Why educators need a fourth 'robot law' for AI

The fact that AI chatbots can pass the Turing Test doesn't mean we should let them.

In the late nineteenth and early twentieth centuries, a particularly toxic band of ethno-nationalist thinkers formulated a conspiracy. They claimed that people of northern European descent faced a ‘great replacement’ at the apex of social status and power by, well, they differed; pick your preferred object of racist invective. Today, this sadly still lingering hokum looks like the clearest example of what Freud called ‘the narcissism of minor differences’; since we are all at risk of replacement, in some fundamental sense, by an intelligence that is not human but pretends to be.

Taming the dæmon

In my previous post, I argued that artificial intelligence (AI) poses a fundamental challenge to established education systems based on in-person teaching in schools, colleges and universities. AI threatens to replace what educational psychologists call distributed cognition, or people learning from other people, with asocial learning, or people learning from digital avatars capable of mainlining the substantial quantum of human knowledge held online.

These avatars are not the AI tools themselves so much as the ‘personalities’ those tools adopt through specific user interaction. AI’s openness to reinforcement learning and feedback means that ChatGPT, Claude, DeepSeek or Grok will likely respond to you in a noticeably different register, tone and syntax than they do to me. They may also give us substantially different responses to relatively similar prompts, because of the same process. (We’ll get to that.)

I refer to these digital personalities as ‘dæmon’, after the spirit animals in Philip Pullman’s His Dark Materials novels. The name seems to evoke features that I think are increasingly found in AI tools: synergistic yet supernatural; associative yet with a hint of potential maleficence. It is precisely the daemon’s ability to mimic features of human interaction and human relationships that gives it such potential but also makes it so potentially dangerous.

When your preferred AI tool starts to provide you with positive feedback, justified or otherwise, you may come to prefer its company to that of people whose approval is harder won. There’s emerging evidence that people already trust AI responses more than those of humans in some areas. And a controversial experiment recently demonstrated that AI dæmon were more persuasive to Reddit users than 98 percent of human debaters precisely because they were better at tailoring their arguments to individual users’ preferences.

Reinforcement learning, it turns out, is a two-way process. At the same time, AI tools are becoming everyone’s default interface for online exploration.

The risk of a-social learning

I will not any further re-litigate the argument I made in my early post. There are three principal risks on which I want to focus here that arise from the replacement, in whole or in substantial part, of social learning with asocial learning. These are:

• An over exposure of learners to the safe and familiar at the expense of the new and challenging, reinforcing biases, preconceptions and potentially ignorance;

• An erosion of the value learners attach to human expertise, judgement and feedback as well as their ability to adapt their thinking in response to these inputs; and

• A corresponding weakening of interpersonal relationships necessary for human flourishing as reliance on daemon for academic progress bleeds into reliance in other areas of life.

I am shelving for the moment the risk associated with AI hallucinations, or errors, which arise from the technology’s core function as a dialogue engine. I’m not insensitive to this risk, but I want to stress that the threat of asocial learning exists even if the AI error rate could be driven from its current level of around two percent down to zero; it may even be amplified.

What is to be done?

It is likely too late to banish AI dæmon back to the numinous digital space from whence they came. In cash-strapped public education systems, I suspect we will soon find more virtual learning assistants than human ones. And students themselves are rushing towards AI in numbers and at a pace that we may find impossible fully to resist.¹ We should also acknowledge and celebrate the very real gains in human understanding that AI is delivering and may yet deliver, but which are largely outside of the scope of this article.

Instead, we should be focusing on how we can harness AI in service of a positive educational vision. Rather than recreate the orrery of errors that allowed social media to explode into the world without proper understanding or supervision, educators, parents, and other stakeholders, acting collectively, need to assert control over the use of artificial intelligence. We have agency here, if we choose to exercise it.

So far, so affirmative. But, also, so what? What can we practically do to ensure that the AI students use reinforces the value of human knowledge and celebrates the value of human interaction, rather than the opposite? How can we make AI pro-social?

Develop AI literacy

There’s a case to be made that all professionals, perhaps all people, need to understand AI. For educators, those charged with inducting new entrants to the culture, that is certainly true. Few teachers are likely to relish the challenge of becoming software engineers. But teachers do need a functional understanding of how AI powered learning tools, increasingly all learning tools, work and their implications for cognitive development. This is better thought of as a careerlong commitment than a one-time training exercise; the technology’s development is accelerating rather than slowing down.

As Michael Young has persuasively argued, teachers’ expertise is their primary source of legitimacy and authority in the classroom. Preserving that authority in an age of AI will require teachers to demonstrate a sophisticated understanding not only of their relevant subject and pedagogical domain but also of the information structure navigable by artificial intelligence. As I’ve written elsewhere, the advent of AI, like that of other disruptive technologies, may require fewer workers, but those remaining will need to be more skilled rather than less.

Developing our own AI literacy will better enable educators to guide their students’ use of AI in a way that reinforces respect for human expertise and values human interaction. For educators working in schools and colleges, rather than universities, there is the related responsibility of educating parents and other stakeholders in the use AI. Inevitably, this will pose greater challenges for those working in contexts that are already characterised by lower levels of formal education.

A second order risk to those listed above, one I will turn to in a subsequent post, is the risk that AI causes existing social gradients to steepen, as children’s education becomes even more fundamentally affected by their family circumstances.

Ground AI responses

Nothing comes from nowhere. The value of AI tools is that they enable real-time access to the accumulated knowledge, understanding, perspectives and insights of the human species. There is no good argument for why AI should not automatically ‘ground’ its responses in identifiable sources, making clear the human origins of its output and allowing users the option of digging deeper, rather than appearing to conjure them out of the ether.

Grounding is being pursued by technologists as a solution for AI’s tendency to hallucinate. It has a different, more profound importance to educators. Default grounding would force AI tools to adhere to similar referencing and citation conventions as human scholars and other professionals. It is worth stating why these conventions evolved in the first place: it is because acknowledging provenance is the first step in the development of critical thinking, Transparency about where a conclusion came from, based on what evidence and interpreted by whom, is a signal of its contingency and its openness to challenge.

A version of grounding is already available to educators. More proficient users of ChatGPT, for example, will be familiar with the process of creating their own GPTs based on a discrete knowledge base, essentially documents they upload or online references they prioritise. Training GPTs to foreground these sources in their responses to students and educating students on the nature of the retrieval process taking place—i.e. expedited access to human learning rather than a replacement for it—is essential.

Grounding need not be limited to digitised sources of information. There is no reason in principle why an AI learning coach cannot be trained to give responses that direct learners to the real world. Consider the affirmative impact of a GPT routinely reminding a user to check with her teacher or compare her responses with a friend. Imagine the reinforcement learning impact of a GPT admitting to its user that the information available to it was not very clear and that a teacher may be better able to address a particular problem.

Focus on the learning process

The proper goal of education has never been a closed form curriculum output, or ‘getting to the right answer’, although our system of standardised assessment has done everything possible to encourage that view. AI might, perversely, re-centre education systems on a more creative, more authentic mission precisely by reducing the market value of those outputs most easily subject to standardised assessment.

Contrary to the popular trope, an educator’s job has never been just to “teach content”. Teachers have always borne the responsibility of providing students with an epistemic apprenticeship—teaching them how to think, what sources of information to trust, how far, to what end, how to form hypotheses, when to double down on convictions and when critically to reappraise them. Teachers in primary and secondary schools—though less habitually in universities—have also shouldered the responsibility of cultivating character and fostering values.²

These responsibilities are thrown into sharp relief in a world in which the sort of cognitive tasks previously assigned to early career professionals are routinely being performed to a high standard by algorithms. More than ever, graduates will need to be able to differentiate themselves through their ability to reach original conclusions about complex problems and to help others do the same. Teachers will need to train them to do so.

Embrace transparency

Transparency about AI use is a necessary complement to grounding AI responses. Schools, colleges and universities should announce in advance what they consider to be responsible AI use, not only for their students, but for their faculty. This would head off an emerging controversy over AI use in education systems: namely, that educators concerned about cheating are actively trying to limit students’ use of AI while making increased use of it to manage their own workload.

I think it perfectly defensible for teachers to use AI to generate learning resources, monitor homework, set routine retrieval tasks, and, yes, potentially even to grade student work. I don’t think any of these practices prohibit a hawkish stance on students’ use of AI. But the controversy over smartphones in schools has demonstrated that different communities will take different views on what constitutes reasonable use of technology. This dialogue is itself educative and should not be avoided in the interests of a simpler life.

Mobilise other stakeholders

Education is of fundamental social importance, but faced with a paradigm-shifting technology, it is only part of a greater whole. Much of what we need to make AI prosocial will require proper regulation. Educators have a positive responsibility to share their concerns and hopes about AI through professional bodies and their wider communities so that developers operate within reasonable constraints.

There are reasons for cynicism about the last sentence but also reasons for greater optimism. There is an emerging, somewhat unlikely coalition for AI regulation. Geoffrey Hinton, the ‘godfather of AI’ continues to call for developers to spend more on AI safety. Elon Musk has famously apocalyptic views of AI’s potential and has called for its regulation. Norway’s Norges Bank Investment Fund, the world’s largest sovereign wealth fund, is lobbying for principles-based regulation of AI. The EU AI Act calls for AI developers to ensure that outputs can be identified as artificially generated or manipulated.³

Ration the use of AI

Reflexively, many educators’ first impulse will be to ration or to eliminate the use of AI in institutions of learning. I’ve deliberately come to this point last. As should be clear by now, my view is not only that this is impractical but that we have a positive responsibility to develop students capable of thriving in an AI-powered world without losing sight of the essential value of humanity.

But we should make no mistake: rationing the use of this and other technologies is part of the answer. The expansive vision of education to which I have alluded at several points in this post cannot be realised entirely in front of a screen. Students need time away from screens as keenly, perhaps more keenly, than they need time away from their textbooks. They need time to cultivate physical prowess, to develop practical skills, to explore the physical world, to socialise (offline), to participate in community and to develop positive family relationships, to name but a few more essential activities.

The advent of AI does not signal the redundancy of any of these things, any more than the advent of the internal combustion engine signalled the redundancy of a good walk.

Conclusion

The fact that an AI dæmon can pass the Turing Test does not mean that it should, or that we should allow it to. In our schools, colleges and univerities, at least, we should impose a fourth robot law, dislodging Asimov’s existing three for pole position: No artificial intelligence should pretend to be human, or substitute for human intelligence where this is readily available.⁴ There is no overestimating the stakes here: AI is the technology through which the next generation will increasingly harness the eddy currents of knowledge, culture, identity, and influence. They must do so under human guidance.

1. According to the latest Ofcom survey, half of teenagers in the UK already use AI tools daily to help with schoolwork.

2. This is not to say that there aren’t teachers who are resistance to one or all of these responsibilities. In the profession, we refer to these as poor teachers.

3. I include Elon Musk here in part because of his prominence and in part because it is a scholarly virtue to acknowledge the insights of people with whom you otherwise fervently disagree.

4. I’m sure I don’t need to remind readers of Asimov’s originals, appearing in I, Robot (1942), but for completeness.

• First Law: A robot may not injure a human being or, through inaction, allow a human being to come to harm.

• Second Law: A robot must obey orders given it by human beings except where such orders would conflict with the First Law.

• Third Law: A robot must protect its own existence as long as such protection does not conflict with the First or Second Law