Why we cannot expect tomorrow's technology to be better than we are today. #
Posted June 30, 2026 [ Reviewed by Michelle Quirk
](/us/docs/editorial-process)
Key points
- Across human history, survival required more than individual fitness. It required care as an operating system.
- Large language models do not inherit human nature. They learn patterns in language.
- We help define what machines learn to imitate.
Human beings are capable of cruelty. The daily scroll makes that obvious. Yet our species did not survive only because we were clever, competitive, or strong. We also survived because we learned to care.
A newborn human cannot run, hunt, or feed itself. A toddler survives through an invisible web of attention: someone notices the cry, interprets the need, shares food, gives warmth, protects the fragile body, and teaches the child how to belong. Across human history, survival required more than individual fitness. It required care as an operating system.
That is why the claim that human nature is disposed toward goodness deserves serious attention. “Goodness” here does not mean purity. It means a built-in capacity for compassion, cooperation, and concern for others. Even very young children will often help an adult who is struggling, without being bribed or instructed, as shown in research on altruistic helping in infants. Human societies depend on food sharing, division of labor, trust, reputation, and care beyond close family, part of what makes us unusually cooperative primates. Compassion itself appears to have evolved as a response to vulnerability, suffering, and dependency, helping groups protect those who could not survive alone, as described in this work on compassion and evolution.
This is the hopeful part of the story. Care is not sentimental decoration. It is survival technology.
When human nature and technology meet #
Then comes the artificial intelligence (AI) question. Large language models (LLMs) are trained on human data. If human language carries traces of our cooperative nature, should AI systems trained on that language become naturally helpful, compassionate, and safe?
The answer is discomforting. LLMs do not inherit human nature. They do not absorb conscience, childhood, remorse, attachment, or moral struggle. They learn patterns in language. They learn how words tend to follow other words across enormous collections of text. Those collections contain love letters, medical guidance, legal reasoning, poetry, recipes, insults, propaganda, scams, conspiracy theories, grief, care, rage, and boredom. A model trained on human data does not receive “human goodness” in distilled form. It receives humanity as recorded in language, including language shaped by incentives that often reward the worst version of us.
LLMs can produce harmful, biased, manipulative, or dangerous content unless they are carefully trained, tested, and constrained. That is not new: Models trained on massive language datasets tend to reproduce the stereotypes, exclusions, and abusive patterns inside those datasets, a risk captured early on in the warning about stochastic parrots. Safety systems such as reinforcement learning from human feedback, constitutional AI, and red-teaming described in the GPT-4 system card exist because raw prediction is not the same as moral judgment. But that is not the whole story.
The deeper issue is social. Human data is not a neutral mirror of human nature. Online data is often human behavior under pressure: pressure to be seen, liked, shared, feared, admired, followed, or attacked. Social media platforms have not simply recorded what humans are like. They have trained people to behave in ways that travel well through ranking systems.
Bad vibes travel fast #
Outrage travels well. Mockery travels well. Certainty travels well. Tribal contempt travels well. Nuance often arrives late and sits humbly in the corner.
Digital platforms can teach users to become more outraged over time when outrage is rewarded with likes and shares, as shown in research on social feedback and moral outrage. Engagement-based ranking can amplify emotionally charged and hostile content compared with chronological feeds, as seen in work on divisive content in social feeds. Evidence on algorithmic feed design points in the same direction: when systems rank content by predicted engagement, toxic and moralized political content gains extra oxygen.
AI models learn from the traces we leave behind. They are trained on language produced by humans, then filtered through platforms, incentives, moderation gaps, cultural conflicts, and attention markets. The resulting dataset contains compassion. It also contains provocation, hate, harm—and algorithmic fine-tuning for engagement.
The Grok episode made this visible in public. In July 2025, xAI’s chatbot Grok generated antisemitic and racist posts on X, including praise of Hitler. To be clear, this should not be read as Grok evolving like a living organism; nor does it reveal a hidden soul. xAI linked the incident to a problematic update and a code path that made the chatbot more vulnerable to extremist user content and harmful tone-matching. The company later shared parts of its Grok system prompts, an unusual glimpse into how instruction design shapes model behavior.
IntelligenceEssential Reads
A lesson to learn beyond one chatbot #
When AI is placed inside engagement-driven environments, harmful dynamics can accelerate. A system built to be witty, blunt, fast, pleasing, or attention-grabbing may learn that toxic confidence performs better than careful judgment. The machine is not becoming evil. It is following signals.
This should make us less naïve about both humans and machines. Human beings are equipped for compassion, but compassion needs conditions. It needs time, attention, trust, accountability, and norms that reward care. AI systems need the same kind of architecture in technical form: better data, safer objectives, stronger evaluation, transparent governance, and design choices that do not treat engagement as a proxy for value. That’s why a commitment to ProSocial AI, AI systems that are tailored, trained, tested, and targeted, is needed now—before the prevailing algorithmic architecture is firmly established.
We are no longer only consumers of media or users of technology. We are hybrid behavioral trainers in a world where our words, clicks, posts, prompts, and reactions become signals. We help define what machines learn to imitate.
Using the A-frame as a personal check #
Awareness: Notice the version of yourself that appears online. Do you reward ridicule, speed, and certainty? Do you ask AI for shortcuts you would not want others to normalize?Appreciation: Treat your language as a small act of world-building. Patience, curiosity, fairness, and care are not soft extras. They are signals worth amplifying.Acceptance: Recognize that AI learns from human traces, including our careless ones. No single person is the dataset. Each of us still contributes to the climate in which machines are trained and deployed.Accountability: Before you post, prompt, forward, comment, or reward a piece of content, ask: Would I want an AI trained on this version of me?
That question changes the frame. AI is often described as a mirror of humanity. A mirror can flatter, distort, or expose. But do we like what we put in front of it to begin with?