The Deeper Problem With the Foundation Models from BigTech and Everyone Else
Even the languages that are well-represented face serious issues with foundation models as they stand today.
Hallucination is the politest word in AI for “making things up with total confidence.” Every major foundation model does it. Claude does it less than most; others do it more. In high-stakes domains — medicine, law, financial advice, public administration — a model that fabricates citations or misremembers facts is not just unhelpful. It is dangerous.
Bias is structural, not incidental. If the training data reflects historical prejudices, the model will too. A model trained mostly on English-language Western internet content will have a particular worldview — and that worldview will be delivered with the authority of a trillion-parameter certainty.
And then, there is a myth about how a person using a foundation model has access to real-time information.
The Knowledge Cutoff Problem
Every foundation model is trained on a snapshot of the world — a dataset frozen at a specific point in time, known as the knowledge cutoff. After that date, the model has no awareness of what has happened: no new scientific discoveries, no geopolitical shifts, no updated regulations. But the challenge runs deeper than the cutoff date itself. There is an often-overlooked gap between when training data is collected, when the model is actually trained, and when it is finally shipped to users. This pipeline can take anywhere from several months to well over a year. By the time you interact with a model, the knowledge baked into it may already be 12 to 24 months old — a significant lag in a field that moves as fast as AI.
Training-to-Shipping Gaps: Real Numbers
The data from leading labs makes this gap strikingly concrete. OpenAI’s GPT-5.2, for example, has a knowledge cutoff of August 2025 but was released in December 2025 — a gap of roughly four months. GPT-5.5 Instant carries an October 2025 cutoff yet shipped in May 2026, meaning users interacting with it today are working with knowledge already seven months old. Anthropic follows a similar pattern: Claude Opus 4.6 and Claude Sonnet 4.6 both carry a training cutoff of August 2025 and were released in February 2026 — a six-month gap. Claude Opus 4.8, their most current model, was trained on data up until January 2026, and even that represents months of lag by the time it reaches everyday users. Notably, Anthropic distinguishes between a “training data cutoff” and a “reliable knowledge cutoff” — acknowledging that even within the training window, knowledge of very recent events is thinner and less reliable.
The Real-Time Data Myth
One of the most persistent and quietly dangerous misconceptions in AI literacy is the belief that working with an LLM means working with live, real-time information. It does not. Unless a model is explicitly augmented with retrieval tools, live search, or external APIs, every response it generates is drawn entirely from its frozen training snapshot. This matters enormously in practice: a leader making strategic decisions, a researcher assessing the competitive landscape, or a professional reviewing regulatory requirements may receive answers that are confidently worded yet factually outdated. Confidence in tone is not a signal of currency. Treating an LLM as a real-time oracle is not just a technical misunderstanding — it is a risk management failure, and closing that literacy gap is precisely why AI education at every organisational level is no longer optional.
Foundation models are genuinely extraordinary things. The speed at which they have advanced is almost without precedent in the history of technology. But they are not neutral tools descending from some objective scientific heaven. They are built by specific people, in specific countries, with specific data, on specific infrastructure that consumes specific amounts of water and electricity. Understanding what is inside them — and what is not — is not an optional technical detail for specialists. It is a literacy requirement for anyone who wants to make good decisions in the world that is being built around them.