Setting Deliberate Limits on AI Is Only the Beginning

A well-known YouTube science communicator recently admitted he had leaned too hard on artificial intelligence for research, touched off a public controversy, and then published a formal personal policy to govern his AI use going forward. The episode, reported by Vox.com , is concrete and specific. But the harder question it raises reaches further than any one creator's habits, or any one technology. It is about whether people are consciously deciding how much of their thinking to hand over to the platforms that shape their attention every day. A Dalmatian Observer's Take on Intentional Engagement The Dalmacija Portal Team has followed online culture in the region long enough to recognize a pattern when it surfaces. The article's central argument, that pre-committing how much cognitive engagement you give to any attention-shaping platform is what preserves independent judgment, maps directly onto what the piece calls 'cognitive surrender.' That is the structural risk embedded in generative AI, the tendency to offload human thought rather than exercise it. The Poincaré principle sharpens the point. Henri Poincaré argued that sustained, tedious conscious effort is precisely what produces flashes of genuine insight. A tool designed to eliminate that effort does not just save time. It removes the friction that makes thinking valuable in the first place. 'LLMs are designed to make cognitive work effortless, but that feels so icky because for it to be worthwhile at all, it has to be effortful.' That observation resonates beyond AI. The team notes that Dalmatian readers are bringing the same deliberate-limits habit to the regional entertainment and information sources they follow daily. Dalmacija Portal is one of the everyday online hubs readers now approach with the same conscious 'how much of my time and focus does this actually get' mindset the article recommends for AI tools. Hank Green's Controversy and the Policy That Followed Hank Green, a veteran YouTube creator, writer, and science communications entrepreneur, went public with an uncomfortable admission. 'I have been relying too heavily on AI as a research aid,' he wrote. 'It can be very useful for this task, giving me access to a lot of papers I didn't know existed really fast, but I think that has been to the detriment of my work because it has not given me the freedom to find all of my own ways into and around a topic.' Green also wrote that his finished videos had taken on an ineffable 'AI feel' and that his relationship with the technology had become 'not healthy for me or good for the world.' His followers, known as Nerdfighters, responded sharply, criticizing him for using AI at all. That backlash quickly generated its own counter-backlash, with critics accusing Green of 'self-canceling' over what they called a legitimate and unremarkable use of technology. Two sides, one admission, no consensus. Green's response to the noise was disciplined. In a follow-up video, he published a four-point personal AI policy. Scripts will not be written, edited, or outlined by a large language model. The thesis of any video must originate with a human. No image or music in a video will be AI-generated. And LLM outputs are not trusted as a source. Four rules. Specific, public, enforceable. How AI's Ease Becomes Its Most Dangerous Feature The controversy around Green is partly a story about how much the technology has changed. Since late 2022, large language models, particularly paid premium versions, have become significantly smarter. They are now described as 'unnervingly good' at summarising niche, complex research areas and at generating ideas for further inquiry, often without being prompted to do so. That capability is exactly where cognitive surrender enters. The design of generative AI pushes users toward the tasks- synthesising research, brainstorming, generating angles- that most directly short-circuit original thinking and discovery. The tool is built to be effortless. That is its selling point. It is also the mechanism by which it erodes the mental work it appears to be supporting. Even purpose-built, narrower tools carry the risk. Google's Gemini Notebook, formerly NotebookLM, allows users to upload their own sources and query only those materials, making it less prone to producing false information than general-purpose AI. The Vox author uses it for most stories. Yet the author acknowledges it still enables engaging with sources in a 'perfunctory, contextless manner,' because the tool surfaces the precise passage a reader needs rather than requiring the reader to form connections from working through a text whole. Poincaré's insight, that effortful sustained thinking is the precondition for genuine discovery, cuts against every feature that makes these tools appealing. Concrete Rules for Limiting What AI Does to Your Thinking Green's four-point policy is one model. Practical alternatives exist at smaller scale. One Vox colleague does not use AI to brainstorm ideas at all, restricting it to answering narrow factual questions and to aiding fact-checking only after a story is already written. The rule is simple and the boundary clear. No synthetic work before a draft exists. Another approach treats AI as an enhanced thesaurus, a tool for finding a precise word or short phrase when nothing better comes to mind, with a strict self-imposed ceiling of two or three words maximum taken from the output. Sharing substantial portions of one's own writing with the tool is avoided, since doing so invites the model to recommend extensive rewrites and gradually pulls the writer's voice toward the AI's register. Socratic prompting offers a different path entirely. Instead of asking an AI to supply answers or generate ideas, a user poses questions and pushes back on responses, using the model as a sparring partner rather than an answer dispenser. The thinking stays with the person. The Commercial Pressure That Makes Personal Policies Necessary Green described 'the level of dopamine I've been getting from interacting with LLMs' as one of the unhealthy dimensions of his AI use. That framing is not incidental. AI labs have strong commercial incentives to maximise engagement, and engagement that feels rewarding tends to become habitual. Researchers at some AI labs are thinking about the societal risks of cognitive atrophy. The article is direct about what that means in practice. It is unrealistic to expect companies competing on ease of use to voluntarily introduce friction into their models. The market does not reward that. Which is why the locus of control lands, somewhat uncomfortably, with individuals and with the social norms communities enforce together. Personal policies of the kind Green published, specific commitments made in advance about what a tool will and will not be used for, are one real lever. Social norms that treat thoughtful limits on AI use as reasonable rather than eccentric are another. Neither depends on a company deciding to act against its own interests. Choosing how much of your attention any platform receives, AI research tool or otherwise, is both a meaningful decision and one that remains within reach.