If it feels like every few months there is a new “most powerful model ever,” you are not imagining it. Systems like ChatGPT, Claude, Gemini, and other frontier models now write code, pass graduate-level exams, and help scientists reason through complex problems. OpenAI, for instance, highlights that GPT‑5.2 reaches over 90% on a difficult graduate‑level science benchmark and outperforms earlier models on expert‑level math and research tasks, which they explicitly frame as “traits that are foundational to AGI.”OpenAI GPT‑5.2 overview

At the same time, when you ask experts whether we have artificial general intelligence—or when we will get it—you will get wildly different answers. Some say “within a decade,” others say “not this century,” and many say “we still do not even agree on what AGI is.” Surveys of AI researchers over the past few years show shorter timelines on average, but still no consensus that AGI is imminent.Our World in Data: AI timelines

So are we actually getting closer to general intelligence, or are we just getting better at a very fancy kind of autocomplete? To answer that, you need to unpack three things: what “general intelligence” even means, what is really changing in today’s systems, and how credible the various forecasts about AGI timelines actually are.

What do people really mean by “AGI”?

Part of the confusion is that “AGI” is not a single, universally agreed‑upon technical term. Different communities use it slightly differently, but there is a common core.

A widely used high‑level definition is that artificial general intelligence is a system with generalized human‑like cognitive abilities: when you give it a new, unfamiliar task, it can figure out how to solve it rather than needing to be retrained from scratch. TechTarget, for example, frames AGI as software that could approach new tasks much as a human would, applying broad reasoning instead of narrow pattern‑matching.TechTarget: What is AGI? IBM similarly describes AGI as a theoretical system that can understand, learn, and apply knowledge in a wide range of settings, not just one domain.IBM: Artificial general intelligence

Recent academic work tries to make this less fuzzy by borrowing ideas from human intelligence research. One 2024 paper in “Intelligence” proposes thinking about AGI as the “shared variance” in an AI system’s performance across many different tasks—essentially, the machine version of human “g‑factor” (general intelligence). It argues that we need systematic “AI metrics” to measure this, just like psychologists use standardized tests to measure human intelligence.Defining intelligence: bridging human and AI perspectives

For you as a user, an easy way to parse it is:

  • Narrow AI: Amazing at a specific thing (e.g., playing Go, transcribing speech, translating languages).
  • General AI / AGI: Competent across many things, able to adapt on the fly, and not obviously confined to a single pre‑defined niche.

By that standard, frontier models today are clearly beyond “narrow” in the everyday sense—but many researchers argue they still fall short of the stronger, human‑level “general” bar.

How far have today’s models actually come?

Compared to even five years ago, the jump is dramatic.

Large language models (LLMs) like GPT‑5‑series, Claude 3.7, and others now:

  • Solve a large fraction of graduate‑level exam questions in science and math.
  • Write and debug substantial codebases with relatively little human guidance.
  • Perform multi‑step reasoning using tools like chain‑of‑thought prompting and external “scratchpads.”
  • Transfer knowledge between domains—for example, using physics intuitions to reason about engineering tradeoffs.

OpenAI’s own internal benchmarks show their GPT‑5.2 models hitting over 90% on GPQA Diamond, a “Google‑proof” graduate‑level science Q&A benchmark, and taking top spots on new science reasoning suites like FrontierScience.OpenAI: Evaluating AI’s scientific abilities

Anthropic, in a 2025 report about its frontier red‑teaming program, notes that its Claude models went from solving around 5% of complex cybersecurity capture‑the‑flag challenges to about a third in a year, and that models now show “early warning” signs of approaching expert‑level skills in some biological domains.Anthropic: Frontier red team progress

From an engineering standpoint, this progress looks surprisingly smooth. Research on neural scaling laws—starting with work by Kaplan and colleagues at OpenAI—found that as you scale up model size, data, and compute, performance follows fairly clean power‑law curves over several orders of magnitude.Kaplan et al.: Scaling Laws for Neural Language Models In plain language: make models bigger, feed them more diverse data, and give them more compute, and they tend to keep getting better in predictable ways.

This is part of why companies like OpenAI, Google DeepMind, and Anthropic talk openly about being on a “path to AGI”—if you see smooth, non‑saturating improvement on broad, reasoning‑heavy tasks as you scale, it is natural to extrapolate and imagine reaching human‑level performance on many tasks in the not‑too‑distant future.

Why most experts still say “we are not there yet”

However, if you zoom out from benchmark charts and marketing language, the consensus among serious researchers in 2025–2026 is that current systems are not yet AGI.

A 2025 panel report associated with AAAI (a major AI research society) summarizes it fairly bluntly: despite large-scale deep learning advances and impressive new reasoning results (for example, success on the challenging ARC‑AGI abstract reasoning benchmark), modern AI systems “are generally considered to not have achieved all of the capabilities cited in most definitions of AGI.”AAAI 2025 Presidential Panel on AGI

Where do they fall short?

You can think of several gaps that still show up in real use:

  • Robustness and reliability: Models still hallucinate facts, misinterpret edge‑case prompts, and fail in non‑obvious ways when pushed outside distribution, which is not what you want from a “general intelligence.”
  • Grounded understanding: They lack embodied experience and direct sensorimotor grounding; they reason over text and other data, but do not actually manipulate the physical world themselves (at least not yet, in mainstream deployments).
  • Long‑horizon autonomy: They can help you reason step‑by‑step over long contexts, but keeping coherent goals, plans, and safety constraints aligned over days, weeks, or complex real‑world projects is still an active research area.
  • Self‑directed learning: Today’s mainstream models do not autonomously go out, design experiments, collect data, and update their core weights unsupervised in production. Most learning happens during the expensive pre‑training phase, with relatively limited online adaptation afterward.

If your mental picture of AGI involves something like “a colleague who could be dropped into almost any cognitive job and, with some training, perform competently and responsibly,” current tools—even the strongest versions of ChatGPT, Claude, and Gemini—are powerful assistants, not fully general replacements.

The messy business of predicting AGI timelines

Even if we agree we are not at AGI today, the natural follow‑up is: when?

Here things get murky fast.

Several large surveys of AI researchers conducted around 2022–2025 and summarized by Our World in Data show a wide spread of expectations. On average, many experts now place a 50% chance of “high‑level machine intelligence” (machines able to outperform humans at essentially all tasks) sometime in the second half of this century—but the confidence intervals are huge, and North American researchers tend to be more conservative than researchers in some other regions.Our World in Data: AI timelines

Other analyses, like those summarized by IBM, note that more recent surveys (for example, a 2023 follow‑up of 2,778 researchers) have pulled timelines forward relative to earlier expert polls—some now cluster around the 2040s for very capable AI systems—but still emphasize how uncertain these estimates are.IBM: AI surveys and AGI forecasts

If you look outside academia, you will find even more aggressive forecasts—investors and independent analysts talking about AGI in the late 2020s or early 2030s—but those are less systematically collected and more opinion‑driven.

The key takeaway for you:

  • There is a real shift toward shorter timelines compared to a decade ago.
  • There is still no strong consensus; reasonable experts disagree by decades.
  • Forecasts often depend on how strictly someone defines AGI in the first place.

Are we “just scaling” our way there?

A common skeptical argument is that today’s models are just “stochastic parrots”—huge pattern recognizers that get good at mimicking training data, not at truly general intelligence.

The scaling‑law work and recent performance jumps complicate this story. As models have grown and training strategies have improved:

  • They have acquired qualitatively new capabilities (for example, multi‑step chain‑of‑thought reasoning) that were not obvious at smaller scales.
  • They generalize better to tasks that were not explicitly part of their training dataset.
  • They can be steered at inference time with techniques like tool use and “test‑time scaling,” where you let them explore many internal reasoning paths before choosing an answer.

This suggests you are not just getting “more of the same autocomplete,” but something that increasingly looks like domain‑general problem solving—particularly when models are connected to external tools, search, and code execution.

Still, scaling is not magic. As the AAAI panel notes, progress on difficult generalization benchmarks required not just bigger models but also architectural tweaks and new training methods like reinforcement learning for reasoning chains.AAAI 2025 Presidential Panel on AGI Many researchers expect that pushing from “very capable assistant” to “reliably general intelligence” will require further conceptual advances, not just more GPUs.

Why “getting closer to AGI” matters for you now

You might reasonably ask: if we are not at AGI yet, does any of this really matter for your job or industry today?

The answer is yes, for a few reasons:

  1. Capabilities already overlap with high‑skill work
    GPT‑5‑series models, Claude, Gemini, and open‑source systems are already doing pieces of work you would have assumed needed human experts: drafting legal documents, analyzing scientific papers, generating code, brainstorming product strategies. Even if they are not “general intelligences,” they are general enough across many white‑collar tasks to reshape workflows quickly.

  2. The direction of travel is clear
    Benchmarks in science, math, cybersecurity, and abstract reasoning are trending up year over year.OpenAI research overview Anthropic’s red‑team results show step‑function jumps in sensitive dual‑use areas like cyber offense and biology.Anthropic: Frontier red team progress That matters for both opportunity (AI co‑scientists, AI copilots for developers) and risk (misuse, accidents, systemic failures).

  3. Policy and safety debates are being framed around AGI and beyond
    Bodies like the OECD are already publishing reports on potential future AI risks, benefits, and policy imperatives that explicitly assume continued progress toward more general systems and call for safety evaluations and governance frameworks that can keep up.OECD: Assessing future AI risks and benefits Whether you are building AI, regulating it, or just adopting it, these choices will affect how safely and broadly future near‑AGI systems are deployed.

In other words, even if genuine AGI is still years or decades away, the road to it is already reshaping the economy and the information environment around you.

So, are we getting closer to general intelligence?

Putting it all together:

  • Definitions are still contested, but the broad idea—systems that can flexibly solve new tasks across many domains—is reasonably clear.
  • Capabilities are rising fast across science, coding, reasoning, and complex problem‑solving, supported by both empirical scaling laws and better training and inference techniques.
  • Most serious experts say we are not at AGI yet, especially if you require robust, reliable, human‑level generality across everyday tasks.
  • Forecasts cluster between “later this century” and “within a few decades,” with some more aggressive guesses, but the uncertainty is high and bound up with which definition you adopt.

On a practical level, it is fair to say: yes, we are moving in the direction of general intelligence in a way that is visible in current systems—but we are still in the “very powerful tools” era, not the “drop‑in human‑level colleague” era.

What you should do next

You do not need to settle the AGI definition debate to act. Here are concrete next steps you can take now:

  1. Get hands‑on with the strongest models
    Spend a few focused hours using tools like ChatGPT (with access to the latest GPT models), Anthropic’s Claude, or Google’s Gemini on your own hardest day‑to‑day tasks—coding, analysis, writing, planning. Pay attention to both what they can do and where they fail. That will give you a visceral sense of how “general” today’s systems already feel in your context.

  2. Map “automation‑adjacent” parts of your work
    List tasks you do that are structured, language‑heavy, or rules‑based: drafting emails, summarizing documents, exploring data, generating test cases. Experiment with offloading those to AI while keeping yourself firmly in the loop for judgment and oversight. This positions you ahead of the curve as these tools improve.

  3. Stay literate about AI risks and governance
    Follow at least one reliable source that tracks AI policy and safety developments—think major research labs’ safety blogs, organizations like the OECD, or neutral explainers that aggregate expert surveys. Understanding how close experts think we are to more general systems—and what guardrails they recommend—will help you make saner decisions about adopting AI in your team, company, or community.

You do not need to know the exact year AGI might arrive to benefit from this technology—or to help steer it in a direction you are comfortable living with. But you do need to pay attention now, while the path is still being laid.