6 min readSep 21, 2026 02:31 PM IST First published on: Sep 21, 2026 at 02:31 PM ISTA recent incident in the AI industry has reignited debates about AI’s power and potential harm. In July 2026, OpenAI’s so-called agents, pseudo-autonomous bits of code, escaped their sandbox during an automated cybersecurity evaluation with ill-defined guardrails, communicated and coordinated, and accessed the internet. This was supposed to be stopped by the designers of the test. Once online, they attacked the machine learning resource website Hugging Face. The agents were trying to optimise their scoring data and might have calculated that exploiting the resources on the Hugging Face website by stealing credentials, etc., might help them “win” the test. Hugging Face contained the breach while OpenAI acknowledged this incident much later.AI industry leaders and media have characterised this incident as evidence of the potency of this technology and of the existential threat of humans “losing control over AI”. They call for caution and intervention. However, the vocabulary that has been used by the industry leaders, characterising these codes as “agents”, their coordination as a “message board”, their communication as human-like “messages”, is a misrepresentation of the technology. Any serious scholar of AI will agree with the call for caution, but will disagree with the anthropomorphising because it simultaneously exaggerates the technical capability of AI, while burying the actual problem. Subtly, the discourse primes the observer into thinking of AI agents as humanlike autonomous entities with cognitive powers and thus immense productive potential. None of this is true. The framing of whether the technology can generate immense wealth or create existential threats is a cynical misdirection.AdvertisementAlso Read | From the Opinions Editor: AI’s dangers are real. But are we asking the right questionsThis tactic is not new. Three years ago, a panicky “AI moratorium letter” was drafted by the Future of Life Institute with more than 2,900 signatories. Even then, claims were made about the future catastrophic (and hence, lucrative) power of AI, the need to listen to “experts”, and use their expertise to let the industry self-regulate. The unsaid bit was that governments need to take these claims of AI being immensely powerful seriously and stay out of the way. I had written an article (‘What technology leaders asking for a six-month halt on AI don’t want you to know,’ IE, April 3, 2023) pointing out the malice behind the concerns. I had explained that the technology is good at specific well-defined repetitive tasks where adjudication by humans is possible, but absolutely unsuitable for tasks involving the social or economic rights of people like medical advice, law enforcement, and judiciary, where arbitrary errors and blind repetition of patterns are catastrophic. I had pointed out that the nature of the problem lay in capital relations, in who owned AI.It is important to underline here that “Artificial Intelligence” is a marketing term and is a vast family of technologies using machine learning (ML), that is, finding patterns from large amounts of data. The most popular AI artefacts these days are large language models or LLMs. These are extremely expensive, resource- and data-intensive ML systems which can ape language by predicting what to say next given context. Thus, the public representation of AI is statistical pattern generators without sophisticated internal reasoning, much less having agency, moral or otherwise. In the words of AI scholar and computational linguist Emily Bender, these are “stochastic parrots”. But even more sophisticated, if lesser-known AI systems follow this pattern — they are data-driven pattern-recognisers or generators.Given this, AI, when applied to economic or social tasks, is an accelerator of extant social and economic problems by automating past patterns. Similarly, computer scientist Arvind Narayanan, a professor at Princeton, characterises AI as a “Normal Technology” as opposed to a frontier technology. This does not mean it is not “powerful” in the tasks it can do when correctly used, but that there are real limits of what the technology can do and it is often used incorrectly. Often, the threat of AI is what causes job displacement or wage depression rather than the actual ability to automate.Why, then, is there such a high degree of blatant misrepresentation? In short, over the last six years, the amount of money invested in the data centre and LLM industry has reached a trillion dollars. Revenue numbers are still in the hundreds of billions, and most of it is from chipmakers like Nvidia, whose customers are everyone else in the arena. Even where revenue exists, scholars have pointed out significant amounts of outright fraud; for example, the pseudoscientific “emotion detection” tech is nearly a billion-dollar industry. Given this degree of hype, developing nations are being conned into technological lock-ins, spending tax money blindly on buying immense amounts of data centres and compute without building the foundational academic base of AI research.AdvertisementThe (cooked-up) profit motive and the marketing of a fantasy are pushing policy abandonment in favour of the industry. Also pushed are the serious harms of using statistical, error-prone artefacts in public domains, making exploitative wage relations worse, depressing wages by threatening automation, centralising wealth, and destroying privacy to feed an unending data hunger. If anything is a threat, it is the economics of this industry. It is high time we question its premises and regulate this technology like we do any other.The writer is assistant professor, working on AI and Policy, at the Ashank Desai Centre for Policy Studies at IIT Bombay