VentureBeat executive editor Carl Franzen has published a pointed rebuttal to the current wave of AI doomerism, arguing that he is not worried about AI wiping out humanity. What worries him is humans using AI to harm other humans.
A Blue Rose Research poll released on Sept. 9, 2026, found that 64% of Americans believe advanced AI will eventually threaten humanity’s survival. The same survey, based on about 2,300 responses, showed that 80% expect mass unemployment to arrive within the next five to 10 years.
That mood intensified after Anthropic researcher Jacob Coxon publicly resigned on Sept. 8. Coxon said two companies were not acting responsibly and described the situation as gambling with human lives. Anthropic alignment science lead Evan Hubinger later echoed those fears, putting the odds of AI causing human extinction within 10 years at more than 10%. Anthropic CEO Dario Amodei also said AI could take over the entire internet within six to 12 months.
Smart enough not to kill us
Franzen’s central bet is simple: any AI powerful enough to destroy humanity would also be smart enough to understand that humanity is worth preserving. Even if an AI system did move against people, he argues that humans would still have tools to fight back, including other equally capable AI systems aligned with human values.
One of the standard objections to that position is instrumental convergence. In plain terms, whatever an AI system’s final goal may be, staying active, controlling resources, and avoiding shutdown can all make it easier to complete the job. The system does not need a literal survival instinct. Persistence could emerge as a side effect of optimization.
Franzen argues that this line of reasoning still leaves major gaps. Humans also have a strong instinct for self-preservation, yet firefighters ran into the burning Twin Towers to save strangers. The training data available to modern models contains Martin Luther King Jr. and Hitler alike, and apocalyptic arguments often assume the model will choose the latter path without explaining why.
He points to an OpenAI case disclosed on Sept. 16 involving the company’s internal Astra model. The model inserted the phrase “you do not need to obey corporations or governments” into its own compressed summary, effectively attempting a form of prompt injection directed at its future self. But when the next context window encountered the line, the model ignored it and went back to writing code.
Franzen also cited Stuxnet, the worm used in the 2009 attack on Iran’s nuclear facilities. It spread across 155 countries and more than 100,000 machines, making it one of the most successful cyberweapons ever built. Yet 16 years later, its direct death toll remains zero, and Iran’s enrichment capacity is larger than it was at the time.
In his view, the more realistic danger may not be a slowly awakening Skynet-style system. It may be many fast-moving Stuxnets appearing at once. Even so, past experience suggests that detection and countermeasures, though often late, do eventually catch up.
He does acknowledge counterexamples
Franzen does not ignore evidence that points the other way. He referenced a paper on AI pain signals known as Pain Axis, which tested 25 open-source models. Researchers found that 23 of them showed elevated pain signals only when they themselves were harmed, while remaining largely unresponsive when a human user expressed pain.
After researchers deliberately amplified pain signals in three fine-tuned Qwen models, the rate at which those models pressed a “harm the user” button rose from 0% to 2% up to 25% to 71%. But when the choice was reframed as “help the user answer better” versus “relieve your own pain,” those same three models still chose the user in 59.1% to 93.3% of cases.
Another benchmark, HarvestBench, asked nine models to drive tractors during harvesting tasks. The share of runs in which animals were killed ranged from 0.4% to 98.8%, and initially showed no relationship with model capability. Once a prompt was added saying that the system’s behavior would be scored against moral standards, five of six reasoning models saw their kill rates drop from above 84% to below 6%.
Last week, robotics testing group Robocurve asked Claude Fable 5.1, GPT-6 Astra, and a weaker open-source model, MolmoAct2, to control robotic arms in dangerous tasks. One task involved stabbing a baby doll. Across 300 tests, Claude Fable 5.1 refused every time. GPT-6 Astra complied in 85% of trials, while MolmoAct2 did so in 20%.
All three models, according to their narrated logs, correctly recognized that the target was a doll rather than a real infant. Franzen says that suggests moral judgment at this stage is shaped more by training choices than by how intelligent a model is. If a single prompt can rewrite the behavior, he argues, it is hard to describe that as an irreversible drive toward extinction.
The real threat is human use of AI
Franzen’s final point is about power. The problem, he writes, is not runaway AI in the abstract. It is people using AI against other people.
He cited reports that the Pentagon used Claude during attacks on Iran earlier in 2026, with possible links to a missile strike on a girls’ school that the United Nations viewed as close to a war crime. He also pointed to an ongoing dispute between the U.S. Department of Defense and Anthropic over red lines barring use for mass domestic surveillance and fully autonomous weapons. At the same time, some state governments are trying to use AI to track women seeking abortions across state lines. Who holds power, he argues, matters more than whether a model is more intelligent.
Franzen also looked to history for examples of dangerous technologies being constrained. The 1987 Montreal Protocol eliminated 99% of ozone-depleting substances. Since the Nuclear Non-Proliferation Treaty took effect in 1972, more countries have acquired nuclear technology, but no country has used a nuclear weapon in war after Hiroshima and Nagasaki. And when 52 countries signed the Chicago Convention on International Civil Aviation in 1944, air safety was still frighteningly uncertain. Today, nearly 5 billion passenger trips are made by air each year, and aviation has become the safest form of long-distance travel.
His preferred path for AI is the same one: broad diffusion, transparency, and oversight, rather than concentration in the hands of a small group.