What is AI? A clear, evidence-based answer

Almost every company now says it uses AI. Your phone, your CRM, your bank and your kid's homework helper all claim it. When a word describes everything, it starts to describe nothing, and that makes it hard to judge what you're actually buying, building or worrying about.

This article gives you a working answer. Not a slogan, but a definition you can apply, a short map of how the pieces fit together, and a snapshot of what the best evidence says about where the technology stands in 2026. It's the foundation for everything else this series will cover.

The short answer

The most widely used definition today comes from the OECD, updated in late 2023 by a group that included Stuart Russell, co-author of the field's standard textbook. It says an AI system is "a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as content, recommendations, or decisions that can influence physical or virtual environments"¹. The European Union's AI Act adopted nearly the same wording as its legal definition, adding that such systems operate "with varying levels of autonomy" and "may exhibit adaptiveness after deployment"².

That sentence is dense, so here it is in plain terms. An AI system:

• Works toward a goal. Sometimes a person spells the goal out ("flag fraudulent transactions"). Sometimes the goal is implicit, learned from examples or feedback rather than written down.
• Infers rather than follows a script. This is the key word. A traditional program executes rules a developer wrote. An AI system works out how to turn an input into an output, usually from patterns it learned in data.
• Produces something useful. A prediction, a recommendation, a decision, or (newer) content like text, images and code. The word "content" was added in the 2023 update specifically to cover generative AI³.
• Affects the world. Its outputs change something, whether a web page, a loan decision or a robot arm.

The practical test is inference. If a developer could list every rule the system follows, it's ordinary software, however clever. If the system learned its own way of mapping inputs to outputs, it's AI.

Where the idea came from

The question is older than computers most people have used. In 1950 the mathematician Alan Turing asked "Can machines think?" and proposed replacing that unanswerable question with a practical one: can a machine hold a conversation well enough that a person can't tell it apart from a human⁴? Five years later, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon wrote the funding proposal for a 1956 summer workshop at Dartmouth. It coined the term "artificial intelligence" and rested on the bet that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it"⁵.

For the next few decades, most AI work tried to do exactly that: write the rules of intelligence down by hand. These "symbolic" or "expert" systems worked in narrow settings and broke outside them, because the real world has more exceptions than anyone can write rules for.

The approach that eventually won was different. Instead of telling machines the rules, researchers gave them examples and let them find the patterns. Richard Sutton, a pioneer of reinforcement learning, summarized seventy years of this history as "the bitter lesson": general methods that learn from data and scale with more computing power have beaten hand-crafted human knowledge again and again⁶. That lesson explains most of what has happened since.

Four layers people often confuse

"AI," "machine learning," "deep learning" and "generative AI" are used interchangeably, but they are nested, like Russian dolls.

• Artificial intelligence is the whole field and the goal: machines doing tasks that would require intelligence if a person did them. It includes old rule-based systems as well as today's models⁷.
• Machine learning is the subset of AI where systems learn from data instead of being programmed rule by rule. Your email spam filter and a bank's fraud model are classic examples.
• Deep learning is the subset of machine learning that uses large, layered neural networks. Its breakthrough decade was the 2010s, summarized by Yann LeCun, Yoshua Bengio (based in Montreal) and Geoffrey Hinton (based in Toronto) in a landmark 2015 review⁸. It's what made computers good at recognizing images and speech.
• Generative AI is the subset of deep learning that creates new content. Most of today's systems, including ChatGPT, Claude and Gemini, are large language models built on the "transformer" architecture introduced by Google researchers in 2017⁹.

When a headline says "AI," it now usually means the innermost doll. But most AI actually running inside businesses (forecasting, fraud detection, recommendations) is still the plainer machine learning in the second doll. Knowing which one someone means is the first step to evaluating their claim.

How today's AI actually works, in one paragraph

A large language model is trained on a vast amount of text to do one thing extremely well: predict what comes next. Given "The capital of France is," it learns that "Paris" is likely. Do that across trillions of words and the model picks up grammar, facts, reasoning patterns and styles of writing, because all of those help it predict better. After this "pretraining," developers fine-tune the model with human feedback and other techniques so it follows instructions, answers helpfully and declines harmful requests. Newer "reasoning" models are also trained to work through a problem step by step before answering. Later articles in this series take each of those steps apart.

What powers the progress is scale. Epoch AI, an independent research group that tracks the field, estimates that the computing power used to train frontier language models has grown about 5x per year since 2020, while the algorithms have become roughly 3x more efficient per year, meaning the same performance needs a third of the compute a year later¹⁰. Two exponential curves stacked on each other is why capabilities have moved so fast.

What AI can do in 2026: the evidence

A few well-sourced measures give a fair picture.

Capability is rising quickly on hard tests. Stanford's 2026 AI Index, the most comprehensive annual review of the field, reports that top models now score around 50% on Humanity's Last Exam, a benchmark of expert-level questions designed to be too hard for AI, up from single digits a year earlier¹¹.

AI can now handle longer tasks on its own. METR, a nonprofit that evaluates AI systems, measures the length of software tasks (in human working time) that models can complete with 50% reliability. Across 2019 to 2025, that "time horizon" doubled roughly every seven months, and faster since 2024. In its January 2026 update, the best model measured could complete tasks that take a skilled human around five hours, though METR stresses the confidence intervals are wide¹².

Adoption is broad. The AI Index reports that generative AI reached about 53% of the global population within three years of ChatGPT's launch, faster than the personal computer or the internet¹³. In McKinsey's 2026 global survey, nearly nine in ten respondents say their organization regularly uses AI in at least one business function¹⁴. Canada is further behind but moving fast: Statistics Canada found 19.2% of Canadian businesses used AI to produce goods or deliver services in the year to Q2 2026, triple the 6.1% of two years earlier¹⁵.

Money is pouring in. Global corporate investment in AI passed US$581 billion in 2025, more than double the year before, and more than 90% of notable models now come from industry rather than universities¹¹.

What AI is not (yet)

The same evidence also shows clear limits, and they matter as much as the headline numbers.

It is "jagged." AI capability doesn't rise evenly across tasks. In a field experiment with 758 Boston Consulting Group consultants, published this year in Organization Science, those using AI on tasks inside its capabilities completed 12.2% more tasks, 25.1% faster, with markedly higher quality. On a task just outside that frontier, consultants using AI were 19% less likely to reach the correct answer than those working without it¹⁶. The same pattern shows up in the 2026 AI Index, where models that ace graduate-level science questions still struggle to read an analog clock reliably¹¹. The problem is that the frontier is invisible: nothing warns you when you've crossed it.

It can be confidently wrong. Language models sometimes produce fluent, plausible statements that are false, known as "hallucinations." A 2025 analysis by OpenAI researchers argues this isn't a mysterious glitch: standard training and evaluation reward a confident guess over an honest "I don't know," much like a multiple-choice exam with no penalty for guessing¹⁷. That makes hallucination reducible but not something to assume away.

Adoption is not the same as impact. Nine in ten organizations use AI, yet only about 6% of McKinsey's respondents qualify as "high performers" who attribute more than 5% of their earnings to it¹⁴. Most of the value still depends on redesigning how work gets done, not just switching a tool on.

It is not a mind in the human sense. The OECD definition deliberately describes what AI systems do, not what they are. Whether any AI system understands, intends or experiences anything is a live debate this series will cover separately. For practical decisions, it's safer to describe systems by their measured behavior than by human metaphors.

What this means for you

A good definition is only useful if it helps you decide something. Here are four questions to ask whenever a product, pitch or headline says "AI":

1. Does it actually infer, or is it following rules? Ask what the system learned from and what it produces. Plenty of "AI-powered" products are well-designed rules engines, which may be fine, but should be priced and trusted accordingly.
2. Which layer is it? A classic machine-learning model, a generative model, or a mix? They fail differently. The first tends to degrade quietly when data changes. The second can produce convincing errors.
3. Is your task inside the frontier? Test it on your own work, including the hardest 10% of cases, before relying on it. The jagged frontier means a great demo tells you little about the edge cases.
4. How will you know it's working? Decide the measure (time saved, error rate, revenue) before you start. The gap between adoption and impact is mostly a gap in measurement and workflow design.

Sources

1. Russell, S., Perset, K., & Grobelnik, M. (2023, November 29). Updates to the OECD's definition of an AI system explained. OECD.AI. oecd.ai/en/wonk/ai-system-definition-update
2. European Union. (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 3(1). artificialintelligenceact.eu/article/3
3. OECD. (2024). Explanatory memorandum on the updated OECD definition of an AI system. OECD Artificial Intelligence Papers, No. 8. oecd.org/en/publications/explanatory-memorandum-on-the-updated-oecd-definition-of-an-ai-system_623da898-en.html
4. Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460. doi.org/10.1093/mind/LIX.236.433
5. McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (1955/2006). A proposal for the Dartmouth summer research project on artificial intelligence. Reprinted in AI Magazine, 27(4), 12. doi.org/10.1609/aimag.v27i4.1904
6. Sutton, R. (2019, March 13). The bitter lesson. incompleteideas.net/IncIdeas/BitterLesson.html
7. Russell, S., & Norvig, P. (2021). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.
8. LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521, 436–444. doi.org/10.1038/nature14539
9. Vaswani, A., et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems 30. arxiv.org/abs/1706.03762
10. Epoch AI. (2026). Trends in artificial intelligence (accessed September 24, 2026). epoch.ai/trends
11. Stanford Institute for Human-Centered AI. (2026). The 2026 AI Index Report. Figures as reported in IEEE Spectrum, "Stanford's AI Index for 2026 Shows the State of AI" (spectrum.ieee.org/state-of-ai-index-2026). hai.stanford.edu/ai-index/2026-ai-index-report
12. METR. (2026, January 29). Time Horizon 1.1. metr.org/blog/2026-1-29-time-horizon-1-1
13. Stanford HAI (2026), as reported in Search Engine Journal, "AI adoption outpaced the PC & internet." searchenginejournal.com/ai-adoption-outpaced-the-pc-internet-dive-into-the-stanford-report-data/572305
14. McKinsey & Company. (2026, August 25). The state of AI in 2026: On the road to ROI. mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
15. Statistics Canada. (2026, June 11). Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026. www150.statcan.gc.ca/n1/pub/11-621-m/11-621-m2026010-eng.htm
16. Dell'Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Organization Science (Articles in Advance, March 11, 2026). hbs.edu/faculty/Pages/item.aspx?num=64700
17. Kalai, A. T., Nachum, O., Vempala, S. S., & Zhang, E. (2025). Why language models hallucinate. arXiv:2509.04664. arxiv.org/abs/2509.04664