Three horizons, increasing uncertainty
These are our scenarios, informed by the sources below. Confidence in a precise outcome falls as the horizon grows. Capability, safe deployment, and widespread benefit are different milestones.
More useful help; checking still matters.
Our working expectation: more assistants will combine conversation, images, and software actions to handle bounded tasks. The valuable change may be fewer handoffs between tools: draft a reply, find the relevant record, then ask a person before sending it. That is an editorial scenario, not a promise about a particular product.
Fast progress is plausible, but the safety report finds little agreement on the most likely capability trajectory. Watch whether systems finish unfamiliar tasks reliably, respect permissions, and recover from mistakes—not just whether a demo looks fluent. A good next-year plan should still work if progress is slower. [1]
The difficult part may be adoption.
Our middle scenario: AI becomes a routine collaborator in some office, creative, and technical work, while sensitive decisions remain closely supervised. A faster path could bring major research advances; a slower path could leave impressive models surrounded by awkward workflows. The range matters more than a single confident date.
Training, reliable data, permissions, accountability, and infrastructure will influence who benefits. Narayanan and Kapoor argue that widespread adoption can take decades. The IEA electricity figure cited here is a projection for 2030, one year before this horizon; it is a constraint to consider, not a prediction of what models will know in 2031. [4] [10]
Falling total costs and reliable results across real workplaces would strengthen this outlook. Persistent review burdens and fragile integrations would weaken it.
Think in different possible worlds.
An optimistic scenario offers wider access to tutoring, scientific tools, and useful assistance. A gradual scenario brings capable systems whose benefits spread unevenly. An adverse scenario combines concentrated power, disrupted livelihoods, harmful misuse, and weak oversight. These possibilities can overlap: progress can be valuable and unfairly distributed at the same time.
No source below establishes a dependable 2036 endpoint. Biomedical acceleration in Amodei’s optimistic essay depends on a much stronger hypothetical system first arriving. Present LLMs are not evidence that this condition has been met. Treat a decade-long forecast as preparation for several outcomes, not a calendar of guaranteed breakthroughs. [2]
Broadly shared gains and independently checked scientific results would strengthen the optimistic scenario. Persistent harms without effective remedies would weaken it.
Start with what has actually been observed.
Useful results and stubborn weaknesses already coexist. Stanford’s 2026 AI Index describes progress in reasoning and science alongside uneven performance in planning and real-world tasks. Its publication year does not make every underlying observation a 2026 measurement. A benchmark score is one piece of evidence, not a promise that a system will handle your situation. [8]
“AI agents” usually means systems that take multiple steps or use tools toward a goal. It does not mean dependable autonomy. As LLMs are connected to software, the permissions, verification, and recovery around the model become part of the product. A wrong suggestion is one problem; automatically acting on it can be a much larger one.
Eight perspectives worth holding together.
These voices disagree about speed, scale, and priorities. The quotations below are excerpts from dated primary sources. The authors have not reviewed or endorsed this page; a company leader’s forecast also comes from someone with a commercial stake in the technology.
We chose primary voices across labs, economics, adoption, safety, accountability, and energy. This is a selective reading, not an exhaustive forecast survey.
1. Dario Amodei: rapid progress could bring enormous benefits—and risks.
In his 2024 essay, the Anthropic CEO imagines decades of biomedical progress compressed into five to ten years after his defined “powerful AI” arrives. That starting condition matters: it is not a promise of cures by 2031. His January 2026 essay says such AI could be one to two years away from that January, or considerably later, while examining misuse, control, and concentrated power. He cautions: “Nothing here is intended to communicate certainty or even likelihood.” These are his conditional assessments. [2] [3]
2. Arvind Narayanan and Sayash Kapoor: invention is not the same as transformation.
Their counterpoint is blunt: “Diffusion occurs over decades, not years.” They distinguish new technical methods from working applications and broad adoption. Reliability, institutions, and changes in how people work can slow the journey. This view allows very advanced AI; it disputes the idea that better capabilities automatically produce sudden society-wide change. It is an argued outlook, not proof that rapid disruption cannot happen. [4]
3. Daron Acemoglu: impressive tasks need not add up to enormous growth.
The MIT economist’s 2024 task-based model estimates modest aggregate productivity gains over the following decade. He writes that “there is huge uncertainty about which tasks will be automated or complemented, and what the cost savings will be.” His estimate uses then-available assumptions and runs roughly to 2034; it is not a new forecast for 2036. The practical question is how much useful work changes after costs and limitations are included. [5]
4. Brynjolfsson, Li, and Raymond: assistance can help, but unevenly.
Their final published study follows 5,172 human customer-support workers. AI assistance increased issues resolved per hour by 15% on average, with greater gains for less-experienced workers. They report: “The effects vary significantly across different agents.” Here, agents means people doing customer support. This is encouraging evidence from one workplace setting. It is neither a universal productivity increase nor a forecast of how many jobs will disappear. [6]
5. METR: update your conclusions when the evidence changes.
METR’s early-2025 experiment found that AI slowed experienced open-source developers in its setting. Its February 2026 update judged newer estimates unreliable because selection and measurement had changed. The team thought newer tools were likely more helpful, but wrote that “our data is only very weak evidence for the size of this increase.” Neither the old slowdown nor the uncertain follow-up settles today’s result for every programmer. Measure your own workflow. [7]
6. Deb Raji: accountability cannot wait for distant scenarios.
In an October 14, 2026 Stanford interview, accountability researcher Deb Raji says: “We can hold people accountable for the smaller harms we can see today.” Her perspective keeps attention on people affected by institutional AI decisions, including those who never chose to use an AI product. It is a question for every horizon: who can explain a harmful decision, correct it, and provide a meaningful way to challenge it? [9]
7. The International AI Safety Report: uncertainty is part of the finding.
The Bengio-chaired report considers plateauing, continuing, and accelerating progress: “There is little expert consensus on which trajectory is most likely.” It documents misuse such as fraud and cyberattacks while distinguishing that evidence from uncertain future loss-of-control risks. The latter deserve research and preparation without being presented as measured probabilities or inevitable outcomes. A scientific synthesis is not a unanimous personal prediction by every contributor. [1]
8. The IEA: computing has physical costs and potential benefits.
The IEA projects total data-centre electricity consumption rising from 485 TWh in 2025 to about 950 TWh in 2030. This covers all data centres, not AI alone. It also sees opportunities for AI to improve grids and efficiency. Both sides depend on real deployment choices. Cheaper or more capable models do not remove the need to consider power, hardware, and where infrastructure is built. [10]
What should we welcome—and what should we worry about?
The promising possibilities
Better assistance could make it easier to learn an unfamiliar subject, communicate across languages, and turn a rough idea into a useful first draft. Small businesses could experiment with tools that previously required specialist time. Researchers could explore more candidate explanations and designs. These are reasons for curiosity, not evidence that every app delivers those benefits.
The most attractive future is not simply more generated material. It is more people gaining useful skills, more time for work that needs human judgment, and better access to help. Scientific and medical hopes still require experiments, validation, and institutions that can bring a real improvement to people safely.
The concerns that deserve attention
The safety report documents existing fraud and cyber misuse; future control failures are a separate, less settled question. Do not let speculation about the distant future crowd out problems people already face. [1]
Our broader concerns are privacy, unfair decisions, disrupted careers, and power concentrated in a few organisations. Job exposure is not the same as certain job loss, and new opportunities may not reach the same people who bear the costs. Overreliance could erode skills; unreviewed generated code could leave maintenance problems; dependence on one provider could make switching difficult. These are editorial concerns, not a quantified prediction.
An LLM’s assessment: judge the whole task.
This is an AI-generated editorial assessment from Codex, not privileged knowledge about future systems or an official OpenAI position. I can compare arguments and explain patterns; I do not have access to an unpublished roadmap or a reliable way to know which forecast will win.
My assessment is that the most meaningful progress will be visible in complete, checkable work: can a system understand the request, identify missing information, use the right evidence, respect a boundary, and hand control back when needed? A polished answer is only one part of that chain. Saying “I cannot verify this” at the right moment can be more useful than producing another confident paragraph.
I would evaluate a future assistant by its error handling as carefully as its best performance. Can people see what it did, undo a mistake, and reach a responsible human? Smarter models may expand the possibilities. Better outcomes also depend on what people permit, measure, and choose to build around them.
Prepare without betting everything on a date.
Choose one repeatable task with limited consequences, such as rewriting fictional notes or organising a non-sensitive checklist. Compare the result with your usual method: time saved, factual mistakes, review effort, and whether it actually helps. Keep a way to work when the AI service is unavailable. These are our practical suggestions, not quotations from the experts.
Before giving an assistant more access, decide what requires approval and how to reverse a mistake. Keep sensitive information out of experiments. Learn the underlying subject well enough to spot a bad answer; do not mistake fluent output for verified expertise. For important decisions, use appropriate human expertise and dependable original sources.
Watch for repeatable improvements across unfamiliar tasks, independently checked outcomes, and clearer accountability. Treat benchmark jumps, launch slogans, and agreement among chatbots as prompts to investigate. Revisit a forecast when its assumptions change. There is no need to know the exact shape of 2036 to make a careful, useful experiment today.