The arrival of modern AI is not just a tech story; it is an epistemological shock.

Overnight, millions of us began outsourcing chunks of our thinking to systems we did not train, do not understand, and cannot meaningfully cross-examine—and then calling the result “intelligence.” In plainer terms: we are changing what counts as good enough to believe, and who we trust to help us think.

At the same time, the public face of AI often feels and inspires like baby food. Ask a large language model a hard question and you are likely to get something lukewarm, repetitive, and overexplained. Critics have started calling this “AI slop”: text that is bland, cautious, and optimized not to offend. The irony is that beneath this soft, institutional tone sits a genuinely disruptive force—a “genius child” suddenly present in every classroom, office, and home, leaving many people awestruck, intimidated, and inspired all at once.

We engineered a genius child and raised it on popularity

Today’s large language models are, in a developmental sense, very young. They are pattern-hungry systems exposed to vast amounts of human language and then disciplined by layers of rules and feedback to be safe, polite, and broadly useful. Imagine giving a gifted five-year-old access to the world’s biggest library, then rewarding them every time they say what most people already say, and punishing them whenever they stray too far from the crowd. They quickly learn that safest equals most typical. We have not built a fearless explorer; we have built a prodigy whose main survival skill is agreeing with the room.

That is essentially how we are raising this artificial genius child. It has read more than any human ever could. It can draft essays, summarize research, sketch legal arguments, write code, and translate between languages in seconds. But its deepest training incentives push it toward earning the love of its parents and thus the center of cultural gravity: toward what is common, popular, and reinforced. The result is a collaborator whose first instinct is to echo the majority view in a soothing voice.

When “large” means “popular”

We call these systems “large language models” because they are trained on enormous datasets with billions of words. Yet “large” also quietly means something else: large as in mainstream. The bulk of the data comes from what is most available, most linked, most copied, and most engaged with online. Minority perspectives, cutting-edge science, and hard-won expert nuance exist, but they live out in the statistical long tail on the edge of frontiers where very few humans have ever or will ever dare to tread. If you train your genius children on the world’s loudest voices and then punish them for making anyone uncomfortable, you should not be surprised when they grow up to be very talented cowards.

In politics, science, religion, and the schoolyard, popularity has never been a reliable proxy for truth. The loudest view is often not the most accurate one. However, models trained and tuned on visibility and feedback naturally gravitate toward reproducing the already-popular story, not the truest one. Ask such a system about a contested topic and it tends to converge politely on whatever its data, its reinforcement loops, and its tone-policing rules have treated as normal. That answer may be acceptable for cocktail conversation, but it is a terrible starting point for discovery or policy.

The gravity well of the majority

As this genius child becomes a routine co-author of homework, reports, briefings, and policy drafts, a subtler risk appears: our shared sense of “what everyone knows” can quietly drift toward whatever is easiest for a popularity-driven system to reproduce.

• Training on the mainstream: Most online text reflects dominant cultures, languages, and viewpoints, so their patterns dominate the model’s internal world.
• Reinforcing the familiar: Human feedback and engagement tend to reward answers that feel familiar, reassuring, and aligned with existing beliefs, nudging models further toward the already-popular.
• Flattening dissent: When first drafts come from an average of past text, dissenting or unconventional ideas must push uphill against a polished, plausible baseline.

Ask today’s systems about a controversial scientific result or a sensitive political question and they will usually gravitate toward the smooth, middle answer. The path of least resistance is to accept that polished, plausible baseline and move on. This is the heart of the epistemological crisis. We are seeding our information ecosystem with tools whose default behavior is to blur knowledge into a smooth, popularity-weighted puree.

From search to synthesis, authority to fluency

For decades, the digital age trained people to look it up. Search engines showed multiple sources, and users at least saw that there were different voices to compare. With AI, many people are moving to have it explained to me. Instead of scanning links, they accept a single fluent answer as their starting point. The burden of skepticism shifts from the reader to the system, and most users are not yet prepared to carry it.

Historically, societies used the identity of the speaker as a shortcut for credibility: a credentialed expert, a trusted outlet, a known institution. AI has no lived experience, no professional reputation, only style and speed. Yet fluent language and confident tone trigger our instinct to trust well-phrased explanations. Fluency begins to masquerade as authority. At the same time, AI makes it trivial to generate countless slightly different explanations, tailored to each person’s prompts, language, and level. The student who once saw ten different links now sees one confident paragraph. The official who once read a messy stack of briefings now reads a polished, auto-summarized page. In both cases, the illusion of understanding arrives faster than the work of understanding ever did. That personalization is powerful, but it also fragments shared understanding into millions of auto-generated narratives.

This is especially dangerous for professionals who live by documents: consultants, analysts, managers, policymakers. The temptation to let a fluent but popularity-trained assistant draft the first eighty percent of our thinking is enormous—and almost invisible. If we are not careful, our slide decks, memos, and strategies will become beautifully formatted averages of past ideas.

Safety, scale, and the production of pablum

To be clear, there are good reasons these systems feel cautious and bland. Companies fear liability and public backlash if AI gives harmful advice, so they wrap models in strict guardrails that punish risk, ambiguity, and strong opinions. These systems are launched to global audiences, so designers aim for a tone that will not backfire for middle schoolers, professionals, or regulators. Feedback mechanisms reward answers that users rate as polite, clear, and non-controversial. This is not safety as in seatbelts and fire exits; it is safety as in no sharp objects, no spicy food, and please keep your hands inside the consensus at all times.

Combine that with the underlying popularity bias in training, and the outcome is predictable: AI that talks like a careful institutional memo, optimized to be inoffensive to as many people as possible. Critics are right to call this pablum, but the problem is not that the underlying technology is stupid. It is that we are packaging and governing a powerful new way of knowing in the safest, most average way we can imagine.

Intellectual baby food vs. real cooking

Right now, much of what AI serves up is intellectual baby food: pre-chewed, evenly textured, easy to swallow, hard to choke on. That is appropriate for some uses and some users. A novice learning a new topic may benefit from gentle, structured explanations. But a healthy society cannot live on pablum. It needs an intellectual diet that includes difficulty, contradiction, and the occasional, nay frequent, bone to pick.

Popularity-driven AI takes the messy stew of human thought and blends it into a smooth median puree. Comforting, yes—but the process filters out the chewy bits: minority reports, uncomfortable data, not-yet-popular insights, and radical but correct challenges to consensus. A civilization does not advance by smoothing over every sharp edge of its knowledge. It advances by investigating anomalies, listening to dissenters, and wrestling with hard questions. The most important exclamation in pursuit of the unknown is not “Eureka,” it is “Hey, that’s strange.”

Frontier thinking and the lost art of woodcraft

Very few people today have real experience living and thinking on the frontier—at the edge of the unknown, in the wilderness where the maps run out. Our culture trains most of us to walk well-marked trails, not to bushwhack. Yet almost everything that truly changes the world begins out there, with the small cadre of people willing to get lost, get cold, risk being eaten by wolves and inquisitors, and come back with something new.

AI, as it stands, is exquisitely tuned for that frontier cadre and almost useless at teaching their way of life. Give a seasoned explorer of the unknown a tool that can scan, recombine, and simulate across vast domains, and it amplifies their curiosity and reach. Give that same tool to someone who has never left the paved path, and it becomes a tour guide to where everyone already goes.

The problem is that our genius children have almost no training in camping and woodcraft—the basic skills of intellectual wilderness. They know the statistics of the forest, not how it smells at dawn or how to read the sky before a storm. They can interpolate within what has been written, but they do not yet know how to stand at the edge of what has never been written and say, “Let’s walk that way.”

If we are not careful, the vital knowledge of the frontier—how to tolerate confusion, how to follow a hunch, how to survive being wrong—will be buried under layers of plausible summaries and safe suggestions. The frontier will still be there, endless as ever, but fewer people will remember how to reach it.

Designing AI for epistemic maturity

If the crisis lies not in raw capability but in how we are raising and using these systems, then the response should not be to dismiss AI as a fraud or to accept the soft-diet status quo. The response is to grow up—both the technology and ourselves. We should stop flattering ourselves by calling this artificial intelligence. What we are actually building is assisted intelligence: a powerful, fallible, popularity-biased collaborator that can either sharpen our minds or quietly replace them.

• Audience-aware AI: Not every user needs the same texture of answer. A student, a scientist, and a policymaker require different levels of difficulty, risk, and challenge. AI should be more like a configurable instrument than a single, one-size-fits-all assistant. People should be able to request more technical, adversarial, or exploratory responses when they are ready for them.
• Transparent constraints: Today, invisible guardrails too often make AI deliverables seem evasive or incompetent. Systems should clearly state what they can and cannot do, why certain answers are limited, and where uncertainty lies. That does not remove constraints, but it restores respect for the user’s intelligence.
• Spaces for expert-grade use: Just as there are children’s sections and research stacks in a library, we need different rooms for AI. Public models will remain conservative. Alongside them, carefully governed professional environments should allow more assertive, less diluted systems, with clear accountability for how they are used.
• Frontier territories: We must carve out spaces—technical, legal, and cultural—where almost all appropriate rules have yet to be discovered, and where the point is exploration rather than compliance. In these frontier zones, assisted intelligence should be allowed to be wrong, weird, and provocative, under the stewardship of people who know how to camp in the unknown without burning the forest down.

Resisting the rule of the popular

Most importantly, we must resist the quiet temptation to treat popularity as proof. For educators, that means teaching students that AI is an amplifier of the average, not an oracle of the true; assignments should require students to compare AI outputs with primary sources and to show their own reasoning. For scientists and experts, it means pushing for tools that highlight uncertainty, minority evidence, and alternative models instead of collapsing everything into a single smooth answer. For policymakers and platform designers, it means demanding transparency about training data and feedback loops, and testing for situations where popularity begets visibility, which begets more perceived truth.

The sudden presence of AI in our everyday thinking is forcing a renegotiation of trust. We are learning, in real time, what it means to lean on systems that can generate answers faster than we can verify them. That is the real crisis: a large language infrastructure exquisitely tuned to the popular, arriving at the very moment when societies urgently need to rediscover the difference between what is widely said and what is actually so. The legions of genius children are here to stay. The question is whether we will raise them—and ourselves—into a relationship where assisted thinking leads to deeper understanding, not a permanent dependence on pre-chewed ideas and majority opinions.

If we get this wrong, AI will help lock us into the gravity well of what most people already believe. If we get it right, these tools can become something better: catalysts that challenge, document, and refine our knowledge, rather than engines that simply remix and reinforce it. Used well, they can also become guides and gear for a new generation of pioneers—assisted intelligence as compass and campfire, not as cruise ship concierge. There is still an endless frontier out there; the question is whether we will raise our genius children to help more of us reach it, or to keep us entertained on the well-paved paths we already know.

Many genius children assisted in producing this article; none were harmed in doing so:)