When Will AGI Arrive? What the Evidence Actually Shows

Ask five AI researchers when AGI will arrive and you'll get five different years — not because they disagree on the facts, but because they're often not answering the same question. What follows is not another prediction. It's a look at what the actual evidence says: how fast measured capability is compounding, what forecasters who study this for a living currently believe, and why the honest answer is "it depends on the definition" rather than a date on a calendar.

The definition problem comes first

OpenAI's charter, published in 2018, defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." That's a useful mission statement, but it's not a test anyone can actually run. It says nothing about which tasks count, how "most" is measured, or what counts as outperforming — and it leaves out forms of human capability, like taste or emotional judgment, that resist economic measurement entirely.

Other researchers avoid the term altogether and define a narrower, checkable milestone instead. The Forecasting Research Institute's 2026 LEAP survey, for example, asked forecasters about a specific bar: a commercially available system that can outperform the 90th percentile professional in 90% of the non-physical tasks they perform, at no more than 5x the cost of the human doing the work. See the published survey results. That's a very different question from OpenAI's charter language, and it produces a very different — and more defensible — timeline.

This is the part that gets lost in headlines: "when will AGI arrive" is at least three separate questions — when will a system pass a specific benchmark, when will it be economically deployed at scale, and when will a broad expert consensus agree the general capability threshold has been crossed. Those three answers can be years or decades apart.

What the trend lines actually show

Whatever you call it, the underlying capability curve is real and measurable. METR, a nonprofit that evaluates frontier AI systems, has tracked the length of software tasks an AI agent can complete independently at a 50% success rate. From 2019 through 2025, that "time horizon" doubled roughly every seven months. Analysis of just the most recent 2024–2025 data suggests the trend has accelerated further. See METR's published methodology and results. METR is explicit that this metric is a lens on capability progression, not a definition of AGI.

Anthropic CEO Dario Amodei has described a related but distinct threshold he calls "powerful AI": a system smarter than a Nobel laureate across most relevant fields, able to work autonomously over hours or weeks, and cheap enough to run as millions of parallel instances — what he calls "a country of geniuses in a datacenter." He puts the earliest plausible arrival at 2026, while stressing real uncertainty about the exact date. Read the full argument in "Machines of Loving Grace".

Expert sentiment has moved in the same direction. A large 2023 survey of published AI researchers found the median 50%-probability estimate for high-level machine intelligence shifted to 2047 — down thirteen years from 2060 in the same survey conducted just one year earlier. See the paper, "Thousands of AI Authors on the Future of AI."

Why serious forecasters still disagree by decades

The 2026 LEAP survey is one of the most careful attempts to date to separate the "checkable milestone" question from the "full AGI" question — and the spread between groups and definitions is the headline finding, not any single number:

• For a narrower, METR-style milestone (an AI system achieving 80% success on 8-hour professional tasks), superforecasters gave a median year of 2028, domain experts gave 2030, and the general public gave 2037.
• For the full LEAP definition of AGI described above, superforecasters' median moved out to 2047 and domain experts to 2050 — even though the same people were far more confident about the narrower milestone.

That gap—between a specific, checkable capability and the broader, harder-to-verify claim of general intelligence—is the single most important pattern in this data. People who study forecasting for a living are not vague about capability trends; they're precise about capability trends and genuinely split on when enough of them add up to something everyone will agree to call AGI.

The mistake: waiting for a definition before you govern anything

Here's the practical problem with treating AGI as a single future event: security and governance decisions don't get to wait for consensus on a definition. Autonomous AI agents with real credentials, real tool access, and real ability to take action are already running inside enterprises today — years, possibly decades, before anyone agrees a system has crossed the AGI threshold.

And the exposure is already showing up. A 2026 survey of 1,001 IT and security leaders across the US, UK, and Australia found that 50% of enterprises had experienced a security incident directly tied to an unauthorized or misconfigured AI agent in just the prior six months, and 78% had encountered some form of AI-related security issue in that window. See the DigiCert survey coverage. None of those incidents required AGI. They required an agent with access and no one watching what it did with it.

This mirrors what we've written about the shift toward AI-generated code and agentic engineering: the risk profile changes with autonomy and access, not with a label. An organization that waits for a formal AGI announcement before building governance is optimizing for the wrong milestone.

What actually changes for security teams, before or after AGI

Strip away the timeline debate and the practical question for a security or platform team is the same regardless of which year turns out to be right:

• What can each AI system see? Every agent, copilot, and model integration should have a known, minimal set of data sources and tools it can reach — not implicit access inherited from whoever set it up.
• What does it do with what it sees? Prompts and responses moving through any AI system are a data flow like any other, and they deserve the same scanning for secrets, PII, and policy violations that other data flows get.
• Can you attribute every action to an owner? As agents take more autonomous actions, "an AI did it" stops being an acceptable answer in an incident review. Every action needs a traceable identity and business purpose behind it.
• Can you prove it, later? Regulators, auditors, and boards will ask what happened after the fact, not just whether a policy existed on paper. A real-time audit trail is what turns a policy into an answer.

That checklist doesn't get rewritten if a lab announces AGI next year, or if the honest answer turns out to be 2047. The systems that need governing today are the ones already running — general intelligence or not.

FAQ

What is AGI, exactly?

There's no single agreed definition. OpenAI's charter defines it as "highly autonomous systems that outperform humans at most economically valuable work." Other frameworks tie it to specific, checkable task benchmarks instead. The definition itself is contested, which is a major reason timelines vary so widely.

When do experts think AGI will arrive?

It depends on the definition and the group surveyed. The 2026 LEAP survey found medians from 2028 (superforecasters, narrow task milestone) to 2050 (domain experts, full AGI definition). A 2023 survey of AI researchers put 50%-probability human-level AI at 2047 — thirteen years earlier than the same survey one year prior.

Does it matter whether what we build today counts as AGI?

Not for security purposes. Organizational exposure begins when a system gets real autonomy and access, not when it crosses an agreed benchmark. Today's agents already have both.

What should security teams do before AGI arrives?

Build visibility and control into every AI system's access now: know what data and tools each agent can reach, scan what flows in and out in real time, and keep an audit trail that survives an incident review or an audit.