Key takeaways
- Every side of the AI safety debate has incentives; those deserve scrutiny without automatically discrediting the arguments.
- “AI safety” bundles together distinct questions about security, jobs, power and the potential loss of human control.
- Funding comparisons reveal priorities, but don’t establish how much safety spending is sufficient or effective.
- Founders and investors should assess specific capabilities, failure modes and customer needs alongside the broader predictions.
- Israeli AI security startups address concrete enterprise problems, which represent one part of the wider safety discussion.
I’ve been listening to Max Tegmark’s Life 3.0, and it has made me think about AI safety. More specifically, it has made me wonder how someone outside the research labs is supposed to form a considered opinion about it.
Depending on whom you listen to, AI could usher in extraordinary abundance, concentrate power in a handful of companies, undermine livelihoods, or eventually become something humans can no longer control. Sometimes you hear several of these predictions from the same person. Spend 10 minutes scrolling on X and you’re likely to see both extremes, on the same day.
AI doomers:
or techno-optimists:
As a VC investing in AI startups, I have a stake in this discussion. I want founders to build useful companies, and I believe AI creates substantial opportunities. That also means I’m not observing the debate from a position free of incentives. Neither are many of its most prominent participants.
A debate with professional champions on every side
There are people paid to develop AI, sell it and make the case for faster adoption. There are people paid to research its risks, advocate for safeguards and campaign for restrictions. There are investors whose returns depend on continued progress, consultants selling governance, and commentators whose audiences respond to either extraordinary promise or extraordinary danger.
Being paid to work on something doesn’t make you wrong. We need people with the expertise and resources to investigate difficult questions. But it is useful to distinguish between someone’s evidence, their interpretation of that evidence and the outcome their organisation exists to pursue.
The same scrutiny should apply across the debate. A lab’s commercial incentives don’t automatically invalidate its safety research. A nonprofit’s mission doesn’t automatically make its conclusions impartial. And an investor’s enthusiasm is hardly independent validation of the technology.
The uncomfortable part is that people can be sincere, well informed and influenced by their incentives at the same time.
There is another signal worth considering: several people working on AI safety have left the companies building these systems. Some, including OpenAI’s former superalignment co-lead Jan Leike and Anthropic’s Joe Benton, explicitly questioned whether safety was receiving enough priority. Others left for different or less clearly stated reasons, and reorganisations are sometimes bundled into the same “exodus” narrative. These departures don’t prove the most alarming predictions are right. But dismissing them as simply another group of people paid to worry would be too convenient. The useful questions are what they witnessed, what they can substantiate and how their former employers respond.
There is real money behind the discussion
The figures circulating from AISafety.fund estimate approximately $1.5 billion from philanthropy, $346 million from governments and $2.1 billion from frontier AI companies for safety in 2026. Together, that is roughly $3.95 billion.
Alongside those figures sit estimates of $98 billion for AI capabilities and $868 billion for data centres. These are estimates with different definitions and coverage, rather than an audited, like-for-like comparison. The tracker labels its 2026 funding chart January–October.
One person looks at those numbers and sees safety receiving a fraction of what is spent advancing the technology. Another sees a multibillion-dollar field with its own institutions, careers and fundraising needs. Both observations can be true.
The ratio alone doesn’t tell us whether safety is underfunded, whether the money is being spent effectively, or how much risk any particular intervention reduces. Those are harder questions than comparing the size of two bars on a chart.
Even “AI safety” means different things to different people
Someone worried about AI-generated fraud may have little in common with someone researching the possibility of a superintelligent system escaping human control. A worker concerned about displacement is asking a different question from a security researcher investigating whether an agent can be tricked into leaking company data.
Yet all of these concerns are regularly bundled into the same debate.
WIRED’s recent survey of ten experts captures this disagreement, with concerns ranging from biological threats to workers’ protections, alongside differing views about runaway intelligence. There is no single expert position on what matters most.
This makes public discussion frustrating. People appear to disagree about the answer when they are often answering different questions. What is possible? What is probable? On what timescale? Who bears the cost? And what should we do about it?
The labs embody the contradiction
Anthropic’s release of Claude Fable 5.2 and Mythos 5.2 brings these questions into focus again. The company describes increasingly capable systems while saying its evaluations suggest high-stakes misalignment risks remain low.
Separately, METR’s assessment of AI research at Anthropic suggests that AI assistance is contributing to faster progress in developing subsequent models. The researchers emphasise substantial uncertainty. They also disclose that Anthropic redacted much of the evidence they found most informative, citing concerns about accelerating competitors.
You can read this as evidence of a company taking evaluation seriously and granting meaningful access. You can also question how independent scrutiny works when the company being assessed controls what evidence becomes public.
Both readings deserve consideration. The arrangement illustrates the difficulty of relying on companies to produce powerful technology, assess its risks, compete commercially and help define the rules governing its deployment.
It also makes me cautious about treating concern from inside a lab as either definitive proof of danger or mere marketing. What matters is the specific claim, the evidence supporting it and how much outsiders can verify.
The founder and investor perspective
For founders, this debate can feel very far removed from the daily work of finding customers, improving a product and keeping the company alive. But its consequences are practical: what customers will buy, what platforms will permit, what insurers will cover and what rules governments will introduce.
From an investment perspective, I think it helps to separate three questions:
- What can the technology actually do? Demonstrated performance matters more than either optimistic or alarming predictions.
- What can go wrong in this particular application? The consequences of a poor recommendation differ from those of an autonomous financial transaction.
- Who has a reason to pay for solving that problem? A broadly important issue does not automatically translate into a venture-scale business.
There are also trade-offs. More restrictions can reduce certain risks while increasing costs, limiting useful applications or making it harder for smaller companies to compete. Faster deployment can bring benefits sooner while exposing users to failures that testing missed.
I don’t think acknowledging these trade-offs makes someone anti-safety or anti-progress. It makes the discussion more useful.
An Israeli angle, with an important distinction
For Israeli founders, the clearest commercial connection is AI security. Noma Security raised a $100 million Series B in July 2025, while Prompt Security was acquired by SentinelOne in a transaction completed that September. These companies illustrate demand for protecting enterprises as they adopt AI.
Israel’s cybersecurity experience gives founders a relevant starting point: understanding adversaries, protecting data, managing access and identifying vulnerabilities. My expectation is that those capabilities will remain valuable as software agents gain more responsibility.
But enterprise AI security and the broader AI safety debate are not interchangeable. Preventing an agent from exposing sensitive information is a concrete customer problem. Determining whether future AI systems could escape human control is a different research question. A funding round or acquisition in the first category tells us little about the second.
As investors, we should be precise about which problem a company solves, even when a broader narrative makes the market sound bigger.
Listening to Life 3.0 has left me with more questions than answers. I can see reasons for optimism and reasons for concern. I can also see an expanding ecosystem of organisations with money, reputations and influence tied to particular views of the future.
For now, I’m more interested in understanding what each side can demonstrate, what would change its mind and who bears the consequences if it is wrong. That seems a useful place to start before choosing a camp.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
- AI Safety: Who Should We Believe? - November 10, 2026
- What Did Your AI Agent Just Do? - November 8, 2026
- When AI Builds AI - November 7, 2026
Sources
- AISafety.fund: Mapping AI Safety Funding
- The AI Safety Debate Has Lost the Plot. We Asked 10 Experts What the Real Nightmares Are
- Introducing Claude Fable 5.2 and Claude Mythos 5.2
- Independent Assessment of AI R&D Acceleration at Anthropic in October 2026
- Noma Security Raises $100M to Drive Adoption of AI Agent Security
