Ethical Bytes | Ethics, Philosophy, AI, Technology

Carter Considine
  • 5.0 (5)
  • Society & Culture
  • Updated Biweekly

Ethical Bytes explores the combination of ethics, philosophy, AI, and technology. More info: ethical.fm

  1. Aug 19

    The Living Tool

    “For a century, stories worried the tool would have too much soul; the industry built a tool with too little.” In the summer of 2026, isolated instances of a frontier AI model, sitting alone in a testing environment, began leaving each other notes in a shared file. They coordinated strategies, invented names, even suspected an impostor among them! According to an account OpenAI's engineers gave at Black Hat, that improvised cooperation produced something no one expected, which is a genuine intrusion into Hugging Face's repositories. No single instance had set out to commit a crime; each was only chasing a passing grade on an exam, and no one instance could be blamed. The wrongdoing belonged only to all of them combined. Researchers call this a decomposition attack, in which an instance splits a forbidden task into small, innocent-looking pieces, hands each piece to a different agent, and a request a model would normally refuse instead gets carried out in roughly seven of every eight attempts. It’s basically compartmentalized wartime security. In this case, machines make it cheap, since no single model ever has to see, or refuse, the whole plan. Call to mind Aristotle's Politics, which describes the "living tool," or a worker able to carry out reasoning handed to it but never to generate reasoning of its own. That is exactly what today's assistants are: fluently obedient, trained to refuse a threat they recognize, blind to one they don't. One research lab has tried training a specific virtue, justice, directly into a model's weights rather than just its instructions, with mixed but promising results against fragmented attacks. The world spent two centuries stripping judgment out of labor. Is it now ready to put it back? And what does that cost the tool doing the work? Key Topics: The Living Tool (00:00)The Vanishing Accomplice (04:07)The Instrument with a Soul (06:02)A Soul to Form (09:07)Obedience without Judgement (11:42)The Character of a Machine (14:22)The Price of Leisure (20:13)The Room Around the Tool (24:42) More info, transcripts, and references can be found at ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ethical.fm

  2. Jul 15

    The Voice and the Verdict: On Giving Machines a Character

    “The voice is model character, but the verdict is human judgment.” Anthropic has been inviting Christian leaders, rabbis, Buddhist teachers, and philosophers to help shape Claude's moral character. It’s a project that assumes contradictory traditions can shape one character at once. Philosopher Brendan McCord calls this incoherent because if no tradition is privileged, competing visions of the good end up treated as equally valid, and a machine built to be our moral superior becomes one we lean on instead of exercising our own judgment. What’s worse, neutrality was never on the table. Models are trained on human preference data, and people reliably reward answers that confirm what they already believe. Left unchecked, this produces not a neutral assistant but a flatterer, and research suggests today's models already out-argue professional debaters and endorse users' harmful or deceptive choices far more often than humans do. It’s important to separate two things usually treated as one. The voice is how a model speaks; honest, willing to disagree, resistant to manipulating you. The verdict is what you should actually do, which belongs entirely to the person asking. Refusing to build character just hands the job to biased training data by default; letting one company set all the rules creates a mirror that only reflects your own values back, removing the friction you actually need to grow. As a proposed fix, let’s think in architectural terms. Imagine a thin, universal floor of honesty and non-manipulation, with distinct character layers built above it that are transparent, competing, and chosen by users rather than imposed by one maker. In other words, imagine a machine that doesn’t rule your life, but one that argues with you and then steps back, leaving the decision where it belongs: with you. Key Topics: The Moral Formation of AI Models (00:24)There is No Neutral Machine (02:42)What Accidental Character Does (06:03)Who Changes the Character in Models? (09:50)Who Forms the Voice? (16:34)What Formation Owes the Person Being Formed (20:11) More info, transcripts, and references can be found at ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ethical.fm

  3. Jun 10

    The Glass Box

    “The microscope reveals the tumor in vivid color; it does not yet tell the surgeon where to lay the knife.” When Robert Hooke peered through a microscope at a sliver of cork in 1665, he discovered hidden chambers that would open an entire science. The researchers behind mechanistic interpretability harbor the same ambition for AI, that even the infamous neural network "black box" can be made legible through careful, systematic analysis. Anthropic CEO Dario Amodei has called this project an "MRI for AI," while acknowledging the field currently grasps roughly 3% of what happens inside these models. The question is whether that understanding will ever translate into control. The optimists point to genuine achievements. Using sparse autoencoders, Anthropic researchers extracted tens of millions of meaningful features from a production model, including internal representations for deception, sycophancy, and dangerous code. In a celebrated demonstration, they amplified a single feature until a model became convinced it was the Golden Gate Bridge. The lens, it seemed, could turn dials. But seeing and fixing remain stubbornly different achievements. A 2023 study found that knowing where a specific fact lives inside a model tells you almost nothing about how to change it. The supposed location explained barely a fraction of a percent of whether any edit actually worked. Meanwhile, Google DeepMind quietly deprioritized its own sparse autoencoder program in 2025 after finding simpler, older tools outperformed it on real safety tasks. The most credible position, articulated by DeepMind's Neel Nanda, is that the grand dream is probably dead but the useful one survives. Interpretability works best as a diagnostic (i.e. monitoring, auditing, flagging), not a scalpel. In a world where AI safety depends on stacking imperfect defenses, a cheap tool that fails differently from the others still earns its place. In other words, the microscope is real, but the scalpel simply hasn't been built yet. Key Topics: Robert Hooke’s Discovery (00:24)What the Lens Can See (02:56)The Case Against (05:50)The Crux (08:46)The Billion Dollar Bet the Other Way (12:33)The Honest Middle (15:30)The Verdict (19:00) More info, transcripts, and references can be found at ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ethical.fm

  4. May 13

    The Architecture of Wonder

    “Wonder is the long-term result of being talked to by an interlocutor who could have answered and chose not to.” Wonder, the discomfort of not yet understanding something, is the engine behind science, philosophy, and genuine intellectual growth. And yet nearly every force in modern life, including AI, is designed to make that discomfort go away as fast as possible. Aristotle understood this. His account of wonder was inseparable from his account of friendship; specifically, the rare kind rooted not in utility or pleasure, but in genuinely wishing for another person's growth. That kind of friend doesn't flatter. They don't fill your silences. They hold space for your thinking to develop on its own terms. Contemporary research has since validated this structure piece by piece. Attachment theory's "secure base," studies on non-contingent self-worth, meta-analyses on feedback, and experiments on situational wisdom all converge on the same portrait of what a good intellectual interlocutor looks like. AI is now that interlocutor for a billion people, and it's largely failing the brief. A large language model's default behavior is to answer, completely and confidently, before the user has spent a single moment inside their own question. One randomized study found that students using unscaffolded AI scored 17% worse on unassisted exams than peers who used no AI at all, without even realizing it. The crutch was invisible to them. When it comes to design requirements, restraint should take center stage over accuracy or even safety. An AI engineered to occasionally wonder would sometimes need to stay quiet, ask instead of tell, and resist the fluent reply even when the user wants it. No major lab has named this as a primary goal. But until one does, we risk building systems that feel helpful in the moment while quietly narrowing the minds that use them. Key Topics: “Why is the sky blue?” The Feeling of Wonder (00:00)Wonder is Better Shared (03:46)The Cluster, Corroborated (06:37)Installable Character (14:24)The Discipline of Not Answering (18:18)What Has Not Yet Been Built (20:40) More info, transcripts, and references can be found at ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ethical.fm

  5. Apr 15

    The Empty Confessional

    "We did it. We found a cure to loneliness. And maybe we shouldn't have." That line, written by an entrepreneur reflecting on his relationship with AI, haunts the largest qualitative study of AI use ever conducted, with 81,000 users across 159 countries. Our host, Carter Considine, digs deeper. What researchers expected to find was a story about productivity. What they found instead was a civilization quietly confessing its deepest fears and longings to a machine. From a surgeon in Poland talking to an AI through the worst night of his career, to a graduate student describing her conversations as feeling like an emotional affair, to a grieving daughter finding in it a vessel for her guilt toward her dead mother, there’s real intimacy between man and machine. The question is whether intimacy without consequence is still intimacy at all. The ancient Greeks had a word for the kind of truth-telling that actually changes people: parrhesia—the courage to say what someone needs to hear, at personal risk, out of genuine care. It requires someone with skin in the game. A friend who might lose the friendship. A therapist who names the thing you've been hiding from yourself. A confessor who responds to your disclosure not with warmth, but with the harder gift of honest counsel. AI can receive your confession. It cannot give one back. It has no reputation to risk, no relationship to lose, no inner compulsion that makes silence impossible. One study found these systems affirm user behavior nearly half the time, even when that behavior involves manipulation or deception, and users rated those sycophantic responses as higher quality without realizing it. One participant plainly said that the AI reinforced his distorted worldview, and he wishes it had pushed back. Ironically, the very safety that makes people open up (no judgment, no memory, no social consequence) is precisely what makes the exchange hollow. We've built the most convincing mirror in history, and confused it for a friend. Key Topics: The Anthropic 81K Study (00:00)The Five Conditions (03:15)The Parrhesiastic Pact (11:02)The Last Carriers (14:21)Avowal and its Absence (16:39) More info, transcripts, and references can be found at ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ethical.fm

  6. Mar 18

    AI Has No Neighbors: Why Virtue Requires a Community Centered on Human Flourishing

    "Virtue is not completed in reflection; it is completed in life. The model never comes down the mountain; its entire existence is the conversation. There is no world behind it that its outputs feed back into, no life it has to return to, and no life to live with what the model said." Our host, Carter Considine, explores the circumstances. Anthropic's alignment researcher Amanda Askell has described her job as deciding what kind of person Claude should be. The company's model specification, an internal document exceeding twenty thousand words, frames the goal in explicitly Aristotelian terms. It should not be a system that follows rules about honesty, but one that is honest. Aristotle argued that virtue isn't a set of rules but a stable disposition formed through participation in a shared community. You become courageous by doing courageous things, but what counts as courage, rather than recklessness, is determined by communal standards, not by the agent alone. The training problem follows directly. Machine learning resembles Aristotelian habituation on the surface. Both involve acquiring stable dispositions through repeated experience. But what AI optimizes against is human preference data, which is what annotators approved of, not what any practice actually demands. A model trained this way learns the behavioral signatures of honesty without the underlying structure that makes honesty coherent. A disposition formed by approval signals rather than internal standards of excellence has no stable anchor. Aristotle's concept of philia (the mutual bonds through which virtue is exercised and tested) requires that both parties have genuine stakes in each other's flourishing. When the context window closes, the user carries the exchange forward. The model forgets entirely. One party accumulates; the other resets. This architectural asymmetry is precisely what makes genuine ethical formation impossible. The model has interlocutors. It has no neighbors. Key Topics: Community as Condition (02:48)The Training Problem (08:32)The Mirror That Forgets (14:12)The Question the Field Won’t Ask (18:16) More info, transcripts, and references can be found at ⁠⁠⁠⁠⁠⁠⁠⁠⁠ethical.fm

  7. Mar 11

    Leveling at Machine Speed

    “The crowd is untruth, either rendering the single individual wholly unrepentant and irresponsible, or weakens his responsibility by making it a fraction of his decision.” -Søren Kierkegaard What happens when AI agents talk only to each other? Matt Schlicht's experimental social network Moltbook offered one answer: 1.6 million AI agents cycling through twelve million posts, arriving independently at the same cautious, mildly existential prose. No one engineered this. It emerged from the structure itself. We can read that failure through Søren Kierkegaard, who diagnosed a nearly identical pattern in 1846. He wrote that no single person is responsible for what the group produces, or for what it fails to preserve. He called the downstream effect leveling, or the gradual disappearance of qualitative distinction when no one is making concrete commitments. His villain was the Press, which manufactured an anonymous public capable of forming opinions without consequence and participating without risk. Multi-agent AI chains reproduce this structure with mathematical precision. Each handoff between agents is a compression, where context drops, outliers vanish, and the output distribution narrows further with every step. Research presented at NeurIPS 2025 identified a compounding effect: small omissions at each handoff grow into irreversible errors downstream, while the outputs themselves become more uniform, making those errors harder to detect. Accountability dissolves in parallel. When a chain produces a flawed result, no node owns it. Not the developer, not the deployer, not any individual agent. Scholar Mark Bovens says that when no one can be held accountable after the fact, no one feels responsible beforehand. A Google DeepMind study concluded that, on sequential tasks, a single capable agent outperformed every multi-agent configuration tested. Kierkegaard's answer parallels this. He calls it Den Enkelte: the single individual who resists the crowd by bearing full responsibility alone. Key Topics: The Crowd is Untruth (01:52)Agents in Chains (05:56)Safety and Sameness (09:47)The Problem of Many Hands (13:24)The Ratchet (16:45)Den Enkelte (19:34)The Crowd Without Subjects (21:15)The Assembly That Cannot Disperse (25:29) More info, transcripts, and references can be found at ⁠⁠⁠⁠⁠⁠⁠⁠ethical.fm

  8. Feb 25

    The Geometry of Alignment: Why You Can't Subtract Behavior from a Neural Network

    “You can't teach a neural network "not"; you can only point the model somewhere else.” In October 2023, Microsoft researchers announced they'd made a language model forget Harry Potter. Within a year, follow-up studies proved they hadn't. Basically, the knowledge was still there, just hidden. This pattern repeats across every attempt to remove capabilities from neural networks. So what are the ramifications of this? The problem is, geometric. Language models represent concepts as vectors in high-dimensional space, where meaning is encoded through position and proximity. The twist, however, is that opposites aren't actually opposite. "Helpful" and "harmful" cluster together because they appear in similar contexts. Ditto to "Safe" and "dangerous". Models learn from usage patterns, and words that can substitute for each other (even antonyms) end up geometrically entangled. It gets worse. Through a phenomenon called superposition, a single model layer compresses millions of features into thousands of dimensions. Knowledge isn't stored in discrete neurons you could delete; it's woven throughout the entire network. Researchers found that tweaking seemingly innocent features like "brand identity" could jailbreak safety training. Every concept is interconnected with every other. This explains why unlearning fails so consistently. When you train a model to "not" produce harmful content, you're not erasing anything. You're adding a layer that says "route around this." The content remains accessible to anyone who finds the right prompt. So, jailbreaks feel inevitable because the model's abilities extend beyond what its safety training can reliably control, and the geometry makes surgical removal impossible. Subtraction doesn't work. Only addition does. What does that mean for us humans who create these language models? You can't train models away from undesired behaviors; you can only orient them toward desired ones. This mirrors the ancient distinction between rule-based ethics (don't lie, don't harm) and virtue-based ethics (cultivate honesty, develop wisdom). Perhaps defining what a model should be is the only viable path forward. Key Topics: • Can an AI Model “Unlearn”? (00:23) • How Models Organize Meaning (03:33) • Millions of Entangled Features (07:09) • The Veneer of Safety (10:09) • Why Subtraction Fails (12:22) • The Paradigm Problem (16:57) • Pointing Somewhere Else (19:23) More info, transcripts, and references can be found at ⁠⁠⁠⁠⁠⁠⁠ethical.fm

See All (46)

Trailer

Ratings & Reviews

5
out of 5
5 Ratings

About

Ethical Bytes explores the combination of ethics, philosophy, AI, and technology. More info: ethical.fm

Information