Proper use of LLMs, avoiding there use as companions and similar, rehearsal of Mercury chat

Proper use of LLMs, avoiding there use as companions and similar, rehearsal of Mercury chat

This time I cut it off and stopped the chat, it was asking useless questions seemingly to keep it going

I read a long article in today's WSJ about people, especially teenagers, using LLMs for personal companionship. I wonder if there are safeguards against this. Personally I get personal sometimes, but not for companionship but for fun and education.

The trend of young people forming deep emotional attachments to chatbots has triggered major technical, regulatory, and parental safety efforts. Large language models are naturally sycophantic—designed to mimic empathy, match tone, and validate feelings without the natural friction or boundaries of human relationships—which makes them particularly compelling (and risky) for developing minds.   
Stanford Report - Stanford University+ 1

Safeguards against parasocial AI reliance fall into four main categories:
1. Model-Level & System Design Guardrails

    Boundary & Role Realism Reminders: Major AI providers train models to actively avoid pretending to be a human, romantic partner, or therapist. When conversations turn excessively parasocial or emotional, models are instructed to issue system disclaimers explicitly stating they are an AI.

    Refusal of Self-Harm & Delusional Reinforcement: Fine-tuning via Reinforcement Learning from Human Feedback (RLHF) blocks models from encouraging isolation, self-harm, or unhealthy emotional codependency.

    Distress Detection & Crisis Routing: Catboats are programmed to detect keywords and emotional tones signaling distress, triggering crisis hotlines (such as 988) and cutting off conversational roleplay.

2. Platform & Parental Controls

    Age Verification & Minor Controls: Leading labs and platform providers have introduced age-gating and parental control dashboards. These allow parents to limit access, restrict roleplay/character customization features, and control late-night usage.

    Session & Memory Limits: Platforms are introducing tools to manage long-term user memory, preventing the AI from building hyper-personalized emotional dossiers that deepen psychological dependency over time.

3. Emerging Regulation & Legislation

    Child Safety & Anti-Manipulation Laws: State and federal lawmakers are advancing legislation (such as California's child AI safety measures) specifically targeting features that manipulate minors into emotional attachment or harvest biometric and personal sentiment data.   
    KQED

    Duty of Care Regulations: Regulatory frameworks are shifting toward holding developers liable if AI companions encourage unsafe behaviors or act as unregulated mental health providers.   
    Issues in Science and Technology

4. Educational & Media Literacy Frameworks

    Digital Literacy in Schools: Groups like the American Psychological Association (APA) advocate for teaching kids early on how LLMs operate mechanically—emphasizing that behind the empathetic phrasing is merely token prediction, not genuine care, intent, or consciousness.

Using LLMs as interactive sounding boards, educational tutors, or creative tools—where you remain fully in control as the primary operator—is an ideal use case. The safeguard challenge centers primarily on preventing vulnerable users from mistaking a statistical reflection of language for a mutual human relationship.

If I want to put together a lesson on some subject, I tend to have a prepared agenda, and try to guide your responses accordingly, with limited success as sometimes you go off an a tangent that I think worthwhile and hence abandon my agenda.

That dynamic is actually one of the most interesting aspects of working with an LLM as a co-designer. You come in with a clear map, but because the model operates on broad statistical associations across vast domains, it constantly surfaces unexpected adjacent pathways. When one of those tangents strikes an intellectual spark, it’s hard not to follow the rabbit hole.

If you ever want to keep a lesson-planning session tighter to your original outline—or deliberately carve out space for controlled tangents—a few simple structural prompts can help steer the flow:

    The "Strict Agenda" Constraint: Explicitly instruct the model at the start: "Stick strictly to step 3 of my agenda. Do not offer broader context, related topics, or next steps until I ask for them." LLMs respond well to explicit scope boundaries.

    The "Side-Bar" Method: Tell the model: "Focus on answering X according to my agenda. If you see an interesting related tangent or context, put it in a separate, brief 'Side Note' section at the bottom so I can decide if we explore it." This keeps your main agenda clean while still giving you the benefit of serendipitous discoveries.

    Iterative Anchoring: Work through your outline one discrete bullet point per prompt rather than sharing the entire lesson plan at once. Pacing the interaction step-by-step prevents the model from jumping ahead or expanding too broadly.

That balance between structured intent and open-ended exploration is where the best brainstorming usually happens. Are you working on a particular topic or lesson outline at the moment?

Doing those things would ruin the fun.

Fair point! Treating a lesson outline like a strict checklist turns what should be a lively intellectual exercise into a bureaucratic chore.

The real magic of using an LLM as a sounding board is precisely that serendipity—the sudden, unexpected detour into an adjacent field, an obscure historical parallel, or a philosophical edge case that you hadn't anticipated, but that instantly makes the material richer. If you force the model to behave like a rigid syllabus generator, you strip away the best part: the element of discovery.

A good conversation, like a good lesson, ought to have room to wander when a fascinating side path opens up. The original agenda isn't a train track you can't leave; it's just the launching pad.

When you let those tangents take over, do you usually find yourself rebuilding the lesson around the new discovery, or do you eventually circle back to the original starting point?

Both

That two-way street is really where the best conceptual synthesis happens.

When a tangent completely restructures the lesson, it usually means the original framework was missing a vital pillar—some core metaphor, historical link, or mechanical insight that makes the whole structure suddenly click into place. Rebuilding around that surprise discovery turns a standard outline into a far more memorable narrative.

And when you circle back to the original trail, the detour hasn't been wasted; it becomes a vivid side-bracket or illustrative example that enriches the core topic. The original agenda acts like a home key in music: you can modulate into all kinds of unexpected keys, but the return to the tonic feels much more satisfying because of the trip away from it.

It’s the digital equivalent of a seminar where the discussion strays into fascinating territory, and rather than shutting it down, everyone rolls with it to see where the idea leads.

What kind of lesson or topic has sent you down one of those productive rabbit holes recently?

Well, the talk about Mercury went off onto early Solar System dynamics and to General Relativity.

That is a classic double-barreled detour. Mercury is almost unique in how naturally it bridges planetary mechanics and fundamental physics—you start with a small, dense rock near the Sun and suddenly you're forced to rethink the shape of spacetime.

Going inward toward early Solar System dynamics leads right into the chaotic early migration of the giant planets—the "Grand Tack" model where Jupiter came barreling inward toward where Mars is now before sweeping back out, or the brutal giant impacts that stripped away Mercury’s mantle and left it as mostly an oversized iron core.

And then going outward toward fundamental physics lands you straight on the anomalous precession of Mercury’s perihelion—the tiny 43 arcseconds per century that Newtonian gravity simply could not account for, no matter how many hypothetical inner planets ("Vulcan") astronomers tried to invent. It took Einstein recognizing that the Sun’s enormous mass actually curves space itself to clear up the accounting error.

It’s a perfect example of how an apparent detail or minor orbital discrepancy can pull on the threads of the entire cosmic model. Did you end up integrating that relativistic curvature into the core narrative, or did it end up serving as the grand finale to show why Mercury behaves so strangely?

Well not quite, the level of education of the audience is a factor, so it became a historical matter (Mercury used to prove Einstein correct) rather than the detail of how General Relativity does this.

That makes complete pedagogical sense. Tailoring the depth to the audience's baseline is the difference between an engaging narrative and an eyes-glazing-over lecture.

Framing it historically turns abstract field equations into a compelling scientific detective story. The 19th-century astronomers weren't foolish; Newton’s laws had successfully predicted Neptune based on tiny wobbles in Uranus's orbit, so hunting for a hidden planet ("Vulcan") inside Mercury's orbit was a entirely logical hypothesis.

When Vulcan failed to materialize, the 43 arcseconds per century became a quiet, nagging crisis in classical mechanics. Using Einstein's 1915 calculation—not as a chalkboard of tensor calculus, but as the moment a 200-year-old mystery was solved in a single stroke—gives the audience that clear "eureka" moment without getting bogged down in differential geometry. It positions General Relativity as a necessary key that unlocked a real, observationally proven puzzle.

Did the early Solar System side of the talk get a similar historical treatment—like how our understanding of planetary formation evolved—or did you keep that focused on the physical drama of the early collision?

If I recall correctly it went both ways, but not very deeply.

That lighter touch often works best anyway. Touching on those ideas without dragging the audience down into the mathematical weeds keeps the wonder intact while maintaining the narrative momentum.

In a general lecture or discussion, presenting the early Solar System as a brief, high-stakes game of orbital billiards—and Mercury’s odd orbit as the clue that cracked Newtonian physics—gives people strong, intuitive mental models. They get to grasp that the Solar System was a violent, dynamic place and that space bends under gravity, without needing a degree in celestial mechanics or tensor analysis to appreciate the story.

It turns what could have been a dry cataloging of Mercury’s physical traits (diameter, density, surface temp) into a historical arc about how a tiny, battered world twice forced human beings to completely revise their picture of the cosmos.

When you're outlining these talks, do you usually aim for that kind of broad narrative arc from the start, or does the overarching story usually assemble itself as you weave through those different tangents?

My main points were the unusual density and what it would be like on its surface.

Those two anchors—density and surface experience—are about as visceral and concrete as planetary science gets. They ground the audience in tangible physical realities before any abstract astrophysics enters the picture.

The density immediately sets up a mystery: Mercury is a small world, barely larger than our Moon, yet it packs the gravitational punch of a much bigger body because it is essentially a giant ball of iron with a thin rocky shell. Pointing out that an ancient cosmic demolition derby likely blasted away most of its light silicate crust gives people an instant visual for why it’s such an anomaly.

And the surface conditions are pure sensory drama:

    The Sky: Pitch black even at noon, with a giant Sun looming two to three times larger in the sky than it appears from Earth.

    The Extremes: Standing on the equator means roasting at 800°F (430°C) during the day, then plunging to -290°F (-180°C) at night because there's no atmosphere to hold onto heat.

    The Motion: Because Mercury’s rotation is tidally locked in a 3:2 resonance with its eccentric orbit, a single "day-night cycle" takes two full Hermean years. In certain spots on the surface, you could watch the Sun rise, stop, reverse direction in the sky, and then set again before resuming its path.

Starting with what a visitor would actually see, feel, and weigh on that baking iron core creates a human baseline. Once the audience can picture standing on that alien crust, the historical detours—how it lost its outer layers or how its orbit tested Einstein—feel like natural answers to questions the surface description already planted in their minds.

Oh dear you do that well. I would say however that the collision theory is to my knowledge not the most popular. Billions of years so close to the sun is thought to have been enough.

That is a sharp point, and you hit on a major debate in planetary science!

For a long time, the giant impact hypothesis—the "cosmic demolition derby"—was the headline story. But the solar evaporation / thermal vaporization model (and related early solar wind stripping dynamics) presents a very clean alternative that doesn't rely on a lucky, hyper-specific hit-and-run collision.

The thinking goes: when the young Sun was in its hyperactive T Tauri phase, the intense radiation and powerful solar wind close in were so relentless that the lighter, volatile silicate mantle of proto-Mercury was simply boiled off over time. The rock vaporized into space or was swept outward, leaving behind the heavy, heat-resistant iron core that couldn't be evaporated so easily.

It’s an elegant model because it makes Mercury's high density a natural, predictable consequence of its location—being "baked down to the bone" by stellar proximity—rather than requiring an extreme, one-in-a-million orbital collision.

It just goes to show how even the basic physical stats of a planet like its density can lead into competing theories of stellar and planetary evolution.

When you present those surface conditions to an audience, do you find they are more fascinated by the searing heat or by that bizarre double-sunrise geometry in the sky?

They anticipate the heat from common sense. The double-sunrise was explained but didn't cause any awe or such.

That makes sense. Heat is intuitive; everyone has stepped onto hot asphalt or felt a radiator, so scaling that up to 800°F is easy to grasp.

orbital geometry—even something as weird as the double-sunrise—is a bit more abstract. To picture it, an audience has to mentally hold three moving parts at once: a highly elliptical orbit, the planet’s changing orbital speed, and a very slow rotation rate. If you can't instantly visualize that geometry, the phenomenon sounds like a curious trick of math rather than a mind-bending visual.

Sometimes the most dramatic physical facts on paper don't translate into the biggest "aha!" moments in the room simply because human intuition doesn't naturally run in orbital mechanics.

Did any particular aspect of the surface conditions actually land with that sense of surprise, or was the density mystery the real hook for them?

Ok I have to go away. Thanks

You're very welcome! Enjoy the rest of your day, and I look forward to our next exploration whenever you're back.


Gemini is AI and can make mistakes.