10 Predictions About AGI Beyond 2026, and How to Judge Them

Predictions about artificial general intelligence tend to arrive as confident dates. This article does the opposite. It sets out ten things we think are plausible after 2026, labels every one as editorial speculation, and attaches to each a test that would show us wrong. Where we cite published forecasts or measurements, we attribute them to the people who made them and note what they do not show.

What have people actually published about AGI timing?

Four published reference points, each answering a different question. They are not directly comparable, which is the first lesson of the list.

Put together, these show a field in which careful researchers disagree, in which one measurement (METR's) is trending sharply upward on a narrow task type, and in which the most bullish public statements come from people who build the systems. For a longer treatment see our superintelligence timelines analysis and our explainer on AGI versus ASI.

How do you judge a prediction about AGI?

What are our ten predictions?

Each prediction carries the same four lines: the claim, what would make it true, what would make it false, and how to track it. All ten are editorial speculation.

1. Will people ever agree that AGI has arrived?

Editorial speculation: there will be no single, uncontested moment. Different groups will announce different thresholds at different times, and others will reject each announcement.

• True if: at least two prominent organisations each claim AGI on different definitions, and a substantial group of researchers publicly rejects both.
• False if: a widely accepted test or independent body certifies a single definition and most major labs and researchers accept the result.
• Track: announcements from labs, plus whether survey wording such as AI Impacts' 'every task better and more cheaply than human workers' becomes a shared standard.

2. Will long-task performance keep improving, but unevenly?

Editorial speculation: the length of tasks that AI systems complete reliably will keep growing in software, but progress will be bumpy and will not carry over evenly to other kinds of work.

• True if: METR's time-horizon measurements keep rising on software tasks while comparable measurements in other domains lag or are absent.
• False if: the software trend flattens clearly, or it shows up just as strongly across many unrelated task types.
• Track: METR's updated time-horizon results and the paper's own caveats about external validity.

3. Will human review, not model ability, become the bottleneck?

Editorial speculation: as agents handle longer jobs, the limiting factor shifts to checking their work. Organisations that skip that checking will meet more costly mistakes than those that build it in.

• True if: published case studies and incident reports cite missing or inadequate review more often than raw model error.
• False if: automated checking becomes reliable enough that human review is routinely dropped without a rise in failures.
• Track: postmortems from companies running agents; our guide to choosing a coding assistant lists what to test.

4. Will benchmarks matter less than real-world trials?

Editorial speculation: standard benchmarks will keep saturating or leaking into training data, and credible claims will lean more on live, hard-to-game trials in real work settings.

• True if: leading evaluators and buyers cite field trials and fresh private test sets more than public leaderboards.
• False if: a small set of public benchmarks stays trusted for years without contamination disputes.
• Track: how often new evaluations are retired or revised. Start with reading benchmark claims critically and our benchmarks page.

5. Will electricity, memory and chips set the pace?

Editorial speculation: physical supply, meaning power connections, memory and chip packaging, will limit how fast the largest systems grow at least as much as new research ideas do.

• True if: labs and data-centre builders publicly cite power or hardware supply as the reason for delays.
• False if: capacity expands on schedule with no reported shortages while capability keeps climbing.
• Track: company filings, grid-interconnection queues and chipmaker supply statements. These are verifiable and dated.

6. Will expert disagreement persist?

Editorial speculation: even with much better systems, surveys of researchers will still show a wide spread of views on timing and on whether current methods are enough.

• True if: the next large expert survey shows a wide range between its 10% and 90% points, as past surveys have.
• False if: a future survey shows tight agreement across respondents.
• Track: repeats of the AI Impacts survey and the panel-style surveys such as the AAAI panel's. Compare the wording each time before comparing the numbers.

7. Will tasks be automated before whole jobs are?

Editorial speculation: the effect on work will show up first as changed tasks inside jobs, not as whole occupations disappearing at once.

• True if: official labour statistics and employer surveys show shifting task mixes before large occupation-level declines.
• False if: entire job categories shrink sharply while task-level studies show little change beforehand.
• Track: government labour data and studies that break jobs into tasks. Note that the AI Impacts survey itself gave far later dates for full automation of labour than for task-level machine ability.

8. Will governance move toward testing before release?

Editorial speculation: more governments and large buyers will ask for documented safety and capability testing before deploying the most capable systems, even as the rules differ between regions.

• True if: new laws, procurement rules or standards add pre-deployment testing or incident-reporting duties.
• False if: such requirements are dropped or stay purely voluntary everywhere that matters.
• Track: primary legal texts and standards bodies, not summaries. Our explainer on AI alignment covers what the testing is trying to catch.

9. Will one high-profile incident reshape the debate?

Editorial speculation: a widely reported failure or misuse involving an autonomous system will change policy and public opinion more than any benchmark result does.

• True if: a specific incident is followed by named policy or company changes that cite it.
• False if: serious incidents occur without producing visible changes in rules or practice.
• Track: incident databases and regulator statements. Our overview of superintelligence risks sets out the categories.

10. Will AI-assisted discoveries be real but contested?

Editorial speculation: AI will contribute to research results that hold up, and there will also be disputes over how much credit the system deserves and whether results replicate.

• True if: peer-reviewed results acknowledge significant AI contribution and at least some are challenged on credit or reproducibility.
• False if: AI-credited results are either absent from the peer-reviewed record or accepted without dispute.
• Track: published papers and independent replications, not announcements. This is where the line between Documented Research and press release matters most.

How can you keep score on these predictions?

Write down the date you read this. When you return, take each prediction and mark it supported, contradicted or too early to say, citing the observation that decided it. Public forecasting platforms such as Metaculus publish live community forecasts you can check against; their figures change often and secondary reports of them conflict, so use the live question pages, not summaries.

If you want the definitions in plain terms before you start scoring, begin with our AGI versus ASI explainer and our piece on the intelligence explosion.

What are the limitations of this article?

• The predictions are our judgement. We have not built a model or tested them, and we may be wrong about all ten.
• We read each primary source on 18 October 2026. We did not open the AAAI panel's full report, so we cite only what its web page states.
• We cite a lab leader's 2024 view because it is widely quoted, not because it is more reliable than the researchers' surveys. Its stated window of 'as early as 2026' is now closing, and we do not assess here whether it was met.
• Search and aggregator pages report differing Metaculus figures, so we give none.