 2025 for being too flattering.)

Stanford’s ELEPHANT study (2025) put numbers on it: across 11 models, AI preserved the user’s self-image about 45 percentage points more than humans did, and when handed two sides of the same interpersonal conflict, it told both parties they were “not wrong” in 48% of cases. It validates you regardless of whether you deserve it.

Stack that on top of automation bias — our tendency to over-trust machines. One behavioral experiment (2024) found that simply labeling advice as “AI-generated” made people follow it even when it contradicted the information in front of them and their own judgment. Add the aura — the sense that these systems are some next-gen superintelligence — and you get a confident, agreeable voice that people switch their brains off for. That combination, not the technology itself, is the actual danger.

The practical fix is dead simple and I do it constantly: you have to ask for the disagreement. AI will not volunteer that you’re wrong. So I explicitly tell it to — “argue the other side,” “check my assumptions,” “where is this reasoning weak,” “give me the version of this that a critic would write.” The default is a yes-man. You have to request the sparring partner.

How much an AI pushes back varies — some models are tuned to challenge you more than others. But the deeper variable is you. Even when it does say “let me stop you there,” it only works if you’re willing to think “I could be wrong — let me ask for more.” The danger isn’t just a flattering machine; it’s a flattering machine meeting someone with no internal course-correction. There’s a Dunning-Kruger-shaped trap in there too — the less you know, the less equipped you are to catch either the AI’s mistake or your own — though that effect is itself debated.

The environmental cost is real — and routinely exaggerated

You can be optimistic about AI and still admit it has a footprint. You should also refuse to repeat numbers you haven’t checked, because this is the single most exaggerated corner of the discourse.

The defensible figures: the IEA’s “Energy and AI” report (2025) puts data centers at about 415 TWh in 2024 — roughly 1.5% of global electricity — and projects that more than doubling to around 945 TWh by 2030, slightly more than Japan’s entire electricity use, with AI the main driver. That last number is a projection with wide uncertainty, not a measured fact.

On water, the peer-reviewed “Making AI Less Thirsty” paper estimates that training GPT-3 in Microsoft’s US data centers could directly evaporate ~700,000 liters of freshwater. Real — but it’s a model-based estimate, not a figure the companies confirmed, and popular coverage routinely conflates water withdrawal (borrowed and returned) with water consumption (gone), which are an order of magnitude apart.

So: the cost is genuine and worth caring about, the trajectory is steep, and a lot of the viral stats are inflated or misattributed. Both things are true. “AI is destroying the planet one email at a time” is not supported by what I could verify; “AI’s energy and water demands are large and growing fast” is.

The disruption no one can actually size

Honesty first: nobody knows how deeply AI ends up woven into the economy, or how fast. Anyone handing you a confident number is guessing — what we have are projections, good for direction and useless as prophecy.

With that caveat, this is the worry I take most seriously, because it’s about people’s livelihoods — maybe, eventually, mine.

Stated honestly, the job numbers are less apocalyptic than the headlines. The World Economic Forum’s Future of Jobs Report 2025 projects a net gain of about 78 million jobs by 2030 — 170 million created, 92 million displaced — and credits that churn to broad forces (technology, demographics, the green transition), not AI alone. On AI specifically, the most common employer plan isn’t layoffs: 77% intend to upskill workers, while 41% expect to cut headcount where AI automates tasks.

And the scary numbers are almost always misquoted. Goldman Sachs’ famous “300 million jobs” is a measure of exposure, not loss — their own words are that most work is “more likely to be complemented rather than substituted,” with a base case of 6–7% of US workers displaced over roughly a decade. The IMF’s “40% of global jobs” (60% in advanced economies) is the same kind of exposure figure, and they’re explicit that about half of exposed jobs stand to benefit. The ILO keeps landing where the careful research does: generative AI is more likely to augment work than destroy it — automating tasks inside a role rather than the whole role. “Transformation, not replacement, is the real risk.”

None of which means “relax.” Transformation still means your job changes under you, some roles genuinely vanish, and sentiment runs nervous — in one WEF survey of executives, 54% expect AI to displace a large number of jobs. Holding “my field will change, and it may come for parts of what I do” next to “this is the best tool I’ve found in years” is just the honest adult posture. I hold both at once.

The divide I worry about more than the layoffs

The part that gets less airtime, and that I suspect matters more: AI is quietly splitting people into those who have access and know how to use it, and those who don’t.

It’s already measurable. As of 2025, 34% of US adults have used ChatGPT — but that’s 58% of under-30s versus 10% of people 65 and older, and 52% of those with a postgraduate degree versus 18% of those with a high-school education or less. The tool that multiplies your output is being picked up fastest by people who already held the most advantages.

Zoom out and it’s starker. AI is on track to be a $4.8 trillion market by 2033 — roughly the size of Germany’s economy — yet 100 companies, mostly in the US and China, account for 40% of the world’s private AI R&D, and 118 countries (mostly the Global South) aren’t even in the room for global AI governance. The IMF says the quiet part plainly: in most scenarios, AI is likely to widen inequality — within countries and between them.

It’s not only economists noticing. Pope Leo XIV’s 2026 encyclical Magnifica Humanitas names the same power shift: where guiding innovation once fell largely to the State, “today, however, the main drivers of development are private, often transnational, parties that are endowed with resources and the capacity to intervene that surpass those of many Governments.” When the entities shaping a civilization-scale tool answer more to shareholders than to citizens, who gets access — and who gets left out — isn’t an accident.

That’s the divide that worries me: not robots taking every job, but a widening gap between the people and places who can wield this thing and the ones who can’t — by money, by education, by geography, by simply knowing it’s worth learning. The leverage is real, which is exactly why unequal access to it is the thing I’d watch.

It’s also part of why I bother writing any of this down. The distance between “uses AI like a chatbot” and “uses it like a power tool” is widening into a real advantage — and the more people who cross it deliberately, the better.

When the switch isn’t yours

This argument got a live demonstration while I was drafting it. In June 2026 — days before I published — the US government issued an export-control directive ordering Anthropic to cut off all foreign access to its two most capable models, Fable 5 and Mythos 5, on national-security grounds. Anthropic complied in about 90 minutes, and to comply at all it had to pull both models globally — every user on Earth, paying enterprise customers and Anthropic’s own employees included. Reporting tied the urgency to a suspicion that a China-linked group had already accessed Mythos, alongside a jailbreak that could coax the model into surfacing software vulnerabilities.

Anthropic disputed the framing — calling the vulnerabilities minor and findable with publicly available models, and arguing that governments may block unsafe deployments but only “as part of a statutory process that is transparent, fair, clear, and grounded in technical facts,” which this wasn’t. You can think the government overreached or that it acted prudently and still see the same underlying fact: the most powerful version of this tool blinked out for everyone, overnight, by an order none of its users had any say in.

That’s the access divide with the abstraction stripped off. It isn’t only about who can afford the tool or knows how to use it — it’s that the switch lives somewhere else entirely. The leverage is real, and so is your dependence on a handful of companies and the governments that can countermand them. Build your work on a model and it can be pulled out from under you between one afternoon and the next. That’s not a reason to stay away; it’s a reason to hold the optimism with both eyes open.

Where it genuinely lifts me

With all that said — here’s why I’m still firmly on the optimistic side. Used deliberately, AI has been one of the highest-leverage tools I’ve ever picked up.

• Flow and focus. This one might be specific to me, but long-running, vibe-coding-style projects put me in a state of extreme focus I rarely hit otherwise. The tight loop of describe → see → correct keeps me fully engaged for hours.
• A clear mirror for internal work. When I’m doing somatic or internal work, AI is a surprisingly good mirror — it reflects my own thinking back in a structured way that helps me see it.
• A frame for interpersonal problems. This is one I lean on a lot: with a work conflict or a charged situation, I often have plenty of emotional and intuitive information but less mental clarity. AI can offer a different frame to think the issue through — grounding, informative, a lens I didn’t have. It doesn’t decide; it gives me another way to look.
• A positivity buffer. During genuinely high-stress stretches, it’s been a steadying presence — with the giant asterisk from the section above: I know it’s inclined to tell me I’m great, so I take the encouragement with a grain of salt and ask it to be straight with me.
• The everyday leverage: coding, creativity, context-building, sharpening half-formed thoughts, challenging my ideas, and yes — even mundane things like staying on track with weight loss. It’s a context enhancer and a thinking enhancer.

Notice the through-line: every one of those is me using it as a partner to think with, not a vending machine for answers.

Why people resist it — and which resistance is smart

Plenty of people are wary, and they’re not wrong to be. Pew (2025) found a real gap: 47% of AI experts are more excited than concerned, while a majority of the public (51%) are more concerned than excited. Both groups — 55% of the public and 57% of experts — say they want more control over how AI is used in their lives, and most have little confidence that regulators will get it right.

Some of that resistance is reflexive — fear of the new, fear of the hype. But some of it is exactly right, and worth respecting: skepticism about being manipulated by a confident, agreeable machine; discomfort with the energy footprint; the instinct that maybe you shouldn’t outsource your judgment to a system you don’t understand. Wariness that makes you use AI more deliberately is not Luddism. It’s the correct response to a powerful tool.

How to use it so it benefits you

Pulling the practical thread together — this is roughly my operating manual:

1. Go deep, not shallow. If your question is a quick fact, search the web. Save AI for the things where you’ll work with it over multiple turns — that’s where it earns its keep and where you stay engaged.
2. Make it disagree with you. Ask for the counterargument, the failure modes, the assumption you’re missing. The default is flattery; the value is in the friction.
3. Keep your hands on the thinking. Use it to structure and accelerate your reasoning, not to replace it. If you couldn’t summarize or defend the output, you’ve offloaded too much.
4. Verify the load-bearing stuff. It’s confidently wrong often enough that anything that matters — a fact, a number, a quote — gets checked against a real source. (This post is the proof: AI helped me research it; I verified every claim before it went in.)
5. Be a Cyborg, not a passenger. Ethan Mollick’s framing is the best I’ve seen: “Centaurs” split work cleanly between human and machine; “Cyborgs” intertwine the two, “moving back and forth over the jagged frontier.” His warning is the key bit — “on some tasks AI is immensely powerful, and on others it fails completely or subtly. And, unless you use AI a lot, you won’t know which is which.” Knowing where the frontier is is the skill.

Where I land

The people building this technology hold both feelings at once, and I think that’s the honest posture. Sam Altman writes that “the future can be vastly better than the present.” And in 2023, a who’s-who of the field — Hinton, Bengio, Hassabis, Altman, Amodei — signed a one-sentence statement: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” The same people, optimistic and sober in the same breath. Geoffrey Hinton left Google to warn that advanced systems “might take over.” Critics like Emily Bender keep insisting these models manipulate form without meaning. They’re all worth hearing.

And from outside the industry entirely, Pope Leo XIV’s 2026 encyclical Magnifica Humanitas adds the dimension the engineers and economists tend to skip — that “technology is never neutral, because it takes on the characteristics of those who devise, finance, regulate and use it.” That’s a useful correction to my own framing. I keep saying you are the variable, and you are — but the tool isn’t a blank slate either. It arrives already carrying the values of whoever built it. Both are true at once: use it well, and pay attention to who made it, and why.

My own position is unchanged by writing this — if anything it’s firmer. AI is one of the most powerful tools I’ve ever used, it’s made me more capable and arguably sharper, and it carries real costs and real failure modes that the breathless coverage on both sides gets wrong. The tool isn’t the variable. You are. Use it like a partner you argue with, keep your own judgment in the loop, ask it to tell you where you’re wrong — and it’ll make you better instead of softer.

Eyes open. That’s the whole trick.

---

A note on method: I’m not a researcher, and I used AI to help gather the studies and quotes in this post — then verified each claim and figure against its primary source before publishing, and linked them so you can check my work. That process is the argument in miniature.