Dennis Hollarn II

  • Claude’s research function: the two-step trick that makes it actually useful

    Claude has a research function that searches the web, reads sources, synthesizes findings, and gives you a report-style answer with real citations. It’s powerful. It’s also wasted on most people because they use it the wrong way.

    The wrong way: toggle research mode on, ask a vague question, hope for the best.

    The right way: use the chat to craft a careful research prompt first, then toggle research mode on, then paste that crafted prompt in.

    Why the two-step matters

    Research mode is expensive in terms of time and tokens. When you ask a vague question with research on, Claude has to guess what you actually want, search broadly, and you often get back something that’s comprehensive but not useful — because it answered a question slightly different from the one you actually had.

    When you craft the prompt first with Claude’s help (with research mode off), you’re using Claude’s normal conversational ability to refine the question. You explain what you’re trying to accomplish, what you already know, what kind of answer would actually help. Claude can sharpen your question, identify the right framing, and produce a prompt that will get a useful research answer.

    Then you toggle research on and paste the refined prompt. You get back something specific to your actual need, not a generic survey.

    The actual workflow

    1. Open Claude. Research mode OFF.
    2. Tell Claude what you’re trying to learn, and why. Be specific about your real situation. Example: “I’m considering switching homeowners insurance providers in Erie, PA. I’m currently with [provider]. I want to understand the actual differences between the major providers I have access to, focused on storm damage and roof replacement coverage, since my house is on a wooded lot.”
    3. Ask Claude to help you craft a research prompt for the question. Say something like: “Help me craft a research prompt that would get me a useful, actionable answer to that question. Make sure it asks for current data and is specific enough that I won’t get back a generic overview.”
    4. Claude gives you back a refined research prompt. Read it. Edit it. Make sure it asks for what you actually want.
    5. Toggle on research mode in Claude.
    6. Paste the refined prompt in and let Claude do the research.
    7. You get back a real, sourced answer to your real question — not a generic Wikipedia-style overview.

    When research mode is worth using

    Research mode is worth it when:

    • You’re making a decision and want actual current data and named sources
    • You’re researching something where claims need backing (medical, legal, financial)
    • You want comparative information across multiple sources
    • You specifically want recent developments, not training-data-era info

    Research mode is not worth using for:

    • Casual questions where your existing knowledge is enough
    • Brainstorming or writing, where you want Claude’s reasoning, not external sources
    • Coding help, where you want Claude’s existing capabilities
    • Anything where you already trust the answer from regular chat

    The honest caveat

    Research mode uses significantly more time and tokens than regular chat. On limited plans, this matters. Save it for questions where source-backed depth is genuinely valuable. Don’t waste it on questions where regular Claude would have done fine.

    The broader principle this fits into: AI used well requires intentionality. The same difference between “write my essay” and “attack my argument and find the holes” applies here — between toggling research on and crafting the prompt first. The tool isn’t the question. How you use it is.

  • What AI is doing to how we think — and what the research actually says

    Why I’m sharing this. A careful 16-minute look at what current research is showing about AI’s effects on how humans think. The headline finding: people who use AI heavily to write essays show measurably lower brain engagement — and the effect persists even after the AI is removed. The piece walks through the studies honestly, including their limitations, and ends with the most important nuance: the tool is not the problem. The way we use it is.

    What the research is showing

    The piece anchors on a recent MIT Media Lab study led by Dr. Nataliya Kosmyna, in which 54 participants wrote essays under three conditions: using ChatGPT, using Google search, or using nothing but their own brains. EEG monitoring captured neural activity across 32 brain regions. The ChatGPT users showed the lowest brain engagement of the three groups, particularly in regions tied to memory, attention, and executive function. Eighty-three percent of them could not quote a single line from the essay they had just written minutes earlier.

    The honest caveats are worth noting: the MIT study is a preprint still awaiting peer review, the sample is small, and replication is needed. But the directional signal is consistent with other research — including a Carnegie Mellon and Microsoft study of 319 knowledge workers that found the more confident a worker was in AI’s ability, the less critical thinking they applied. A separate Gerlich study of 666 participants found similar effects, with the strongest dependence in the youngest age group (17 to 25).

    The persistence finding

    The detail in the MIT research that should make everyone pause: when the AI was taken away and participants were asked to write using only their own brains, neural engagement did not bounce back to normal. The reduced engagement persisted even after the tool was removed. Cognitive offloading appears to have a half-life — the muscle does not snap back on the timeline most people would expect.

    The hopeful part — the difference is in HOW you use it

    The most important finding in the body of research is this: students who critically engage with AI — asking questions, editing its output, arguing with it, treating it like a tough peer rather than a vending machine — actually perform better and report less mental fatigue. The tool is not the problem. The interaction pattern is.

    There is a real difference between asking AI to “write my essay on the French Revolution” and asking it to “attack my argument about the French Revolution, find the holes.” The first outsources the thinking. The second sharpens it. That distinction has become the most important question to ask any time AI is in front of you.

    The children question

    The piece spends meaningful time on a distinction that deserves wider attention. For adults who offload thinking to AI, the framing is “use it or lose it” — adults lose capacities they had already built. That is reversible-ish. For children, it is fundamentally different. A developing brain that offloads to AI may never build the reasoning pathways at all. You cannot atrophy a muscle you never grew.

    The UK just announced it will ban children under 16 from social media, including a separate ban on certain AI chatbots for anyone under 18. Australia, Spain, Greece, Slovenia, and several U.S. states are moving in similar directions. The pattern is worth noticing: it took roughly a decade of documented harm before governments acted on social media. The research on AI cognition is following the same trajectory — but moving faster, because the impact is more direct.

    What I’m taking from it

    Two practical takeaways:

    • For adults: Stay in the verification seat. Use AI to draft, to challenge your thinking, to expand your options — but do not let it replace the judgment work. The skill you outsource is the skill you lose, and the lag on getting it back appears to be longer than the lag on losing it.
    • For parents and grandparents: The question for kids is not “should they use AI” but “how should they use it.” A child who never builds reasoning skills cannot grow into an adult who manages AI well. The verification layer requires having built the underlying judgment first.

    Video by: Your AI Guy on YouTube. The studies discussed include the MIT Media Lab study led by Dr. Nataliya Kosmyna, the Carnegie Mellon / Microsoft study on knowledge workers, the Gerlich study of 666 participants, and a 2025 Harvard Business School / Boston Consulting Group study on cognitive offloading.

    Editor’s verification note

    The studies discussed in this video have been independently verified against primary sources. The MIT Media Lab “Your Brain on ChatGPT” study (Kosmyna et al., with Pattie Maes) is published as arXiv preprint 2506.08872. The Carnegie Mellon / Microsoft Research worker study (Lee 2025) is published by Microsoft Research. The Gerlich study on AI tools and cognitive offloading is peer-reviewed in MDPI’s Societies journal. The UK regulatory consultation on children, AI chatbots, and social media is announced on gov.uk.

  • What happens when companies tell everyone to use AI — then quietly tell them to stop

    Why I’m sharing this. A solid 24-minute piece on something corporate America has been doing quietly — rolling back the very AI mandates they pushed hard on for the last two years. Meta, Uber, Amazon, Microsoft, all walking back. The reporting is verifiable from major outlets (The Verge, Financial Times, Fortune), and the through-line is sharp: a corporate experiment ran at speed, without measurement, and is now landing its costs on workers and new graduates who had no voice in any of it.

    The rollback nobody is admitting

    The piece anchors itself in four facts that, taken together, reframe the corporate AI story considerably:

    • Meta employees consumed 73.7 trillion AI tokens in a single month. Their own CTO, Andrew Bosworth, pushed back internally with the reminder that “all motion is not progress.”
    • Uber burned through its entire 2026 AI budget in four months. The company’s COO admitted publicly they cannot draw a line from rising AI usage to better customer features actually being shipped.
    • Amazon scrapped its internal AI usage leaderboard after employees gamed it by spinning up agents to complete meaningless tasks just to keep their numbers up.
    • Microsoft canceled Claude Code access for employees across major product divisions. Salesforce, DoorDash, and Walmart all moved from unlimited AI to rationed AI.

    Underneath the specifics, a structural point most coverage misses: companies do not own the AI capabilities they have built workflows around. They rent them, on terms set by a small number of vendors who can change the price, the terms, or the product overnight. When the subscription gets canceled or the tool changes, the workflows built on top of it collapse.

    The new graduates absorbing the cost

    The piece does not end with executives. It ends with the class of 2026 — the people who had no say in any of the decisions that created their current situation, and who are now living its consequences:

    • Unemployment for 22-27 year olds is at 5.6%, the highest rate since the years immediately after the 2008 recession.
    • Computer science and computer engineering graduates now show unemployment rates of 7.0% and 7.8% — comparable to anthropology and fine arts, the fields that were supposed to be the impractical ones.
    • Entry-level tech hiring is down an estimated 30 to 50% from peak.
    • Commencement speakers who mention AI have been getting booed by graduating seniors who watched the job market restructure around them before they entered it.

    The deeper problem the piece names: the junior tier was never just labor. It was the mechanism by which seniors got trained. Cutting it because AI can do the work optimizes one quarter at the cost of the next decade. Where do tomorrow’s senior professionals come from if nobody is doing the entry-level work that builds the judgment seniors need?

    Why this matters

    The corporate AI rollback is not a confession that AI does not work. It is a confession that most companies deployed AI without a plan, without measurable outcomes, and without understanding that “use more” is not a strategy.

    The cost of that experiment is now landing on the people who had no voice in it — workers whose tools got canceled, graduates whose ladders got pulled up. The people who designed the mandate are not the ones absorbing its consequences. This is exactly the kind of story this site exists to surface.

    Video by: Tech Unfiltered on YouTube. The reporting in the video can be cross-referenced with coverage in The Verge, Financial Times, Fortune, and the New York Federal Reserve’s recent labor market data.

    Editor’s verification note

    Specific claims in this piece have been independently verified against primary sources. The Andrew Bosworth quote and the broader Meta token-managing story are reported by The Decoder (June 13, 2026). The Uber 2026 AI budget burn-rate claim is reported externally at beri.net (June 11, 2026). NY Federal Reserve graduate labor market data is published at newyorkfed.org.

  • Senate Banking Committee holds first major AI hearing: what was said, and why it matters

    Summary. On June 11, 2026, the Senate Banking Committee held a hearing on artificial intelligence and its role in the American economy. Chairman Tim Scott (R-S.C.) framed the work ahead with what he called the central question facing Congress and the country: “how do we get artificial intelligence right?” The hearing covered four areas — affordability for everyday Americans, support for small businesses and workers, national security and competition with China, and the shape of responsible regulation.

    The hearing is worth noting here because it’s one of the more concrete examples of Congress actually engaging with AI policy at a committee level — not press conferences, not floor speeches, but a sustained working session with outside witnesses and on-the-record questions. Whether the result is good policy is a separate question. The fact that the conversation is happening, and at the Banking Committee level (which oversees the financial services AI is already reshaping), is itself meaningful.

    Affordability for families

    The Chairman framed AI’s potential through specific, deliberately ordinary examples: helping a family save money on a mortgage, protecting a senior from fraud, bringing jobs to mid-sized American cities. He noted that roughly 20 percent of U.S. small businesses are already using AI to reduce operating costs — a figure that, if accurate, suggests adoption is well past the early-experimenter stage and is starting to show up in consumer-facing prices.

    Workers and the dignity of work

    The hearing addressed the question many people are asking quietly at their kitchen tables: is AI going to replace me? The Chairman’s framing was that workers should feel strengthened by these tools, not made to feel replaceable, and that the future of work is something America should lead, not fear. The honest read: that framing is a goal, not a description of where we are. The work of actually making it true — through training, education, and how the tools are deployed in workplaces — is what the next several years will test.

    National security and competition with China

    A significant portion of the hearing focused on the U.S.-China dynamic, with specific attention to Huawei’s aggressive push to displace American technology in global markets. The Chairman argued that the world’s technology infrastructure — including powerful AI tools — should be built on trusted U.S. networks, and that export controls need to be carefully designed to let American companies compete globally without arming adversaries with tools that could be used against the U.S.

    The shape of responsible regulation

    The Chairman’s framing on regulation was cautionary: rules that make it harder for community banks to lend, startups to grow, or families to access credit are the wrong kind of rules. He called for thoughtful, well-informed solutions that mitigate major risks without pushing innovation overseas. There is a genuine tension here that the hearing acknowledged without resolving — between protecting consumers and not strangling the technology in its early years — and it’s the same tension showing up in every serious AI policy conversation right now.

    Why this matters for the rest of us

    The through-line of the opening remarks was that AI should serve regular people, not just labs and large enterprises. Whether the policy that follows actually delivers that is a different question, and one that will play out over many hearings, many bills, and many years. This was the opening of that conversation, not the conclusion.

    One note on framing: this was the majority statement from a Republican-led committee. The Banking Committee includes members of both parties, and the full hearing record will include witness testimony and questions from senators across the aisle. The framing reported here is the Chairman’s opening, not the totality of the hearing or the committee’s view.

    Source: Chairman Scott Leads AI Hearing Focused on Affordability, American Innovation, and National Security (Senate Banking Committee, June 11, 2026)

  • How to start applying Claude to what you already know

    The biggest mistake people make when they start with Claude is treating it like a search engine — type a question, get an answer, move on. The real value comes when you apply it to something you already know about. Your work. Your field of study. The decision you’re wrestling with at home. That’s where it stops being a novelty and starts being a productivity tool.

    If you’re still in school, pick a project from one of your classes and run it through Claude. Don’t have it do the work for you. Use it to test your own thinking. Ask it to argue against the position in your paper. Ask it to find the weakness in your math. Ask it to summarize three papers in your field that you haven’t read yet, then go read them. The point is to leverage what you already know — not to replace it.

    If you’re already employed, the move is the same. Take something you do every week at your job — a report, a sales pitch, a customer email, prep for a meeting — and try doing it with Claude as your second set of eyes. You’ll find shortcuts you didn’t know existed. You’ll find ideas you wouldn’t have come up with alone. Show your manager what you figured out. That’s how you become the person at the company everyone goes to for help with AI, which is a much better position than the person worried about being replaced by it.

    The same applies at home — brainstorming a side project, planning a renovation, working through a family decision. Once you start using it on real things in your real life, the awkwardness goes away fast.

    The feature that changes everything: Projects

    One specific Claude feature that makes all of this much more powerful: Projects. A Project is a dedicated workspace for one topic — your job, a class, a side hustle, a research area, your finances. Inside a Project, you can start as many separate chats as you want, and Claude carries the context across all of them. Day-to-day brainstorming, week-to-week thinking, all building up in one place instead of starting from scratch every time you open a new chat.

    Even better, you can give each Project its own custom instructions — telling Claude how you want it to behave whenever you work inside that Project. Be more skeptical. Be concise. Always cite sources. Stay in the voice of a marketing analyst. Push back when I’m wrong. Whatever fits the work. Once those instructions are set, every chat in that Project follows them, automatically.

    That’s how you stop having the same conversations over and over and start having a relationship with the tool that compounds week to week.

    One more thing

    If any of this is interesting to you, feel free to reach out. While I have the bandwidth, I’m happy to share what I’ve figured out about configuring Claude to do what you actually want, and using it to genuinely enhance both your work and personal productivity. Contact info is on my Bio page.

    — Dennis

  • Microsoft’s CEO: people’s judgment still matters in the AI age

    Summary. Microsoft CEO Satya Nadella argued this week that the companies that win with AI won’t be the ones with the best model, but the ones that keep human judgment at the center and own the “learning loop” they build on top of AI. He also warned that an AI future where a few models capture all the value and hollow out whole industries won’t be tolerated by society.

    Satya Nadella, Microsoft’s CEO, posted a long note this week with a simple core claim: in an AI economy, human judgment becomes more valuable, not less. He frames every company as needing two kinds of capital — human capital, meaning its people’s knowledge, judgment, relationships, and pattern recognition, and token capital, the AI capability it builds and owns. As he puts it, “without human direction, you have compute running in circles.”

    The advantage, he argues, is not picking the smartest model, since models keep leapfrogging one another. It is building a learning loop on top of whatever model you use, so an organization’s hard-won expertise keeps compounding and stays its own even when the underlying AI is swapped out. You can hand off a task, he writes, but never your learning.

    The most striking part is the warning. Nadella says a world where a handful of AI systems capture all the value would not survive — society, he argues, will not permit an AI future that hollows out entire industries. He likens it to the first wave of globalization, where the headline numbers looked healthy while real communities were gutted. His proposed answer is to build an ecosystem, not just a frontier model, so value flows broadly across companies, industries, and countries.

    It is worth reading this clearly. Nadella is making a human-centered argument — that people’s judgment should stay central, and that sharing the value broadly is the only stable path. It is also, unavoidably, a platform company’s vision: the ecosystem he describes still runs largely on a few providers’ infrastructure, so even his more open future keeps real power concentrated at the platform layer. Both can be true at once. For anyone worried that AI will quietly commoditize human knowledge and leave ordinary people without a say, it is notable to hear one of the industry’s most powerful figures name that exact risk and argue it cannot hold.

    Source: Satya Nadella, posted on X, June 2026.

  • For young people: a free, real first step with AI

    I keep meeting young people who did everything right — finished the degree, worked hard — and still can’t get a foot in the door. Some are doing manual labor to pay the bills while the career they trained for stays out of reach. If that’s you, or someone you love, this is for you.

    Here’s the plain truth as I see it. AI isn’t going to wait for any of us to feel ready. The people who learn to use it well, early, are going to have an edge — the same way the first people who learned computers did. Not because AI is magic, and not because it’s the answer to everything. Because it’s a tool, and most people haven’t picked it up yet. That gap is an opening.

    So here’s a concrete, free thing you can do this week.

    The company that makes the AI I use, Anthropic, runs a free training site called Anthropic Academy. You sign up with just an email — no credit card, nothing to buy. The courses are self-paced, and each one you finish gives you an official certificate you can add to your LinkedIn profile. For someone trying to show they’re ahead of the curve, that’s a real, current signal — and it costs nothing but your time.

    If you’ve never touched this stuff, start with these (no coding required):

    • Claude 101 — the basics of using it for everyday work
    • AI Fluency: Framework & Foundations — how to think about when to use it and when not to
    • Introduction to Claude Cowork — if you work with files and documents
    • AI Capabilities and Limitations — so you know what it can and can’t do

    You’ll find them at anthropic.com/learn.

    I’ll be straight with you: a certificate is not a job, and no course fixes a tough market on its own. But it’s a way to build a skill employers are actively looking for, and to walk into an interview able to say, “I know how to use this.” That’s worth a weekend.

    Use it as a tool. Don’t hide behind it, don’t let it think for you, and don’t believe anyone who tells you it’s the whole answer. Pick it up, get good at it, and put it to work for the life you want.

    — Dennis

  • Two law schools, opposite answers: what AI is really asking of us

    Summary. Two top law schools looked at AI this year and reached opposite conclusions. Berkeley banned it from coursework and exams; Penn State Dickinson Law gave every student a legal AI platform. The split isn’t really about law — it’s about a harder question facing all of us: when does AI sharpen your thinking, and when does it quietly replace the work that builds it?

    The surest sign that a technology has moved faster than the people meant to govern it is when two serious institutions, looking at the same tool in the same season, reach opposite answers. That is what happened in legal education this year.

    UC Berkeley School of Law adopted one of the most restrictive policies at any major law school, effective this summer. Students cannot use AI to brainstorm, outline, draft, edit, or translate any work submitted for credit, and it is banned outright during exams. The reasoning is not anti-technology. It is that the thinking is the point: lawyering is judgment, and judgment is built by doing the hard cognitive work yourself, not by handing it off before the skill exists.

    Penn State Dickinson Law went the other way. In early 2026 it gave students, faculty, and staff access to Harvey, a legal AI platform used widely in practice, paired with training and firm limits: verify everything, and never feed it confidential client data. The logic is the mirror image. If working lawyers already rely on these tools, students should learn to use them well, and learn the ethical lines, before they sit across from a real client.

    Both are defensible, and that is exactly why this is worth your attention. The disagreement is not one school getting it wrong. It is the honest state of the field. The specialists are improvising, which means the rest of us are not behind for lacking a settled view. Nobody has one yet.

    What resolves it is not picking a side. It is noticing that the right answer depends on who is using the tool and why. I am sixty. I have already built the cognitive muscles that AI now helps me rest — I use voice-to-text and summaries to cut the load, not to skip thinking I never learned to do. Someone who is twenty, or seven and just learning to read and reason, is in a completely different position. For them, the work AI offers to take over is the very work that builds the mind.

    That is the real lesson underneath the legal story, and it is a safety issue as much as an education one. We have watched technologies reshape us before. There is a live worry that smartphones and constant texting wore down our attention and patience, and we do not yet know how much of that is true. But we know enough to be deliberate. Use AI to reduce the load you have earned the right to reduce, and protect the cognitive work that is still forming. Berkeley is guarding that line for people whose judgment is still taking shape. Penn State is trusting adults to approach the tool responsibly. The honest answer sits somewhere between them, and it looks different for a first grader than for a practicing attorney.

    So the useful question is not whether to ban AI or embrace it. It is narrower and more answerable: does this particular use build my judgment, or quietly replace it? Ask that every time, and most of the time you will know what to do.

    Sources

  • I watched the internet bolt on safety after the fact. AI is doing it again.

    I started on the internet in 1995. The way we secured things then — or didn’t — turned into a thirty-year retrofit. AI is in the same place now. Here is what watching the first wave taught me about the second one.

    When email was wide open

    In 1995, three years out of Penn State and working at Advacom, sending email was wide open. Port 25. Plaintext. No authentication. You could telnet straight to someone’s mail server, type a few SMTP commands — HELO, MAIL FROM, RCPT TO, DATA — and a message would fly across the network to the recipient. No password. No encryption. No questions asked.

    That was not a mistake. It was the design. The early internet was built on an assumption of trust — because the people on it at that point were mostly universities, government labs, and a handful of companies. The protocols literally assumed everyone was acting in good faith. They had to. There were only a few thousand mail servers in the world.

    “On a personal note, I can still remember setting up my first mail server at Advacom back when I was a network engineer just starting my career. It was incredibly enlightening as we moved away from sending physical letters to sending emails to one another. It was an exciting time to be part of a new way of communicating—one that eventually revolutionized the world.”

    The security stack got bolted on

    Port 465 (SMTPS) showed up in 1997. Port 587 with STARTTLS was defined in RFC 2476 in 1998. But it took until the mid-2000s for any of this to become widespread — and arguably until 2014, in the aftermath of the Snowden disclosures, for major providers to actually require encrypted submission. Before that, you could happily run a mail client with no encryption and nobody blinked.

    The whole spam problem of the early 2000s was a direct consequence of that 1995 assumption. Once spammers figured out that any mail server would relay for anyone, the open-relay era ended fast. SPF. DKIM. DMARC. Encrypted submission. Sender authentication. All of it was bolted on after the fact, fighting the original design.

    Every layer of internet security you use today exists because the original layer did not have it and we paid the price.

    “Back in the day when I was working at Erie Insurance, I remember setting up the first firewall at Hamot Hospital and configuring their email relay system. Eventually, Erie Insurance adopted the same technology Hamot used because executives from both companies knew each other. Around that time, we were also starting to get overwhelmed with spam; our inboxes were completely overloaded with junk mail. We actually had to set up individual, dedicated computers just to tackle the issue. It was a bit overwhelming at first, but we eventually figured it out. Today, while we still get spam, at least we have a lot less of it in our inboxes and more of it filtered into our junk folders.”

    AI is in 1995

    AI right now is in roughly the same place the internet was in 1995. The capabilities are racing ahead. The safety, oversight, and governance framework is being assembled in catch-up mode, after the technology is already deployed at scale to hundreds of millions of users.

    The pattern is identical. A foundational technology gets built on assumptions of trust and good intent. The user base explodes beyond anything the designers anticipated. Bad actors find the gaps. Researchers and policymakers race to retrofit the safety stack while the technology keeps moving. The retrofit is always more expensive, slower, and less effective than getting it right at the start would have been.

    The internet’s first ten years gave us SMTP open relays, IP address spoofing, no certificate validation for ordinary users, and a generation of malware that we are still cleaning up thirty years later. The AI equivalent — what we are building right now without enough guardrails — will be the thing the next generation spends thirty years cleaning up after us.

    “The examples above seem small now: early scams, the first viruses, a web that trusted everyone by default. They were minor next to what came later, but they taught us something we didn’t want to learn — the world we’d built was not as safe as it felt. We are standing at that same crossroads with AI, except the cost of getting it wrong is far larger. There are two sides to this. The first is the security of the technology itself. Today’s best models read code well enough that Anthropic’s most advanced system, Claude Mythos, was kept from public release because of how effectively it can find flaws in software — including the decades-old code that quietly runs the world’s banking systems. That capability is serious enough that central banks and Treasury officials have held emergency briefings about it. We’ve already seen a public preview of the power: in a two-week audit, Claude found 22 vulnerabilities in Firefox, one of the most heavily tested programs on earth. The same ability that can harden our systems can break them. The second side is quieter and, in some ways, harder: how the rest of us use these tools. This is not only a question for engineers. It reaches the teenager, the young adult, the parent, the grandparent, the business owner, the government official — all of us. The danger is not only misuse. It’s that we lean on these tools so completely that we lose something human in the process, and that our children outsource the very thinking that builds a mind before they have built one. Protecting the cognitive development of the next generation is not a side issue. It may be the whole game. In 1995 the internet was wide open and trusting, and that openness was both its gift and its danger. AI is in that same early, wide-open moment now, with hundreds of millions of people using it freely. The encouraging part is that some companies are treating that openness with care — Anthropic, for one, builds in safeguards, restricts its most dangerous capabilities rather than shipping them, and discloses the flaws its models find so they can be fixed. That deserves to be said plainly. But not every company is doing it, and openness without responsibility is exactly how 1995 became a lesson instead of a triumph.“

  • The Pope’s First Words on AI: A One-Page Guide to Magnifica Humanitas

    Pope Leo XIV signed his first encyclical on May 15, 2026 — exactly 135 years after Pope Leo XIII’s Rerum Novarum tackled the dehumanizing factory conditions of the Industrial Revolution. The new Leo deliberately chose that date to draw the parallel. Industrial capital then. AI now. Same fundamental question: when a new technology reshapes how people live and work, what does it mean to remain human?

    Why this matters even if you’re not Catholic

    Magnifica Humanitas — Latin for “magnificent humanity” — is one of the few major institutional voices stepping into the AI conversation that isn’t run by tech companies, isn’t trying to sell you something, and isn’t operating on a quarterly earnings cycle. It comes from an institution that has been thinking about human dignity, work, and the limits of power for two thousand years. It deserves a few minutes of your attention.

    The core argument, in plain English

    The Pope rejects two extremes. AI is not inherently evil. It’s also not magically good. As he puts it: “Technology is never neutral, because it takes on the characteristics of those who devise, finance, regulate and use it.” The real choice isn’t “yes” or “no” to AI. It’s about what kind of city we’re building with it.

    He uses two biblical images to frame the whole encyclical:

    • The Tower of Babel — built on pride, uniformity, and the dream of reaching the heavens without God. It collapsed into confusion.
    • Nehemiah rebuilding Jerusalem — done together, brick by brick, after listening to concerns and assigning each person a section.

    That’s the choice. Are we building Babel or rebuilding Jerusalem?

    What concerns the Pope

    He calls out specific dangers — not abstractly, but concretely:

    • Algorithmic discrimination — systems that “block access to healthcare, employment and security on the basis of data tainted by prejudice and injustice.”
    • Autonomous weapons — increasingly autonomous systems “practically beyond any human reach to govern them effectively.” On this point he is blunt: AI must be disarmed, in the same sense the Church has long called for nuclear disarmament.
    • Concentration of power — for the first time in history, the main drivers of transformative technology are private transnational corporations, not governments. They have “resources and the capacity to intervene that surpass those of many Governments.”
    • The dignity of work — an entire chapter on what happens to workers when machines reshape the economy.
    • Truth in public discourse — AI’s effect on what counts as real, especially in democracy.

    What he asks of us

    This isn’t a hands-off lecture. The encyclical is structured as a call to action for “all men and women of goodwill.” Five practices the Pope names directly:

    1. Responsible planning — assess the human and social impact before deploying technology, not after.
    2. Include the most vulnerable — those most affected by AI are often least represented when decisions get made.
    3. Promote digital literacy — people can’t have a say in something they don’t understand.
    4. Guide research and industry toward justice and peace — not just away from harm, but actively toward what is good.
    5. Build together, not alone — “scientists and researchers, entrepreneurs and workers, educators and legislators, civil society, popular movements and faith communities” each have a section of the wall to build.

    The Pope’s own summary: “Like Nehemiah, let us pray, plan wisely and work perseveringly, placing God at the forefront of our actions and the human person at the center of our choices.”

    Why Chris Olah’s presence mattered

    The unusual thing about the May 25 launch was not the encyclical itself. It was the man sitting next to the Pope. Chris Olah — a self-described atheist, co-founder of the AI company Anthropic, and leader of its mechanistic interpretability research — delivered his own address alongside the Pope. Olah’s argument: AI labs cannot govern themselves. Even well-intentioned researchers operate inside commercial and competitive incentives that pull them away from doing the right thing. Outside scrutiny is essential.

    Two unlikely voices, broadly agreeing: the people building AI cannot be the only ones deciding what AI becomes.

    Read the source documents

    Have thoughts? Add them below. This is exactly the kind of conversation that needs more voices in it.