The question of when — or whether — AI will reach human-level general capability has been a fixture of tech discourse for years. What's changed in 2026 is that the people making the predictions are now the ones building the systems, the timelines have compressed dramatically, and the definitions remain slippery enough that nearly any outcome can be made to fit nearly any forecast. Scientific American took a careful look at the state of current models and what the research actually says about the path to AGI. The Washington Post published an opinion piece arguing that doomsday predictions are looking sillier as the technology matures. Both are worth reading, partly because they disagree.
The compressed timeline camp is led by the labs themselves. Dario Amodei wrote last year that AI smarter than a Nobel Prize winner across most relevant fields could arrive as soon as 2026 — a statement he has not walked back as that year has arrived. Sam Altman has sketched a roadmap in which OpenAI reaches intern-level research capability by September 2026 and full "legitimate AI researcher" capability by 2028. Demis Hassabis, perhaps the most technically precise of the three, has said AGI is five to ten years away and that "no one really knows" when superintelligence follows. These are meaningfully different timelines from leaders of the three most advanced labs, which suggests either that the term AGI means different things to each of them, or that genuine uncertainty exists even among the people with the most information.
Scientific American's analysis focuses on what current systems can and cannot do. The honest summary is that frontier AI models show striking performance on well-defined tasks — coding, mathematics, reasoning over text — while remaining brittle in ways that are hard to characterise cleanly. They don't generalise the way human intelligence does. They fail on problems that require genuine novelty rather than recombination of training patterns. They are unreliable in ways that are hard to predict in advance. Whether these limitations are fundamental or merely engineering challenges that more compute and better training will resolve is the crux of the disagreement, and no one has a definitive answer.
The Washington Post's contrarian take is that the doomsday framing — AI as existential risk in the near term — has not aged well. The specific predictions made by prominent "doomers" have repeatedly failed to materialise on their stated timescales, the technology has developed more incrementally than catastrophically, and the most immediate AI harms are prosaic (job disruption, misinformation, legal uncertainty) rather than existential. This is a reasonable observation about the track record of specific predictions, though it doesn't settle the underlying question about long-run trajectories.
What's notable about the current moment is that AGI has gone from an abstract philosophical debate to a near-term planning horizon for governments, investors, and regulators. OpenAI is now publishing formal recommendations for policymakers about preparing for AI progress. The Pentagon is fighting over AI safety guardrails in ways that implicitly acknowledge that AI systems might one day need such guardrails. Whether AGI arrives in 2026 or 2036 or not at all in any meaningful sense, the expectation of it is already reshaping decisions being made right now — and that may matter more than the timeline itself.