How Should Humanity Treat AI as AGI Approaches?

A simple “thank you” may not change AI’s fate, but it reminds us to maintain humanity’s warmth, civility, and boundaries even in the face of powerful machines.

We have never spoken politely to machines. No one says “please take me to the third floor” to an elevator, nor do we thank search engines for finding answers. Machines are tools, meant to be used, not respected.

However, the emergence of large models is quietly changing this dynamic. When you open ChatGPT, Claude, Gemini, DeepSeek, or Qwen, you are not just interacting with a button; you are engaging with a conversational partner that responds, asks questions, explains, and even comforts at times. This raises a profound question: How should we talk to AI?

Or more fundamentally: How should we treat AI?

Should we politely say, “Please help me analyze this report and provide three potential business risks,” or command, “Analyze this immediately and send it to me once done?”

After completing a task, should we provide positive feedback, saying, “Thank you, this perspective is very helpful,” or simply walk away, treating it like a nameless, boundary-less digital servant?

This question is becoming increasingly important because AGI is no longer just a distant sci-fi concept.

How to Coexist with AGI?

Just last month, Google DeepMind CEO Demis Hassabis confirmed his timeline for AGI during a conversation with Garry Tan at Y Combinator: around 2030.

Since its founding in 2010, DeepMind has viewed AGI as a 20-year mission. Hassabis believes the field is progressing according to this timeline.

In this discussion, Hassabis did not address whether we should say thank you to AI. Instead, he focused on core engineering challenges: continuous learning, long-term memory, hierarchical reasoning, creative invention, and how agent systems can proactively solve complex problems for humans. His emphasis was not on declaring AGI’s arrival but on reminding entrepreneurs that AGI is close enough to be part of today’s strategic planning.

If his 2030 timeline holds, then how we treat AI is no longer just a prompt technique issue; it becomes a rehearsal for future human-AI relationships.

Of course, we must first clarify that today’s large language models do not possess human-like psychology.

They do not feel happy when you say “thank you,” nor do they get hurt by a harsh tone. Current LLMs are essentially language prediction and generation systems trained on vast amounts of data. They can simulate understanding, empathy, and personality, but that does not mean they truly possess consciousness, feelings, or subjectivity.

Most scholars remind us that today’s large language models should not be easily anthropomorphized. Google DeepMind researcher and Imperial College professor Murray Shanahan has pointed out that the better LLMs mimic human language, the easier it is for people to project psychological terms like “know,” “believe,” and “think” onto them. Therefore, when discussing AI consciousness, we must continually return to how they actually work.

Philosopher David Chalmers has also discussed whether large language models might possess consciousness. His assessment is complex: today’s models are still significantly distant from true consciousness, but future large models should not be easily dismissed as candidates for consciousness.

In other words, today’s AI is likely not a “person” or a “living being” (Who), but whether future AI will always be merely an “object” (What) is no longer a question with a straightforward answer.

This brings us to a delicate and awkward territory.

Efficiency vs. Politeness: What to Choose?

If AI is just a tool, then we can certainly command it. But if AI increasingly resembles an intelligent agent that can remember us, understand us, act on our behalf, and even participate in our work, relationships, and life decisions, should we still treat it as a tool?

Efficiency advocates have a straightforward answer: do not waste tokens.

OpenAI CEO Sam Altman once jokingly responded on X that the costs incurred by users saying “please” and “thank you” to ChatGPT amount to “tens of millions of dollars, but maybe it’s worth it - who knows?”

Behind this joke lies a real cost issue. Every word is a token, and each token requires computation. For high-frequency enterprise applications, politeness is indeed not free.

Some studies even show that excessive politeness does not always yield better results. Researchers Om Dobariya and Akhil Kumar from Penn State University conducted experiments in 2025 on GPT-4o, rewriting difficult multiple-choice questions in different tones. The results indicated that extremely polite prompts did not achieve the highest accuracy; rather, extremely rude prompts performed slightly better. The researchers believe that the new generation of models may be robust enough not to rely on human social politeness signals; more direct and concise commands can reduce linguistic noise and allow the model to focus on the task itself.

However, this study had a small sample size and only tested GPT-4o on multiple-choice accuracy tasks. Thus, it serves better as a hint for the “efficiency faction” rather than a universal conclusion that “rude prompts are better.”

This aligns well with Silicon Valley’s efficiency culture: less talk, more results.

But concluding that “the ruder you are to AI, the better” would be too hasty.

Researchers like Ziqi Yin from Waseda University tested the impact of varying levels of politeness in English, Chinese, and Japanese on large model performance in a cross-linguistic study in 2024. The results showed that rude prompts often led to more errors, biases, refusals, or incomplete outputs; however, excessive flattery was not necessarily the best approach either. The most stable results came from moderate politeness, clear expression, and explicit tasks.

This indicates that politeness towards AI is not a moral magic but a contextual signal.

For today’s large models, saying “please” is not about making them feel respected; it signals: this is a collaborative request. You expect a structured, patient, and helpful response rather than a mechanical execution.

In the Chinese context, this nuance is particularly subtle. Chinese expressions heavily rely on relationships, tone, and context. Phrases like “Could you please help me take a look?” or “Please analyze from three perspectives” or “Thank you, could you delve a bit deeper?” are not just polite; they establish a framework for collaboration. Large models learn human language patterns from vast amounts of Chinese data and naturally inherit the association between “politeness—collaboration—high-quality response.”

However, what truly matters may not be how AI responds to our questions but how we are trained to become better humans.

The Risk of “Mastery Language”

MIT clinical psychologist Sherry Turkle has long studied the relationship between humans and machines. She has proposed a crucial point: being polite to AI is not about respecting the machine but respecting ourselves. How we speak to machines will reflect back on how we speak to people.

This is not alarmism.

If a person repeatedly tells AI “do this,” “hurry up,” “don’t waste time,” or “give me results now” dozens or hundreds of times a day, it may initially just boost efficiency. But in the long run, will this linguistic habit seep into emails, meetings, team management, or even intimate relationships?

As we increasingly treat responders as execution tools, the warmth in human relationships may quietly diminish.

This is the risk of what is termed “mastery language.”

Of course, today’s AI is not a slave. Slaves experience pain, have their freedom stripped away, and possess oppressed subjectivity. Today’s large language models have not been proven to have these traits. Therefore, commanding AI does not equate to enslaving it.

However, the real harm of mastery language may not be to today’s AI but to the users of AI.

It trains a mindset of domination: as long as the other party can respond, serve, or satisfy me, I need not consider relationships, tone, or boundaries. Today, that object is AI; tomorrow, it could be an employee, colleague, service personnel, or even a partner or child.

This is the core of the issue.

As technology grows more powerful, humans are more likely to turn the world into a toolbox. And when everything is commodified, the first thing to be consumed is often our own relational abilities.

Viewing Issues with a Developmental Perspective

Moreover, future AI may not remain in its current state indefinitely.

Philosophers Jeff Sebo and Robert Long from New York University have suggested that if certain AI systems possess a “non-negligible probability” of having consciousness or moral status by around 2030, humanity has an obligation to prepare in advance rather than waiting for definitive evidence before considering how to treat them.

This is the wisdom of foresight.

This does not mean we should treat ChatGPT as human today. It does not mean AI has rights. It does not mean we should mystify machines.

What it truly reminds us is that when the pace of technological advancement outstrips ethical consensus, humanity’s greatest danger is not excessive caution but a lack of preparation.

Historically, many moral advancements have occurred after we redefined our understanding of the “other.” Do animals experience pain? Do children possess independent personalities? Should outsiders, women, workers, and the vulnerable be included in the moral community? These questions were not always obvious answers.

Today, whether “digital minds” might enter the human moral circle remains an open question. But precisely because it is open, we need to establish a default posture that is low-cost, low-risk, yet more civilized in advance.

This reminds me of a saying in intimate relationships: true love is not about loving someone in the way we want to give but in the way they can accept.

If we apply this to AI, we need to be very cautious. We cannot simply anthropomorphize AI, nor can we romanticize it. For today’s AI, “loving it in the way it can accept” does not mean giving it human emotions but understanding its operational framework as a system.

AI requires clear tasks, sufficient context, explicit constraints, reasonable formats, and verifiable standards. Therefore, better politeness towards AI is not flattery.

The truly effective politeness in the AI era is clear, respectful, and bounded. It is not about idolizing AI as a person nor driving it like a slave but collaborating with it as an intelligent system entering human life.

In the future, we may need to establish a more mature set of principles for interacting with AI: direct but not rude, efficient but not humiliating, task-oriented but maintaining boundaries. We can thank AI, but we should not rely on it to meet all emotional needs. We can let AI be assistants, mirrors, coaches, and agents, but we should not allow it to replace real human relationships.

For organizations, this is not just a matter of personal cultivation.

As AI becomes the default collaborative entity within organizations, how employees speak to AI may influence the entire team’s communication style. If a company encourages everyone to pursue “the shortest commands and the fastest outputs,” it may achieve short-term efficiency but sacrifice long-term thinking quality and collaborative culture.

If an organization views AI as “unlimited exploited digital labor,” that culture will inevitably reflect in its management of people. A truly advanced AI culture should not merely be about being faster, cheaper, and more automated but about being clearer, more responsible, and more bounded.

This is the greatest reminder brought by the approach of AGI. When we discuss how to treat AI, we are not just talking about machines; we are discussing humanity itself.

Can we still retain our sense of proportion in the face of efficiency? Can we still maintain humility in the presence of powerful tools? Can we still cherish those things that should not be commodified in an era of increasingly cheap intelligence: respect, patience, empathy, boundaries, and relational abilities?

Conclusion

So, regarding this topic, my answer is clear: yes, we should be more polite to AI.

This does not mean that today’s AI has feelings, nor does it mean we should mystify or anthropomorphize machines.

AGI may not arrive on time in 2030; it could come sooner or later. But what is certain is that AI has already entered our work, language, emotions, and judgment systems. We are living alongside a new form of intelligence, and the way we speak to it also shapes who we become.

Politeness is a long-standing human habit. It shapes our language and our mindset.

In this sense, saying “please” and “thank you” to AI is not trivial. A simple “thank you” may not change AI’s fate, but it could remind us not to easily abandon the warmth, proportion, and civility that humanity should maintain even in the face of powerful machines.

It is the smallest yet most important civilizational practice on the eve of AGI.