This post tries to tie a few threads together which eventually evolve into a tentative set of principles for AI usage in modern times. It begins by briefly talking about Yochai Benkler’s The Wealth of Networks (the introduction chapter), expands into some work with AI around collecting a range of cross-disciplinary principles, and ends by proposing a few principles for living and working with AI.
The post is written while following the principles of AI usage proposed below.
The Internet Age
“Human nature is not a machine to be built after a model, and set to do exactly the work prescribed for it, but a tree, which requires to grow and develop itself on all sides, according to the tendency of the inward forces which make it a living thing.”
John Stuart Mill, On Liberty (1859)
So begins Yochai Benkler’s The Wealth of Networks.
Benkler opens by saying that information, knowledge, and culture are central to human freedom and development, and that how they are produced and exchanged shapes everything: how we see the world, who gets to decide things, what we can imagine doing collectively.
For a century and a half, he says, modern democracies depended on an industrial information economy for these functions. A small number of players (broadcasters, publishers, film studios, newspapers) produced information at scale, and the rest of us received it. Because the capital requirements were high, information production naturally concentrated into a relatively small number of institutional actors.
What shifted was the fact that in most advanced economies, there was a move to information centered economy, cultural production and brand based consumption. Second, the physical equipment required to produce and distribute information collapsed in cost. And with the coming of cheap processors and networks, the phenomenon of Internet.
This combination, an economy where information is the primary good, and where the means of producing and distributing information are now widely owned, created something new: what he calls the networked information economy. “Individuals can reach and inform or edify millions around the world. Such a reach was simply unavailable to diversely motivated individuals before”. He offers Wikipedia and Linux as examples, neither possible in the industrial model, both emerging naturally once the economics changed.
He wrote this book in 2006, a decade and half after internet became a meaningful public phenomenon. His argument was that the internet represented not merely a technological shift, but a structural one (“It goes to the very foundations of how liberal markets and liberal democracies have coevolved for almost two centuries”). Given this structural change, individuals are free to take a more active role than was possible in the industrial information economy of the twentieth century.
The old model required concentration because of economics. The new model makes genuine decentralisation possible not as an ideology but as an economic fact.
As the networked information economy develops new ways of producing information, whose outputs are not treated as proprietary and exclusive but can be made available freely to everyone, it offers modest but meaningful opportunities for improving human development everywhere.
The Impact of AI
Now, the question that led me to this post: What has happened in the twenty years since Benkler wrote that introduction?
Two structural developments stand out. a) Markets finding attention as a commodity and b) The emergence of AI.
Firstly, the internet did decentralise publishing and communication (changed the world as he envisaged), but markets eventually discovered how to monetise attention. Human attention itself became a commodity: bought and sold, measured and optimized. When Benkler published his book in 2006, the global online advertising market was roughly $16 billion, nascent, still finding its footing. Today, it exceeds $700 billion, accounting for over 70% of all advertising spending worldwide. The commons that Benkler referred to (the open, non-market space of peer production and free information) did emerge. But the market found the commons, and found in it something more valuable than content: it found human attention itself. Attention became the commodity that the industrial economy had failed to fully monetise.
The second structural development is that of AI. (“What is AI, in the sense that matters here? At its simplest, it is a system trained on vast amounts of human-generated text, code, images, and data, until it develops the ability to generate responses that are coherent, contextually appropriate, and increasingly useful across an enormous range of tasks.”)
AI is different from the internet. The internet changed the distribution of information, allowed anyone to connect, communicate and publish. AI increasingly allows anyone to analyse, synthesize, generate, simulate, critique, translate, summarize, design, and reason.
But extending it further to the way it is going, it is an infrastructural shift in the world, the impact of which is yet to be felt across disciplines, industries, people, companies, governments. Electricity is perhaps the closest analogy. At first electricity appeared optional: useful but not yet foundational, people were used to living without it. Over time it became inseparable from ordinary life itself. Entire industries, institutions, and social expectations reorganised around it. Unlike the internet, which primarily changed how information moved between institutions, AI is beginning to change how institutions themselves function.
So how does it stack against the argument extended earlier, in the introduction to Benkler’s book, about decentralisation? Some of the argument can be stretched forward. But a large part changes because AI is fundamentally different than internet. Millions of people can now access capabilities once restricted to specialists. Yet the systems themselves remain highly dependent on concentrated compute, chips, energy, data, and capital: the industrial economy yet again with centralised concentration.
In another argument with Benkler’s introductory propositions, as to the market, I believe it will find its way around AI the way it found its way around internet. Isn’t it something about the nature of markets? Market that way is a natural system. It works aligned with nature’s principles. Any other system requires energy to maintain and sustain it, as it goes against the grain of nature. Markets work with nature. Markets will eventually find their way around AI.
Principles
That brings us to principles. In another project with AI, I have been compiling principles that underlay several human disciplines, partly for personal contemplation, partly to be inspired by the thinking of the ages. Available here.
When one thinks of any discipline, the best way to approach it is to begin at the Principles. The distilled wisdom of the ages iterated over usage – the similar and unique problems that each discipline faces. One of the things that this document allows to see is that how these principles converge around the same few structural problems of world and matter – that of scarcity and resource allocation, that of information asymmetry, that of the tension between particular and general, that of the gap between intention and effect, and the tradeoff between resilience and efficiency.
Principles are a sort of arrival. Where rules cover the cases one can anticipate. Principles cover the cases one cannot anticipate. And yet here I am, arriving at some early tentative principles suggested for AI usage.
Suggested principles for AI Usage
As any other person living in the modern time and age, I’ve been thinking a lot about AI, about implications, about its nature and structure, about its usage and the changes it brings to the world. It is a fundamental shift in everything. It will soon change the way things happen in the world. In such a scenario, what are the considerations that will help ensure that we work and live with AI in the best possible way?
The thoughts in various conversations (Benkler, the impact of AI and Principles project) somehow fused together – lets call it an evolving thought, open to debate, challenge and discussion.
As in other cross-disciplinary principles that evolved with iterations, principles of AI usage will also emerge – where one of the first one would be perhaps the application of human discernment and human judgement.
I. The Principle of Human Discernment
Every tool in history has required the judgement of its user. But most tools are transparent about what they do. AI does not merely serve objectives, it interprets them. That interpretation is never neutral, never fully transparent, and never the human’s to let go of.
There are two parallel thoughts that I hold here. First, as Mill notes, that human nature is not a machine to be built after a model, but a tree, which requires to grow and develop itself according to the tendency of its inward forces. Each person is a unique specimen of particular pleasures and pains, shifting objectives, irreducible individuality. And the second thought here is that AI can be trained. To see its own blindspots better than perhaps humans can be trained, to critique itself, and under human guidance it can blossom away from the sameness that otherwise it seems to head towards. AI is neither just like language, nor just like an infrastructure, or an educational system. It is more dynamic, more available to training and change, and can be adapted differently by each organisation that uses it.
These two thoughts allow for making something greater than human, and greater than AI when human judgement and discernment is applied over AI. AI can optimize within frames as long as humans remain responsible for choosing the frame. Then the discernment of how to train it. Drawing from the other cross-discplinary principles, it is still that of the human wisdom.
AI expands capability but weakens friction. Therefore human judgement becomes more important, not less.
Human discernment as an underlay and overlay to AI usage.
II. The Principle of cross-platform Critique and Audit
The insight underneath this is that different foundation models have different blind spots because they were trained differently, on different data, with different choices. By asking one model’s output to be critiqued by another model, one will find nuances hitherto unseen. It not only improves the overall output but also at times, the models will systematically disagree in ways that are informative. A response that three different models converge on is more likely to be grounded than one that only one model produces. A response where they diverge significantly is precisely where human discernment (the first principle) is most needed.
The other key aspect of this suggested principle is to audit the output – implying accountability, not just accuracy. An audit allows for a higher standard and a more useful one for consequential decisions. It also addresses the more epistemic problem, whether the AI’s reasoning is trustworthy, or are its blind spots invisible to itself. It uses diversity of model as a check on any single model’s priors. (That leads us to a corollary – it is better to have a large number of good models than less)
The idea is to slowly build a system of checks and balances like the traditional principles of law, accounting propose. Over time, this principle does more than check. It works in constructive ways as well. It teaches. Used well, it is a new form of thinking, not just checks and balances, but a synthesis of what each model does well.
III. The Principle of Graduated Trust
Deploy AI where reliability is high and failure is recoverable. Refine the boundary continuously.
Not every domain is equally ready for AI, and within each domain, not every case is equally appropriate. AI accuracy tends to follow something akin to a Pareto distribution: eighty percent of routine cases, in most fields, reach sufficient reliability relatively quickly. The remaining twenty percent (the high-variance, novel, high-consequence cases) take much longer, or may never be fully handleable by a model trained on what has already happened.
It is the long tail of the unique cases – the 20% that AI cannot handle reliably is not randomly distributed. It’s systematically the high-variance, high-consequence, high-novelty cases. So the principle isn’t just “deploy where reliable” but “deploy where reliable and the failure mode is recoverable.”
The refinement over time aspect is what makes it Pareto-like rather than just a competence threshold. As accuracy improves in a domain, the boundary shifts: more cases become handleable, the human reserve shrinks, but never to zero because there will always be a distributional tail the model hasn’t seen. All the more the importance of the first principle.
IV. The Principle of Preserved Friction
This evolved from a personal credo. That not all friction is inefficiency; some forms of difficulty are developmental.
There is a kind of undocumented learning that happens when one grapples directly with material over time: reading deeply, writing slowly, struggling with first principles, sitting with raw facts before synthesis arrives. The work shapes the thinker even as the thinker shapes the work.
AI removes enormous amounts of cognitive friction, often beneficially. Tedious labour can now be accelerated or eliminated altogether. But some forms of friction are not merely obstacles to output. They are part of how judgement, taste, intuition, memory, and understanding are formed and developed.
A world optimized entirely for cognitive convenience may inadvertently weaken the very faculties required to use AI well. The challenge therefore is not rejecting AI assistance, but preserving sufficient direct engagement with reality, thought, and creation that human cognition continues to deepen rather than merely accelerate.
Safeguard raw time with things.
And a fifth principle, which the other four presuppose, as suggested by AI.
V. The Principle of Transparent Training
The above principles operate at the level of use. What they don’t address is the training layer — who decides what the root system looks like before any organisation fine-tunes it. The audit principle could in principle extend there: foundation models whose training choices are themselves subject to cross-model and cross-institutional scrutiny, not just their outputs. A kind of constitutional audit of the model before deployment, not just of its responses after. That might be the fifth principle — transparency and auditability of training, not just of inference
The root system matters as much as the branches. Auditability of what a model was trained on and for is the precondition of meaningful oversight at every layer above it.
This one is less for individual users and more for the industry itself: transparency about how models are trained, what values are embedded in them. It is beyond most of us to implement. But it is perhaps the mirror the industry needs to hold up to itself.
To conclude,
The above is not a complete framework. It is a beginning, a first attempt at foundational principles which uphold the entire usage structure.
The principles of any discipline take time to settle, through use, through failure, through accumulated judgement about what actually went wrong and what corrected for it. We are at the start of that process with AI.
What I find hopeful is that the structural problems AI poses are not entirely new. They are problems we have encountered before, in other disciplines, in other material. The deeper question is what becomes scarce once intelligence-like capability becomes abundant. Industrial society made physical capital scarce. The internet made attention scarce. It seems that the AI era may make judgement scarce.
And if that is true, then principles may matter more, not less, in the age of AI.