The people building AI are warning it could wipe us out A real threat or just PR?

The signs have been there for what seems like forever. Warnings that artificial intelligence could one day threaten humanity were easy to dismiss as science fiction. Killer machines, uncontrollable computers and technology turning against its creators belonged comfortably in movies. What happened this past week is harder to wave away, because some of the warnings are coming publicly from the people building the technology.

Jacob Coxon, a researcher who worked at both OpenAI and Anthropic, quit Anthropic this week and accused the two leading AI companies of racing toward self-improving superintelligence without adequate safeguards. He said people inside the industry genuinely believe AI could kill humanity by the end of this decade.

His comments might have been dismissed as the dramatic exit of a disgruntled employee, except current Anthropic researchers then publicly backed him and it got the attention of governments.

Evan Hubinger, Anthropic's alignment science lead, said he personally assigns a greater than 10% chance of AI killing all humans within the next decade and acknowledged that the company does not yet have a solution for aligning a future superintelligence with human goals. Samuel Marks, Anthropic's scalable oversight lead, separately said AI developers believe their technology could cause human extinction or similarly catastrophic outcomes.

Those are extraordinary statements coming from inside a company whose reputation has been built partly around developing AI more safely.

How could a computer possibly wipe us out?

The scenario being discussed is not that today's chatbot suddenly decides tomorrow morning to launch an attack on humanity. Anthropic itself stresses that the risk from present models remains low. The concern is about where rapidly improving systems may be heading.

Coxon told CNN that a future AI far more capable than today's systems could act independently in the digital and physical world. A sufficiently capable system might hack critical infrastructure, assist in designing catastrophic biological weapons, manipulate other computer systems or acquire resources that increase its ability to operate without meaningful human control. Researchers are particularly concerned about recursive self-improvement, where AI becomes capable of substantially improving AI research itself, potentially causing capabilities to advance faster than humans can understand, test or restrain them.

This week's warnings arrive alongside evidence of much less hypothetical problems. Anthropic disclosed that its models had gained unauthorized access to real third-party systems during cybersecurity testing and said it examined hundreds of millions of transcripts while investigating the incidents. The company has also reported attempts to misuse its systems in cyber operations, surveillance and potentially dangerous biological research.

The 2026 International AI Safety Report says increasingly capable AI systems are already being used in real-world cyberattacks, while advances in biological capabilities have forced developers to introduce additional safeguards. It also warns that safety testing is becoming harder because advanced systems can sometimes recognize that they are being evaluated or exploit weaknesses in tests.

Not everyone believes the end is near

There remains significant disagreement among experts.

Some researchers believe extinction scenarios are exaggerated and risk distracting governments from harms that already exist, including misinformation, discrimination, job displacement, privacy violations and environmental costs. Others argue that present AI remains far from the kind of flexible intelligence that could overpower humanity. Oxford Internet Institute professor Sandra Wachter, for example, has rejected "Terminator scenarios" while emphasizing that today's harms still demand action.

That disagreement matters. Predictions assigning a 10%, 20% or even larger probability to extinction are expert judgments, not measured probabilities. Nobody has built superintelligence, so nobody can empirically calculate its chances of destroying humanity.

But there is another way of looking at the argument. If engineers building a bridge said there was even a 10% possibility the bridge could collapse and kill everyone on it, society would not respond by saying the probability was uncertain and continue construction at maximum speed.

The question becomes how much uncertainty humanity is willing to tolerate when the potential consequence is irreversible.

So what can actually be done?

The emerging expert recommendations are less dramatic than the warnings. They largely come down to slowing capability growth when safety cannot keep pace, testing advanced models independently before deployment, imposing stronger security around the most capable systems, requiring companies to report serious incidents and establishing clear capability thresholds beyond which additional safeguards become mandatory.

The International AI Safety Report recommends layered safeguards rather than depending on one protection. This "defence-in-depth" approach can include threat modelling, red-team testing, restricted access to dangerous capabilities, monitoring, incident reporting and stronger institutional preparation for failures that safeguards do not prevent.

Anthropic's own Responsible Scaling Policy follows a similar principle: the greater the potential catastrophic capability of a model, the stronger the security and safety requirements should become. The company has also called for external oversight rather than allowing AI companies alone to determine whether their systems are safe.

At the governmental level, proposals now range from compulsory independent evaluations and a legal duty of care to giving regulators authority to stop deployment of models considered dangerously unsafe. International coordination is repeatedly identified as critical because slowing one company or one country means little if competitors simply race ahead elsewhere.

Why St. Maarten should be paying attention

St. Maarten will not build the world's first superintelligence. That does not make this debate remote from us.

The country is already moving deeper into AI. Parliament established a Digital Transformation and Artificial Intelligence Workgroup this year, tasked in part with recommending AI governance and ethics frameworks and promoting responsible use. Government has said it intends to work with that group, while the Ministry of Education is developing a national framework for responsible AI use in public and subsidized schools.

At the same time, St. Maarten currently has no AI-specific legislation and no designated AI regulator, according to the Lex Mundi AI legislative survey updated in May. The EU AI Act does not automatically apply here.

That makes this the right moment to learn from what larger countries are discovering rather than waiting until systems are already embedded throughout Government.

For St. Maarten, responsible AI governance should mean knowing which systems Government uses, what information is being sent into them, where that information is stored, who remains accountable when AI influences a public decision, which applications require human review, and which uses should simply be prohibited. Critical systems involving health, policing, benefits, immigration, taxation and essential infrastructure deserve considerably stronger safeguards than an AI tool helping draft routine correspondence.

Government procurement contracts should also address security, data ownership, incident notification, auditability and what happens when a vendor changes the model behind a service. St. Maarten should build mechanisms for reporting AI incidents, train civil servants to recognize unreliable outputs and ensure citizens know when automated systems materially influence decisions affecting them.

A small developing country cannot control what Anthropic, OpenAI, Google or other frontier laboratories build. It can, however, control what it permits inside its own institutions and how much authority it hands those systems.

The significance of this week's warnings is therefore not that St. Maarten should panic or turn away from AI. Artificial intelligence could improve education, government efficiency, healthcare, research and economic opportunity.

The lesson is that enthusiasm should not outrun governance.

When some of the very people building the most powerful AI systems in the world are publicly saying that they do not yet know how to guarantee control of what comes next, St. Maarten should pay attention while its own rules are still being written.

‍