Superintelligence in 2027? Why Forecasters Doubt It

Chart-style illustration of AI timeline forecasts for superintelligence in 2027

The case for superintelligence in 2027 got a prominent challenger this week, and it comes from someone who first agreed with the underlying evidence. After reviewing METR’s report on AI acceleration at Anthropic, the forecaster says the data doesn’t support the short timeline now being drawn from it.

The target is the AI Futures Project’s Q3.5 update. It builds on the METR findings and gives TED-AI, short for top-expert-dominating AI, a 35% to 65% probability by the close of 2027. That label covers a system equal to or better than leading human specialists across nearly every mental task.

Faster AI research is not the same as superintelligence

The AI Futures Model treats the jump from a superhuman AI researcher to TED-AI as one more step of AI research. If that holds, any sign that AI is speeding up AI development also points to superintelligence arriving soon.

People with longer timelines reject that link. They argue the second jump is held back by outside factors, such as collecting data in fields beyond AI research. METR’s report covers only automation of AI research, so it says nothing on whether that automation leads to superintelligence. Anyone who expected progress to stall before then has no new reason to expect superintelligence in 2027.

Experimental compute may explain much of the speedup

METR said faster scaling of training compute was not a big part of the acceleration. It did not rule out experimental compute, and that gap matters. Earlier estimates, including calculations by Ajeya Cotra, assumed experimental compute stayed on trend.

Ryan Greenblatt’s model shows what’s at stake. Suppose experimental compute doubled relative to its trend. Overall capability progress would rise 1.24x with no AI uplift at all. It’s also likely that compute did grow above trend. When researchers become more productive, a lab has every reason to give them more compute. Assuming Anthropic didn’t do that would be strange.

The critique’s conclusion: experimental compute is probably about as important as AI labor uplift in explaining the acceleration, not a footnote to it.

A 12x productivity claim doesn’t survive scrutiny

METR cited informal polls in which employees tapped emoji on Slack to rate how much faster the newest model made them. The answers peaked at 12x in October 2026. METR treats those polls as very loose evidence, yet earlier explainers used the figure to show the acceleration could be driven entirely by AI uplift.

Plugging a 7x serial labor acceleration into Greenblatt’s model yields a speedup of roughly 1.6x that can be credited to AI. But a 7x boost has odd consequences. Anthropic would accept losing 98% of its researchers, or 91% of its experiment budget in compute, before giving up its best models. Few people at the lab would agree with that.

OpenAI asked its researchers a similar question. The typical respondent would sacrifice 30% of a team’s compute, or 40% of its staff, to avoid dropping back from frontier AI to GPT 5.2. Under the same assumptions, that implies a serial labor speedup near 1.3x. At realistic uplifts, the model predicts roughly 1.10-1.12x overall acceleration. The true figure is probably higher, the critique allows, but not by much.

Wording may have inflated the polls too. Asked for a multiple, people may say 12x. Few would say they were 1,100% more productive.

DeepMind is back in the race

SemiAnalysis had written DeepMind off as a frontier lab, pointing to staff leaving its reinforcement learning groups and weak compute allocation. The forecaster had agreed and treated AI as a two-horse race between OpenAI and Anthropic.

Gemini Carbon and Gemini Neon changed that. Half a year earlier, the models the public could compare were Mythos Preview, GPT-5.5 Pro and Gemini 3.5 Flash, a lopsided matchup. Now the gap looks far smaller. DeepMind’s research automation results also look closer to OpenAI’s than to METR’s picture of Anthropic. If better internal models drove frontier progress, DeepMind’s catch-up is hard to explain.

Why updates have been lopsided

After METR’s report, some people said they felt physically ill, METR researchers urged readers to take 2027 seriously, the AI Futures Project issued an emergency update, and 80,000 Hours published an explainer. OpenAI’s more pessimistic report on research acceleration drew almost no comparable reaction. The critique reads that as a sign forecasters aren’t weighing all the available evidence.

What the dispute changes

The critique doesn’t rule out superintelligence in 2027. It rates the outcome as unlikely.

The author isn’t a long-timelines skeptic. They estimate we’re about 60% of the way to superintelligence compared with 2022, which points to 2029 by default. Two years sounds small, but it shifts priorities. It bears on whether to focus on the 2028 election, push for a pause now, or build AI safety capacity in China, which only pays off on slightly longer timelines.

The worry is that work paying off in more than six months gets dropped in a rush toward 2027. AI would still be risky if it arrived two years later than predicted, and betting everything on superintelligence in 2027 has real costs.

Source: blog.karthiktadepalli.com

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.