Ask five people what “general AI” means and you’ll get at least three answers. A machine learning researcher will assume you mean artificial general intelligence, the hypothetical system that matches human performance across the board. A product manager at a software company probably means the general-purpose models sitting behind ChatGPT, Gemini and Claude. Someone drafting compliance guidance in Brussels means something narrower still: a legal category defined largely by how much compute went into training a model.
All three usages are legitimate. They are also not interchangeable, and a surprising number of bad predictions come from sliding between them mid-sentence.
General-purpose AI is not the same as general intelligence
The confusion is baked into the word itself. “General” can describe what a system can be pointed at, or how capable it is at everything. Those are very different claims, and only one of them is currently true.
The general-purpose models you already use
A general-purpose model is trained on broad data and shipped without a fixed task. You can ask it to draft a lease, debug Python, summarise a deposition, translate a Portuguese contract, or explain the Krebs cycle, all in the same chat window. That flexibility is real. ChatGPT passed 800 million weekly users in 2025, and most of them are using one model for a dozen unrelated jobs rather than one tool per task.
Contrast that with the narrow AI quietly running most of the world: spam filters, fraud scoring, route optimisation, and protein structure prediction tools like AlphaFold. These systems can be extraordinary at one thing and useless at anything else. AlphaFold earned a Nobel Prize and still cannot write you an email.
Artificial general intelligence, the version nobody has built
AGI is a different claim entirely. OpenAI’s charter describes it as “highly autonomous systems that outperform humans at most economically valuable work.” No system shipping in 2026 meets that bar, and the benchmarks designed to measure the gap keep being instructive. Models that ace the bar exam and score well on graduate-level physics still stumble on puzzle tasks a ten-year-old solves in a minute, whenever the puzzle requires genuine novelty rather than pattern recall.
Why the distinction is not just semantic
In 2024, the EU AI Act turned “general-purpose AI” into a legal term of art. Models trained above roughly 10^25 floating-point operations are presumed to carry systemic risk and face extra obligations: publishing training data summaries, documenting evaluation results, respecting copyright opt-outs. That threshold arrived before anyone had agreed on a technical definition of generality. Regulation moved first.
Marketing moved faster. Nearly every enterprise pitch deck now claims to run on a “general AI platform,” which usually means a wrapper around someone else’s API. Worth asking what the model was trained on, and who is liable when it gets something wrong.
What general-purpose AI actually does well
- Drafting and rewriting. First drafts of almost anything, tone shifts, compressing a 60-page document into something you can read on a train.
- Code. Scaffolding, unit tests, porting between languages, explaining a repo you inherited from someone who left.
- Translation and transcription. Near-human quality for major languages, and good enough to be useful for many others.
- Structured extraction. Turning messy emails, invoices and support tickets into clean JSON you can query.
- Divergent thinking. Twenty angles on a problem in thirty seconds. You still do the filtering.
Notice what’s missing from that list: anything requiring a long chain of decisions where one error compounds. That is the boundary worth internalising.
Where it still falls over
Hallucination remains the headline problem. A model that is right 95% of the time sounds impressive until you remember it has no reliable way to tell you which 5% is wrong. In high-stakes settings, that ambiguity is often worse than an outright failure.
Long-horizon work is the second wall. Multi-step agents that book, file, purchase or negotiate on your behalf fail at rates that make them hard to trust with anything expensive. Give a model a task with a dozen dependent steps and a 90% per-step success rate, and you’re looking at roughly 28% end-to-end reliability. Compounding is brutal.
Memory across sessions is the third. Models still forget you, your preferences and your previous corrections. Every new conversation starts cold unless something external is holding the context.
Three competing bets on the road to AGI
Throw more silicon at it
One camp insists scale is still the whole game. That has driven a hardware arms race, and some of the most interesting work is happening at the chip level, including Cerebras and its wafer-scale approach to inference speed. If reasoning gets cheaper per token, a lot of applications that currently don’t pencil out suddenly do.
Change the architecture and the training loop
Another camp argues the transformer plus next-token prediction has a ceiling, and the interesting progress will come from new objectives, better reinforcement learning, or something closer to how scientific discovery actually works. DeepMind’s track record across AlphaFold, Gemini and a Nobel Prize in Chemistry is the strongest evidence that research taste still beats raw compute in some domains.
Ship capability first, argue about safety later
The third bet is the least subtle. Build fast, release fast, and let regulators catch up. The ongoing feud between xAI and OpenAI is as much a fight about which release philosophy wins as it is about personalities. Whoever sets the default tempo for the industry shapes the rules everyone else has to work inside.
Getting value from general AI without swallowing the hype
The practical move is to treat general-purpose models as extremely capable interns with no memory and occasional confident delusions. Three habits cover most of the gap:
- Verify anything that carries a cost. Numbers, legal claims, medical details, citations. Ask for sources and check them.
- Narrow the task. A prompt that defines the audience, format and constraints beats a clever one every time. There’s a real skill to using these tools well, and it is mostly about specificity and iteration.
- Decide the accountability question before deployment. Building AI systems you can actually defend means knowing who owns an error, how it gets detected, and what happens to the person affected.
Teams that get this right tend to use general-purpose models for the 70% of work that is generation, summarisation and extraction, and keep humans on the 30% that involves judgement, liability or genuine novelty.
What’s actually worth watching
Ignore the AGI timelines posted on social media; they’re mostly marketing with a straight face. Watch three concrete things instead.
Agent reliability. The number that matters is the success rate on long, messy, multi-tool tasks, not benchmark scores on clean ones. When that figure climbs past 95% in domains with real consequences, the economics of white-collar work start to shift in ways that are hard to reverse.
Cost per unit of reasoning. Every order-of-magnitude drop has unlocked a wave of applications that were previously too expensive to run. Track it and you can usually predict what gets built next.
What regulators actually enforce. The general-purpose AI obligations in Europe are the first serious test of whether rules written before the technology matured can survive contact with it. If enforcement turns out to be a paper exercise, expect compliance documents to keep getting longer while the models stay much the same. If it bites, training data disclosure becomes a competitive issue rather than a legal footnote.