Artificial intelligence (AI) is the engineering discipline concerned with building machines that perform tasks. Those actions are typically referred to as intelligent, such as recognizing a face, translating a sentence, proving a theorem, driving a car, or writing code. The contemporary field is largely controlled by one approach — machine learning utilizing extensive artificial neural networks fed by data. Over the past four years, this approach has yielded systems capable of answering PhD-level scientific queries, producing functional software, and engaging in conversations. It hasn't resulted in a system where errors can be predicted adequately to be disregarded.[1][2]
What the field actually does now
Almost all contemporary AI is statistical. A model is a very large function with hundreds of billions of adjustable parameters; training adjusts those parameters so that the function’s outputs match examples drawn from data. For language models the training signal is simply the next token in a sequence, which sounds trivial and turns out to induce a startling range of capabilities, from translation to arithmetic to code generation. The 2017 transformer architecture made this scale practical by letting every part of the input attend to every other part in parallel, instead of processing it step by step.[3]
Two consequences follow. The first is that capability scales with compute, data and parameters in a way that has been roughly predictable, which is why frontier training runs are now industrial projects measured in tens of thousands of specialised chips. The second is that these systems are not programmed in any conventional sense; their behaviour is induced, which is why they can be simultaneously excellent and inexplicable, and why evaluation — not just accuracy on a test set, but understanding what a model will do in a new situation — has become the central methodological problem of the field.[4]
Where capability stands
The 2026 AI Index, Stanford’s annual stocktake, reports that capability is not plateauing. Industry produced more than 90% of notable frontier models in 2025, and several now meet or exceed human baselines on PhD-level science questions, multimodal reasoning and competition mathematics. On SWE-bench Verified, which asks a model to fix real bugs in real repositories, performance rose from 60% to near 100% within a single year. Organisational adoption reached 88%, and four out of five university students now use generative AI.[5]
The geopolitical picture has changed too. The gap between the best American and Chinese models has effectively closed: DeepSeek-R1 briefly matched the leading US model in February 2025, and as of March 2026 the top US model led by only 2.7%. The United States still produces more top-tier models and higher-impact patents, but the assumption that frontier capability would remain concentrated in a handful of American labs no longer holds.[6]
The benchmarks are being solved faster than the deployment problems. A model that fixes 100% of a curated bug set is not the same artefact as a system you would let loose on your production database.
Where it fails
The gap between benchmark and deployment is best documented in the one failure mode everyone has a name for: hallucination. The numbers are worse than the marketing. A 2025 study in Communications Medicine planted a single fabricated detail — a non-existent lab test, an invented clinical sign, a made-up syndrome — in 300 physician-validated clinical vignettes and tested six models across 5,400 outputs. Every model repeated or elaborated on the false detail in between 50% and 82% of cases. A mitigation prompt cut the average from 66% to 44%, and the best-performing model from 53% to 23%; setting temperature to zero made no significant difference.[7]
Domain-specific tools fare better but not well. A preregistered evaluation of commercial legal research products in the Journal of Empirical Legal Studies found that Lexis+ AI, Westlaw’s AI-Assisted Research and Thomson Reuters’ Ask Practical Law AI hallucinated between 17% and 33% of the time, despite vendors’ claims that retrieval augmentation had eliminated the problem, with substantial differences between systems in responsiveness and accuracy. A comparison of chat tools against PubMed for medical literature searches found that the rate of fabricated citations varied enormously between products, from none at all to a large fraction of references.[8][9]
These are not arguments that AI does not work. They are arguments that the technology is reliable enough to be useful and unreliable enough to require checking, which is an unusual and awkward combination for institutions used to software that either runs or crashes.
Governance
Regulation has moved from principle to timetable. The EU AI Act entered into force in August 2024 and applies in stages: prohibitions and AI-literacy duties from February 2025, obligations for providers of general-purpose models from August 2025, and most remaining provisions plus enforcement powers from 2 August 2026, including transparency duties on chatbots, deepfakes and synthetic content. A simplification package adopted in 2026 — the Digital Omnibus on AI — deferred the high-risk rules to 2 December 2027 for stand-alone systems and 2 August 2028 for AI embedded in regulated products, while adding prohibitions on non-consensual sexual deepfakes and child sexual abuse material from December 2026. Penalties reach 3% of global turnover for model-level breaches.[10][11][12]
Elsewhere the approach is more declaratory. The AI Impact Summit held in New Delhi on 18–19 February 2026 closed with the New Delhi Declaration, endorsed by 92 countries and international organisations, alongside a set of non-binding deliverables: guiding principles for skilling, statements on AI and child safety and on AI and climate, frontier model usage commitments and a human flourishing agenda. India framed the summit around “AI for All”, with emphasis on access, equity and sovereignty rather than on binding rules.[13][14]
The open questions
Four are worth watching. Whether reliability improves fast enough for high-stakes deployment, or whether the answer is permanent human verification layered over fluent models. Whether the economics hold: training runs and inference both consume enormous compute, and the industry is still working out who pays. Whether concentration — of compute, data and talent — entrenches a handful of firms, or whether the narrowing US–China gap is the leading edge of a broader diffusion. And whether governance can keep pace with a field whose frontier moves on a timescale of months while legislation moves on one of years, which the repeated deferrals of the EU’s high-risk rules illustrate better than any white paper.[15][16]
None of these questions is settled, and confident answers in either direction — imminent transformation or imminent collapse — are usually a sign that the speaker has stopped reading the evidence. The honest position in late 2026 is that the capability is real, the adoption is real, the failure modes are real, and the institutions meant to govern both are still being built.[17]
Select a year and month above to explore cold archived discussions.
Please select an archived topic above to view the discussion logs.