Can anyone slow down the AI escalator so that we can stop to think?

There is an Indian story revolving around a king who asked his advisors to invent something where his feet would never get dirty when walking around. His advisors got into a frenzy, for a regal wish is like an order with the sword of Damocles hanging over their heads. One of the solutions was to cover the whole kingdom with leather so that no dirt would stick to His Majesty’s feet. But it was clearly expensive and impractical, for this would mean nothing would grow on the land beneath if it was covered fully with leather, or any other thing, for that matter. The ultimate logical solution was found in the invention of footwear. You wear it and your feet stay clean if you so desire.

Today, when large language models (LLMs) are driving the pursuit of artificial intelligence (AI), or artificial general intelligence (AGI) to rival human intelligence, it appears that the solutions we are seeking are similar to covering the earth with leather in order to keep your feet clean. The query the model needs to answer may be minor (Eg: What is the best food for cats?) but to answer it ChatGPT or Gemini use billions of texts stored in multiple data centres in order to figure out your probable answer based on pattern recognition and predictions on which word or phrase is most likely to follow which one. Based on how you word your query, the answer may vary, and sometime may even sound silly or random or wrong - also known as hallucinations.

OpenAI’s ChatGPT, which, a few months ago, reported having over 700 million daily users, requires - according to an IEEE report - 850 Mw of power daily - or 310 Gw annually. That’s more than India’s total non-fossil-fuel-based power capacity of 263 Gw. And this is only one LLM we are talking of. There are literally scores of LLMs being created, often using text and data sourced from the internet, often without paying licence fees to those who created the content in the first place.   

While power consumption will obviously reduce as better and more energy-efficient solutions are deployed in future, there is still a larger question about the utility of such LLMs, and what kind of social outcomes they will lead to. More so when LLM creators are advertising trivial uses like which hanky to use with which suit or to compare the performances of cricketers. If such trivia and nonsense is going to use up enormous quantities of power that could be used for better things, LLMs do not sound like a great idea to me. And when they are used to replace human labour and intelligence wholesale, they will cause incalculable social damage.

The standard, scripted answer is that even previous inventions were disruptive or wasteful in the short-term, but in the long term new kinds of jobs emerge. The argument here, that it all worked out in the past and so it will do so in future, is usually convincing, but for a few contra-indicators. Also, if I am going to use AI to completely stop thinking for myself, what are we really doing? What if this time the disruption is qualitatively different?  

The rush to invest endless amounts in AI has gotten so mad, that no one is stopping to think about its consequences for society. Many observers have pointed out obvious issues, but no one is really listening.

First, note the pace of change. The motor car or washing machine took decades to reach one million users - enough time for those threatened with disruption to learn new skills or professions. ChatGPT reached a million users in five days. Anthropic, which usually sells its frontier AI tech to businesses rather than consumers, is already closing in on OpenAI’s valuations (OpenAI is the owner of ChatGPT) and could soon top $1 trillion (that’s over $1,000 billion). Now consider how much more money that can potentially raise for AI. Anthropic is already testing a new AI agent called Mythos which can discover cyber vulnerabilities and, in the wrong hands, it can cause major businesses to lose huge sums or money, if not collapse altogether.

LLMs and agentic AI tools may not yet be threatening existing jobs wholesale, but as newer versions and AI agents replace human coding skills, they will certainly constrain new job creation, and ultimately threaten existing jobs. Consider the huge destruction of market valuations in software services companies with the mere introduction of Anthropic’s Claude Cowork. Human populations, especially those above a certain age, cannot upskill to keep up with that kind of pace of change.

Second, AI is supposed to make work easier for labour, but it is employers and owners of capital it really empowers. When a company inducts AI into its assembly lines or office work, it is not only seeking to improve labour productivity, but an overall reduction in costs, including labour costs. It is about substituting labour with capital. AI induction that does not work substantially in workers’ interest will ultimately lead to social unrest and unemployment. 

Third, AI will ultimately dumb down human ability, unless we make a deliberate choice to not allow it to do so in certain areas. If PhD scholars, students, and ordinary folks think all the answers to questions, or basic research work,  can be done by asking Google Gemini or ChatGPT or the Chinese DeepSeek, we will stop using human skills to seek better answers. All answers will be the canned ones provided by AI algorithms, including bad answers. To prevent this from happening, those who have to check for integrity will use more AI tools to detect automated work. The Supreme Court recently pulled up a junior judge for citing fake judgments using AI.  

Fourth, the real value of AI is when it is focused on reducing the possibilities of human error, fatigue or effort in specific fields. For example, an AI tool focusing on, say, the legal domain, despite the use of fake cases above, can be really useful for lawyers who need to study previous cases. This reduces human effort in ploughing through years of cases found on legal websites, but will still need human oversight and commonsense to weed out false cases. A doctor’s assistant can input a patient’s symptoms and also prepare a docket of possible diagnoses and special cases for his qualified boss to make a judgment call. Primary healthcare workers can use AI diagnostics in far-flung rural areas to separate small, well-known symptoms and diagnoses that require only standardised treatment from those that need expert scrutiny by real doctors and specialists. 

Put simply, more than LLMs, SLMs (small language models) focused on specific domains are possibly the most useful and less power-intensive. They also allow the person threatened with job loss to upskill in the same domain in which he previously had knowledge, and find new income-earning options in case he is actually laid off. An Economic Times report notes that many startups are using SLMs to deal with cost and privacy hurdles, using less cloud spaces and energy.

As for LLMs and AGI, evidence from the past is that anything that uses up so much capital will ultimately create monopolies. Even today, there is no real challenger to Google search or Maps or Gmail. In the case of SLMs, the chances of large monopolies are less likely, except in specific domains. We should worry, but probably we can live with limited sector-specific monopolies.

Fifth, AI companies may be creating mutually beneficial Ponzi schemes by investing in each other. Nvidia, which makes the graphic processor units (GPUs) that run AI programmes, is investing more than $90 billion in AI businesses (including OpenAI) which could ultimately be buyers of its expensive chips. OpenAI itself is investing $4 bn in AI startups. The incestuous relationship between one AI company and another, coupled with the massive investment rush to avoid losing their leads in AI technology, has pushed four of the five major tech companies (Amazon, Meta, Oracle, and Microsoft) into a negative cash-flow situation.

One wonders whether all the AI hype may one day end up causing a huge crash in valuations in tech, something like the dotcom bust at the turn of the century, at a time when tech is the only sector driving up the markets. If tech crashes, it will bring down a whole lot of other sectors that depend on it, not to speak of the massive erosion in pension fund asset values.

The world needs to pause and rethink, but no one is apparently in the mood to sit back because the fear driving AI is the fear of missing out (FOMO).

But there is no doubt we need to rethink our attitudes to technology: the mere fact that it is available does not mean it must be used or allowed to do real damage. AI cannot be left purely to market forces, and the power of capital to decide its evolution cannot be unchecked. 

In their book Power and Progress: Our Thousand Year Struggle Over Technology and Progress, authors Daren Acemoglu and Simon Johnson show how technological change is sold as some kind of huge driver of wealth and public prosperity, while its real impact lies in how much it empowers and enriches a small group of innovators and users while the vast majority are left with only the perception of benefits and progress. Google, Gmail and Gemini may seem like a huge benefit to people, but it generates billions of dollars in economic power to Alphabet, thus giving it the power to influence almost everyone’s choices. Free tech services often reduce citizens to consumers.

Acemoglu and Johnson say that tech razzmatazz makes us stop questioning the extreme optimism that AI generates, and suggest that there is nothing inevitable about any technology, and society should have the ability to choose which one it can adopt widely and which one can be discouraged. This does not - repeat NOT - imply that we must clamp down on AI or its use. Far from it, AI can have many really good uses, but its progress cannot be left only to those hoping to add several zeroes to their company’s valuations or shareholders’ wealth. 

A simple example: if a company intends to adopt an AI tool, its employees should have a say in choosing which part of AI to adopt and how the productivity gains will be shared between capital and labour. AI tools and models should be subject to commonsense public audits, and those damaging to society should be disincentivised. Though we can never really control technology and its progress - what some country proscribes another may accept - we simply need more societal say in which way technology develops, especially AI. More than governments wielding the axe, it is society which must do so.

This is not an easy thing to legislate or enable in a society where public opinion can easily be hijacked by corporate influencers, but any democratic society must have this at the back of its mind. A senseless rush to create things that damage social trust and human values is not something anyone can, or should, support.

But is anybody listening to saner voices right now?