The sceptics

In one sentence: AI is impressive and useful, but claims about imminent superintelligence, whether utopian or apocalyptic, run far ahead of the evidence. The real questions are more ordinary, and more urgent.

Sceptics are often misread as "anti-AI". Most are not. Many are AI researchers who use and build these systems every day. What they doubt is the story told about AI, not its value.

Three kinds of scepticism

1. Sceptics of the current path to human-level AI

Yann LeCun, a Turing Award winner and Meta's chief AI scientist at the time, said in January 2025: "There's absolutely no way … that autoregressive LLMs, the type that we know today, will reach human intelligence." He added that "we're not even close to matching the understanding of the physical world of any animal, cat or dog", and that scaling "is saturating" (PYMNTS report of CES talk, 8 Jan 2025). LeCun's point is about today's methods. He does not argue that human-level AI is impossible in principle. → Tracker: CR-T03

Gary Marcus, a cognitive scientist and entrepreneur, argued in a widely discussed 2022 essay that deep learning "is at its best when all we need are rough-ready results", and that the so-called scaling laws "aren't universal laws like gravity but rather mere observations that might not hold forever." He called for "hybrid" systems combining learning with symbolic reasoning. The essay ends on a hopeful note: "For the first time in 40 years, I finally feel some optimism about AI" (Nautilus, 10 Mar 2022). → CR-K16

How this has been challenged. Since 2022, language models have improved substantially on many tasks. METR measured the length of tasks AI agents can complete doubling roughly every seven months (METR, Mar 2025). Supporters of scaling point to this as evidence against a "wall". Sceptics reply that benchmark gains don't equal genuine understanding or reliability. The International AI Safety Report 2026 describes capabilities as "jagged": strong in some areas and weak in others (executive summary). This question is genuinely unresolved, and we mark it as disputed in the tracker.

2. Sceptics of existential risk

The AI researcher and educator Andrew Ng said in 2015 that worrying about evil, killer AI was like worrying about "overpopulation on Mars": not impossible one day, but not a productive problem to work on now (The Register, 19 Mar 2015, search-verified; the quote's origin is traced by Quote Investigator). → CR-K17

Arvind Narayanan and Sayash Kapoor (Princeton) argue in "AI as Normal Technology" (April 2025) that:

AI is transformative, but in the way electricity was: its effects unfold over decades as institutions adapt.
"Superintelligence", as usually imagined, is an incoherent concept.
Catastrophic misalignment is the most speculative of the risks.
Existential-risk probabilities lack a sound basis.

They see systemic risks as more pressing: inequality, concentration of power, erosion of trust. They favour policies built on resilience and transparency over attempts to restrict who may build AI (Knight First Amendment Institute, 15 Apr 2025). → CR-K14, CR-K15

3. Critics focused on present-day harms

Emily M. Bender, Timnit Gebru and colleagues, in the 2021 paper "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?", set out risks of ever-larger language models:

environmental and financial costs;
training data too large to document properly, which can encode bias;
research incentives that reward size over understanding;
harms from fluent synthetic text, including misinformation and people trusting outputs too readily.

They recommended careful dataset curation, documentation and more thoughtful research directions (ACM FAccT 2021, search-verified). The phrase "stochastic parrots" (systems that string words together from statistical patterns, without understanding) entered everyday discussion. → CR-K18

Their case, stated fairly

AI has a long history of over-promising. Marcus recalls a prominent 2016 prediction that radiologists would soon be replaced. By 2022, he notes, the consensus was that humans and machines complement each other.
Benchmarks aren't the real world. Passing tests is not the same as working reliably in hospitals, courts or cars. Narayanan and Kapoor argue that many benchmarks lack "construct validity": they don't measure what they claim to.
Diffusion takes time. Even powerful technologies spread at the pace of regulation, training, procurement and trust.
Present harms are certain, future ones speculative. Bias, misinformation, surveillance and labour exploitation are happening now and deserve attention now.
Doom and hype can serve the same interests. Both make AI seem more powerful than it is, which can attract investment or justify concentrating control.

Criticisms of the sceptics

From the safety-focused and those most worried about catastrophe: "It isn't happening yet" isn't a plan. Capability has repeatedly surprised experts, and preparation takes years. The International AI Safety Report 2026 notes systems "are improving in relevant areas".
From accelerationists: sceptics underestimate how fast capabilities and adoption are moving. The UAE, for example, reached 73.3% measured generative-AI use among working-age people by June 2026 (Microsoft, 21 Sep 2026).
A fair point of common ground: most sceptics agree that present harms matter and that AI should be governed. The disagreement is mainly about priorities, not about whether to care.

What would prove the sceptics right, or wrong

Sceptic expectation	Would be supported by	Would be challenged by	Tracker
Current methods plateau short of human-level general ability	Slowing gains on hard, novel tasks; persistent reliability problems; labs shifting to new architectures	Continued rapid gains on long, open-ended tasks (e.g. METR time-horizon trend holding)	CR-T03, CR-K16, CR-K22
Impacts arrive over decades, not years	Gradual productivity and employment change; slow institutional adoption	Rapid, measurable economy-wide shifts within a few years	CR-K14, CR-E03
Existential risk remains speculative	No evidence of systems acting autonomously against human interests outside tests	Real-world incidents of deceptive or power-seeking behaviour	CR-R04, CR-R06
Present harms are the main issue	Documented cases of bias, misinformation and surveillance harms	n/a (both can be true)	CR-R03

What this means for you

Sceptics offer a valuable habit of mind: ask what the evidence actually shows, not what the headline promises. That is useful whether you are reading a bold product launch or a frightening prediction. It is also a reassuring reminder that many of the world's most knowledgeable AI researchers look at the same systems you use and see powerful tools, not minds about to take over.