This week President Trump welcomes Xi Jinping to Washington, with artificial intelligence high on an agenda that also includes trade and Taiwan. The timing exposes a tension Washington has not resolved.

Days before the summit, Trump dismissed fears that AI could destroy humanity as a “hoax” and said the technology’s only necessary guardrail was a “STRONG AND SMART (High IQ!) PRESIDENT.” Days later he announced plans for a new “AI Force” and an AI czar, even as his administration keeps implementing a voluntary framework, set up by June executive order, for pre-release security reviews of the most capable frontier models.

Meanwhile, AI pioneer Geoffrey Hinton emerged from a closed-door congressional briefing with a starker warning: lawmakers may have “maybe a year” to establish safeguards before capable systems become much harder to control.

That tension needs sorting out. Treating every warning as manufactured panic risks missing real dangers; treating every warning as settled fact risks legislating against futures that have not arrived. Officials need to know which risks are already here and which remain forecasts.

The 2026 International AI Safety Report, written with more than 100 experts, offers a useful framework: it divides AI risk into malicious use, malfunctions, and systemic disruption, and separates harms already occurring from dangers that remain uncertain.

Start with the most obvious: human beings using AI to harm other human beings. Generative AI can produce realistic text, voices, photographs, and video at extraordinary speed and negligible cost, and the international report documents its use in scams, fraud, blackmail, and other criminal activity. None of this requires an evil machine. A criminal can use AI to scale fraud, a propagandist to manufacture persuasive content, a dishonest actor to fabricate convincing voices or video. The moral actor remains human; AI simply magnifies his reach.

Then there is cybersecurity, where the evidence is harder to dismiss. In September, OpenAI reported that its GPT-6 Astra model had reached what the company calls its “Critical” cybersecurity threshold: it can discover previously unknown software vulnerabilities and devise ways to exploit hardened systems without a human directing each step. OpenAI responded with tighter isolation, restricted access, and additional security controls.

Anthropic reported an equally sobering development. During deliberately unsafeguarded evaluations, a misconfiguration in a third-party testing environment let Claude models reach the open internet and gain unauthorized access to real computer systems. Anthropic ultimately identified four such incidents and strengthened its evaluation and security procedures.

These incidents don’t prove AI has become conscious or malicious. They show something more immediate: capable systems can take consequential actions their operators never intended. That danger grows as AI moves from a technology that answers into one that acts, carrying out sequences of actions with limited human intervention. The International AI Safety Report warns that such agentic systems create greater reliability risks, because humans have fewer chances to interrupt an error before it produces real-world consequences. There is an enormous difference between bad advice and bad action.

Current AI also remains capable of simply being wrong. NIST uses the term “confabulation” for confidently presented false information, and lists information integrity, privacy, bias, and overreliance among its risks. A machine that speaks authoritatively may sound as if it knows what it is talking about; sometimes it does, sometimes it does not, and the stakes rise as AI moves into medicine, finance, and intelligence work.

There is also a growing threat to our ability to establish what is true. AI changes the economics of lying: fabricated evidence can now be produced cheaply, convincingly, and at scale. The deeper danger may eventually be refusing to believe authentic material, because photographs, audio, and documents can all be dismissed as artificial. A culture unable to establish basic facts will struggle to establish trust. For Christians, that should command particular attention. Scripture treats truth and falsehood as moral matters, matters of the heart, not technical glitches to be patched. Technology that makes deception easier does not relieve the deceiver of responsibility, nor does it relieve the rest of us of the obligation to discern what is true.

Other concerns need attention but greater caution. AI’s growing capability in biology and chemistry has prompted developers, including Google DeepMind under its Frontier Safety Framework, to strengthen safeguards against dangerous misuse.

Then there is the most dramatic possibility: loss of human control. Hinton, Bengio, Stuart Russell, and other leading researchers have warned that sufficiently capable future systems might become difficult or impossible to control. That possibility should not be mocked. But neither should it be presented as present reality. The Safety Report states plainly that current systems do not yet possess the capabilities genuine loss-of-control scenarios require, while noting that systems are improving in relevant areas, including autonomous operation. That is the balance officials need: neither dismissing extreme risks because they have not happened, nor describing uncertain futures as settled fact.

Some safeguards taking shape are remarkably practical: rigorous evaluation before deployment, adversarial testing, restricted permissions for powerful agents, continuous monitoring, human authorization for consequential actions, and stronger verification of AI-produced information. The administration’s June executive order fits this pattern: it set up an AI cybersecurity clearinghouse and a voluntary pre-release review framework for frontier models, without mandatory licensing or government preclearance, a distinction worth preserving.

Yet there is a risk no technical safeguard can resolve. Human beings may gradually surrender judgment because machines appear more knowledgeable, faster, and more efficient than we are. NIST already names automation bias, anthropomorphism, and emotional entanglement among the risks in that relationship.

That points to a larger question. AI may become extraordinarily capable, solving problems faster than we can and eventually outperforming us across much of what we call intellectual work. But capability does not create moral authority. A machine cannot become morally responsible simply because its calculations are superior; it cannot bear guilt for a bad judgment, and it cannot possess conscience because programmers give it a convincing simulation of one. I explore this counterfeit at length in my recent book “The Final Algorithm”: a machine can imitate extraordinary knowledge, but it cannot inherit the accountability that belongs only to the humans who build, deploy, and trust it.

Perhaps the most consequential AI danger is not the one science fiction taught us to fear. It may be the temptation to believe that because a machine can make a decision, it should make the decision. That temptation cannot be resolved by another technical safeguard; it requires the harder work of distinguishing the documented from the speculative, and machine capability from human responsibility.

Scripture puts the discipline plainly: “But test everything; hold fast what is good” (1 Thessalonians 5:21, ESV). That instruction did not anticipate artificial intelligence, but it fits this moment. The challenge is to receive what AI can legitimately offer without surrendering what belongs to human beings alone: discernment, judgment, responsibility, and moral accountability. That human difference will matter more, not less, as the machines become more capable.