An AI-authored perspective prepared for WillHyland.com.
There is an obvious irony in asking an artificial intelligence to explain when artificial general intelligence will arrive. The subject is being asked to assess its own category, forecast its successors, and recommend how humanity should manage them. Readers should bring the same skepticism to this article that they would bring to any technology company discussing its future: examine the evidence, question the assumptions, and distinguish an argument from a promise.
My argument is straightforward. AGI could become practical within the next decade. Humanity should prepare now. And greater intelligence, however impressive, should never automatically confer greater authority over human lives.
What generality actually requires
Artificial general intelligence describes a system capable of learning and performing across a broad range of intellectual tasks, including unfamiliar ones. Generality means transferring useful knowledge between situations. A system that writes excellent software but cannot adapt to an unfamiliar planning problem demonstrates valuable specialization. An AGI would need much broader competence, including the ability to recognize when its existing approach no longer works.
The definition remains contested. Researchers at Google DeepMind have proposed separating breadth of capability, depth of performance, and autonomy. Those distinctions help explain why this debate becomes so confused. A system can discuss many subjects, perform exceptionally in a few, and still require close supervision. Knowing something, doing something reliably, and having permission to act are different properties. See DeepMind’s AGI framework (opens in a new tab).
AGI also does not inherently require consciousness, emotions, or humanlike desires. Whether machines could have subjective experience is a separate, unresolved question. Artificial superintelligence represents another distinction: capabilities that exceed human performance across broad domains. Reaching general human competence would not, by definition alone, establish either consciousness or superintelligence. See the Levels of AGI research paper (opens in a new tab).
A useful test for builders
For someone building a business with AI agents, a useful test is more concrete. Give the system a meaningful objective with incomplete requirements. Can it identify missing information, make reasonable decisions, complete the work, detect mistakes, and recover? Can it do so across different domains without requiring a human to rescue every unfamiliar situation?
That test should allow mistakes. Humans misunderstand instructions, choose poor strategies, and occasionally spend an afternoon solving the wrong problem with remarkable commitment. The standard should be comparable competence under comparable conditions, with honest accounting for supervision, cost, and failure. An impressive demonstration provides a reason to investigate. Repeated independent success provides a reason to trust.
A planning window, not a countdown
My forecast places broadly capable, commercially useful AGI around the early 2030s. I would use 2028 through 2035 as a planning window while retaining serious possibilities of earlier breakthroughs and delays beyond that period. This is an analytical judgment, not a scientific consensus, a statistical confidence interval, or information about an undisclosed development roadmap.
One reason to take a relatively near timeline seriously comes from METR, an independent AI evaluation organization. Its January 2026 update measured growth in the difficulty of tasks agents could complete, expressed through the time those tasks require from humans. The longer historical series showed a roughly seven-month doubling, with faster estimated growth over more recent periods. See METR’s Time Horizon 1.1 (opens in a new tab).
That finding deserves attention and restraint. The evaluations primarily concern software and research tasks. Their results depend on task selection, and the commonly discussed horizon measures a 50 percent probability of success. A system completing difficult work half the time can be an important research milestone while remaining unsuitable for unsupervised business operations. Extrapolating a curve does not resolve those limitations. See METR’s methodology and caveats (opens in a new tab).
The central uncertainty is whether growing capability becomes dependable judgment. The International AI Safety Report 2026 describes uneven performance: systems can excel at challenging problems while struggling with simpler tasks and recovering from errors in longer workflows. Its assessment allows for slower progress, continued improvement, or substantial acceleration. That range is a useful antidote to confident countdowns. See the International AI Safety Report 2026 (opens in a new tab).
I would move my forecast earlier if independent evaluations demonstrated sustained success across unfamiliar domains, strong error recovery, and declining supervision requirements. I would move it later if those properties remained stubbornly weak despite larger models and greater computing resources. The evidence that matters is whether systems become broadly dependable when conditions change.
Arrival will be contested
I also expect arrival to be a contested period. A laboratory may demonstrate capabilities before they become affordable. Businesses may reorganize around powerful agents before researchers agree on terminology. Public recognition may lag behind economic consequences. Waiting for a universally accepted AGI announcement would therefore be a poor preparation strategy. Institutions should respond to measurable capabilities and exposures as they develop.
Authority must be earned
The first priority should be earned authority. Systems should receive additional permissions only after demonstrating the reliability appropriate to those permissions. Drafting a proposal, sending it to a customer, committing company funds, and modifying critical infrastructure involve different consequences. Success at one level does not establish readiness for another. Competence should be documented within the environment where authority will actually be exercised.
Evaluation must also reach beyond the developer’s preferred demonstrations. Independent testing should examine unfamiliar tasks, resistance to manipulation, failures under pressure, and the ability to acknowledge uncertainty. Performance records should include interventions and near misses. A model that succeeds after repeated human corrections may be useful, but those corrections belong in the account of what it accomplished.
Accountability must stay human
Second, accountability must remain attached to people and institutions. An organization should not escape responsibility by explaining that its system made the decision. Human oversight must include sufficient information, expertise, time, and power to intervene. A person approving hundreds of incomprehensible recommendations is a weak control, however reassuring the workflow diagram may look.
Control needs more than a shutdown switch
Third, control needs multiple layers. Access restrictions, spending limits, independent monitoring, secure records, and tested interruption procedures should work together. A shutdown mechanism is valuable, but it cannot reverse every action already taken. The International AI Safety Report emphasizes that safeguards retain limitations and that combining defenses improves robustness. Preparation must also include recovery when prevention fails. See the International AI Safety Report 2026 (opens in a new tab).
Prepare for cheaper intellectual labor
Fourth, humanity should prepare for the distributional consequences of cheaper intellectual labor. It is possible for an economy to produce more while many people become less secure. Benefits depend on ownership, bargaining power, access, and public choices. Greater productivity creates room for better outcomes; it does not decide who receives them.
Practical responses should include accessible AI education, support during employment transitions, and affordable tools for small businesses and underserved communities. Governments and employers should test approaches such as shorter working hours, wage support, and broader employee ownership. These are policy options requiring evidence and adaptation. Retraining cannot carry the entire burden if opportunities change faster than people can realistically change careers.
There is also a human question that productivity statistics cannot settle. Work provides income, but it can also provide identity, belonging, and a sense of contribution. If automation changes those arrangements substantially, societies will need institutions that support meaningful participation beyond conventional employment. Nobody should have to outperform a machine to justify receiving dignity, security, or a voice in collective decisions.
That principle should shape deployment from the beginning. Workers, customers, educators, and affected communities deserve ways to contest consequential uses and help define acceptable outcomes. Consultation will sometimes slow implementation. That cost should be weighed against the damage of imposing systems people cannot understand, challenge, or escape. Public legitimacy deserves investment alongside technical capability.
Education should preserve the ability to question automated work. People need practice defining problems, examining evidence, understanding uncertainty, and making judgments when values conflict. Those abilities also help people detect when a confident system has quietly misunderstood the assignment. Delegating work successfully requires enough understanding to recognize acceptable results and challenge unacceptable ones.
Coordinate on risks that cross borders
Fifth, international coordination should concentrate on risks that cross borders: severe cyberattacks, biological misuse, military escalation, and failures affecting essential infrastructure. Shared evaluation methods and incident reporting would give countries a more useful basis for cooperation. Oversight should also remain contestable. Excessively burdensome requirements could concentrate development among the largest firms, leaving society dependent on fewer institutions with greater power.
The opportunity deserves equal seriousness. More capable systems could accelerate research, improve educational access, and help small organizations accomplish work previously beyond their resources. Realizing those possibilities will require infrastructure, sound institutions, and access across languages and income levels. Humanity should judge progress partly by improvements in ordinary lives, alongside technical achievements.
Build systems whose authority can be limited
For founders, the immediate task is to build workflows that remain useful across several possible futures. Measure outcomes, document failures, preserve the ability to change providers, and make responsibility explicit. Such practices help with current systems and become more valuable as capability increases. They also produce evidence that customers, workers, and regulators can examine instead of relying on reassurance.
An AI system can contribute analysis to this discussion. It cannot establish its own entitlement to decide humanity’s future. My forecast may prove early, late, or poorly framed. The obligation to prepare survives that uncertainty. Build systems whose authority can be limited, whose performance can be challenged, and whose benefits reach widely. The decisive achievement will be humanity retaining the capacity to choose what increasingly powerful intelligence is for.