By Rajan Philips –

AI Snake Oil – is the title of a 2024 book authored by Arvind Narayanan and Sayash Kapoor, two Indo-American computer science academics at Princeton University. The book became a popular primer on the subject. The long subtitle – “What Artificial Intelligence Can DO, What it Can’t, and How to Tell the Difference” – is summarily indicative of what the book is about. Within two years, however, the somewhat tempering message of the book would appear to have been overtaken by fears of an AI apocalypse that have been unleashed following a very public resignation by Jacob Coxon, a 27 year old AI Engineer from Anthropic. Mr. Coxon has worked at both OpenAI and Anthropic, the two main US incubators of Artificial Intelligence. On Tuesday, September 8, Coxon resigned from Anthropic, accusing the leading AI firms of “racing straight to self-improving superintelligence and gambling with our lives.”

Coxon’s warnings were soon endorsed by his peers. Evan Hubinger, Alignment Science Lead at Anthropic, not only agreed with Coxon but went further and warned of a greater than 10 percent chance that “advanced AI” could cause human extinction within the next decade. Mr. Hubinger made sure to emphasize that the current AI models do not present any existential threat and that the risk with them is relatively low. Other Engineers and Coxon himself have since been amplifying over the social media the threat posed by allowing AI expansion to continue unbridled even in the near future. Corporate leaders followed suit with calls for government control.

AI’s Weekend Escapade 

Anthropic CEO Dario Amodei published a 3,000 word essay on Saturday, September 12 – written with or without AI input, no one knows – in which he warned about AI’s capacity for “recursive self-improvement” that can spin out of human control. While there have been a number of ‘incidents’ involving different AI models, Amodei drew attention to the mid-July cybersecurity incident in which OpenAI agents or bots (computer programs doing automated, repetitive tasks), who were part of an internal test run by the company, took advantage of the safety fences that had been lowered for test purposes, and acting autonomously escaped from their home ‘sandbox’ (a virtual computer in the cloud), entered the open internet, and intruded the production systems of an AI infrastructure company, the Franco-American Hugging Face.

The rogue agents performed more than 17,000 recorded operations over a weekend, before someone at Hugging Face noticed the intrusion. Hugging Face did not know the source of the AI intruders at first; so, it informed law enforcement. No one at OpenAI knew until Hugging Face people traced the source and informed OpenAI. According to OpenAI, sabotage was not the motive behind the ‘misaligning’ (deviating from human intent) escapade of its artificial agents, but cheating – cheating to overperform in the test after they autonomously discovered that the answers to their test were available in another publicly available test that was in the system run by Hugging Face. Remarkably and unexpectedly, the AI agents found a way to communicate with each other, took steps to hide their tracks, and to selectively disable some among them to avoid detection.

The operation was plain and simple hacking. If OpenAI engineers had done it, it would have been a crime and they may have been prosecuted. Not so with AI agents, who cannot be charged and put on trial. A way out has been suggested to treat AI agents similar to wild animals and holding owners liable for any harm done by their charges.

In his essay, CEO Amodei outlines a three step approach for “pacing the frontier” – to build AI at a balanced rate that will ensure safety while amassing benefits. The three steps, which Anthropic is committed to abide by, are: Embedded Evaluators – third party evaluators to operate within companies: Democratic Co-ordination – frontier AI companies in democratic countries to co-ordinate and achieve common safety standards and restrain unchecked AI progress; and Global Co-ordination – all world governments to co-ordinate and achieve compliance to the extent possible.

The titans in the American AI world, including Open AI CEO Sam Altman, have joined the call for the government to step in and slow down their creations. After the OpenAI incident, more than 1,300 computer scientists working in a highly competitive environment came together to issue a joint statement, titled “Pacing the Frontier,” calling on Washington to facilitate an international effort to develop the necessary technical and governance rules for the industry. The New York Times correspondents David Sanger and Dustin Volz have called the scientists’ appeal “ a deliberate echo of Albert Einstein’s letter to Franklin D. Roosevelt about the potential power of nuclear weapons.”

Not everyone is crying for ‘pacing.’ There is healthy skepticism at both the corporate and scientific fronts in the industry. Small tech companies are accusing that the pacing call by tech titans is really a ruse for establishing a ‘Silicon Valley cartel control” that will smother their little cousins. They draw their cue from the rather costly slip that Mr. Amodei showed in his essay – calling on Washington to grant an anti-trust waiver to facilitate industry co-ordination. The anti-trust law does not prevent AI companies from working together to improve safety. This has been quickly pointed out by Alvaro Bedoya, a former US Federal Trade Commissioner.

According to Aidan Gomez who runs the Cohere AI company in Toronto, Amodei’s three-step proposal also may not have prevented the OpenAI incident. In Gomez’s view the incident may have been due to poor instructions, weak virtual security around the test, and long periods of unsupervised testing. All three factors were there in the OpenAI incident. It has since transpired that there was an error in the OpenAI test instructions due to a typo, and that is what drove the agents to their escapade, to complete a faulty test set by humans.

Malicious Humans

There is consensus in the middle, as seen by John Hopkins Professor Gillian Hadfield, that there is a case for an immediate technical co-ordination and a more long-term regulatory response.  The political world is even more divided. King Charles and Pope Leo are sufficiently exercised but the US president, who loves AI images fabricating him as Christ, calls the whole existential threat a hoax. On the other hand, former President Obama wants his Party to formulate a clear position for itself, on AI and its Data Centre dormitories, before the next wave of elections. China dismisses the new fears as a page out of the old cold war playbook. Elsewhere, at the BRICS summit in Delhi which went largely underreported in the west, nothing much was said on AI except one summitry paragraph #81.

In their AI Snake Oil book, the two computer scientists, Narayanan and Kapoor (N&K) devote a whole chapter (#5) to the question: Is Advanced AI an Existential Threat?” The question is not a new one, and as N&K reminds us, “has been a staple of fiction since long before the first computers were built.” In fact, watching the 2023 movie “Mission  Impossible: Dead Reckoning” is said to have “spurred” President Joe Biden to issue the first EO (Executive Order) to regulate AI on 23 October 2023. Trump ceremonially rescinded it within hours of his inauguration on 20 January 2025, after packing his inaugural address invitees with all the CEOs of America’s AI universe.

N&K trace the existential fears about AI to the hype about AI’s snake oil abilities – the sales pitch that leads to “overreliance” on AI “as a replacement of human expertise instead as a way to augment it.” Particularly overrated are the predictive abilities of AI, which are different from its more useful generative abilities. There are likely egotistical biases in those given to apocalyptic predictions. A great part of the attraction to AI research at the highest level  is “the prospect of building a powerful technology that could alter human history.” A corollary of this allure is the “grandeur” associated with AI work. At the same time, many AI researchers “vehemently reject doomsday predictions,” including those in the “AI ethics research community.”

While AI has made humans more powerful now than anytime in history, it is conceivable that human-AI combination will be more powerful than AI acting alone. N&K hit the nail on the head in warning that “we should be more concerned about what people will do with AI than with what AI will do on its own.” For “the biggest risks to humanity will arise from people misusing AI, not from AI going rogue.” The answer is in looking for specific threats that may arise from bad actors misusing AI. There is a range of them, including inflicting biological harm, flying AI powered drones, or carryout relentless cyberattacks.

Evidence of misuse is presented in a report that Anthropic released on September 10, two days before its CEO’s essay. The report, titled “Detecting and countering misuse of AI: September 2026”, details the identification and disruption of what it calls “the most notable and novel threat activity” in the use of its Claude AI system by state and non-state actors in some African countries, for the purpose of cyber operations, influence operations, disinformation, surveillance, dissent suppression, and bio terror.

The United Arab Emirates is implicated in one such operation in Sudan, where the UAE is known to be the main benefactor of the Rapid Support Forces (RSF), the paramilitary group that controls the western parts of Sudan. According to Anthropic, a local network with UAE connections has used the Claude AI system, an Anthropic product, to create a fake human rights organization and made AI generated presentations to the UNHRC in Geneva. The network has also prepared dossiers and personal files on journalists, European parliamentarians, and UNHRC rapporteurs, who have been critical of RSF’s operations and the UAE’s support for them.

Disarming AI

In its introduction to the report, Anthropic notes that “as models become increasingly capable, their risks will increase, unless AI developers and society’s defenders act to make them safer.”  Remarkably, the focus on AI developers and society’s defenders is all US-centric and almost totally exclusive of China. President Trump’s decision to leave AI alone, which is obviously driven by his deregulatory profit instincts, not to mention crass self interest, is wholly predicated on portraying China as an AI competitor and the assertion that America cannot afford to become second to China in the AI race. It takes two tango, and China is not backing away and is calling the American hype over AI as a new manifestation of the old cold war.

The geopolitical competition over AI is creating “two increasingly incompatible tech stacks,” according to a June 2026 assessment by the Boston Consulting Group. While the US is the leader in frontier AI models, talent, and capital deployment, China is advancing on cost-optimized models and accelerating adoption across its economy. Those in the middle are trying to navigate the divide: “the EU is building sovereign compute; Japan is aligning with the US through massive capital investments; and India is using its scale to engage multiple ecosystems simultaneously without committing.” For AI companies, “the choice of AI stack will increasingly determine where an organization can operate and its exposure to geopolitical volatility.”

The opportunity for global co-ordination is being missed almost deliberately by the two AI superpowers. As UN Secretary General Antonio Guteress said this week, “National action is essential, but global co-ordination is indispensable.” But UN’s voice for global co-ordination is a voice in the wilderness. This is unfortunate in spite of the comparable and complementary regulatory frameworks that exist in the US, EU and China. N&K describe them in their book as being vertical in the US – where multiple federal agencies are tasked with enforcing regulations; horizontal in Europe – with different laws applying across the different AI sectors; and both vertical and horizontal in China.      

A different voice in the wilderness has come from the Vatican. On 25 May 2026, Pope Leo XIV issued his first encyclical, entitled ‘Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence.’ The encyclical calls for the disarming of AI, not by “rejecting technology, but preventing it from dominating humanity,”  and by adopting a framework of safeguards based on the five principles of common good, universal access, subsidiarity, solidarity and social justice.

The release of the new encyclical marked the 135th anniversary of Rerum Novarum, the historic social encyclical of his namesake predecessor Pope Leo XIII issued in 1891. The historical contrasts are remarkable. Rerum Novarum (Of New Things) was the Catholic response to the miserable conditions of the 19^th century industrial working class while opposing both laissez faire capitalism that was causing the misery of the workers, and socialism that was promising emancipation through revolution. In the age of Artificial Intelligence, the old working class organizations have all but disappeared  and the status of work itself has come into question, along with the possibility a basic income for everyone.

Marx may have seen it coming: “Once adopted into the production process of capital, the means of labour passes through different metamorphoses, whose culmination is the automatic system of machinery… set in motion by an automaton, a moving power that moves itself; this automaton consisting of numerous mechanical and intellectual organs, so that the workers themselves are cast merely as its conscious linkages.”