The Observatory
Autocomplete with ambitions: artificial intelligence in plain English
How today’s AI is trained, where it shines, where it slips, and the questions worth asking.

Photo Credit
Image courtesy of © GoogleLarge language models write, summarise and translate in seconds, and still make things up with a straight face. How they learn, why they hallucinate, how Europe and Britain are writing the rules, and the questions to ask before you trust an answer, checked against the makers’, regulators’ and universities’ own pages in November 2026.
From a parlour game to autocomplete
In October 1950 the philosophy journal Mind published a paper by Alan Turing that opened with a question still argued over today: “Can machines think?” Within a few pages he had dismissed it as “too meaningless to deserve discussion” and swapped it for a game. The imitation game, as he first set it out, had nothing to do with computers: an interrogator, passing typed questions, tries to tell a man from a woman hidden in another room.
Five years later, on 31 August 1955, John McCarthy of Dartmouth College, Marvin Minsky of Harvard, Nathaniel Rochester of IBM and Claude Shannon of Bell Telephone Laboratories proposed “a 2 month, 10 man study” of artificial intelligence for the summer of 1956. Their problems included how a computer might be programmed to use a language. They expected “a significant advance” within the summer. The summer has so far lasted seven decades.
Strip away the chat window and a large language model is a prediction engine. Google’s own developer course calls LLMs “essentially autocomplete mechanisms”: given some text, the model estimates which token, a word or a fragment of one, comes next, then the next, thousands of times over. The engine underneath, the transformer, weighs how much every other word in a passage should colour its reading of each one. On its own, it looks nothing up.
How it learns, and where it slips
Stage one is pre-training. The model reads an enormous body of text and, by predicting the next token over and over, absorbs the grammar and idiom of a language, along with much of what that text says about the world. Nobody hands it the rules of grammar; it works them out from the examples, one guess at a time. Google’s course notes that training a large model can take “multiple months and enormous computational resources and electricity”.
Stage two is fine-tuning on examples of the job wanted; stage three brings in people. In reinforcement learning from human feedback, labellers write model answers and rank the system’s replies. In a 2022 paper, OpenAI researchers reported that people preferred a 1.3-billion-parameter model trained this way to the 175-billion-parameter GPT-3, despite its having 100 times fewer parameters. Good manners, it turns out, can beat sheer size.
Anthropic, which makes Claude, adds a variant it calls Constitutional AI: the model is trained to critique and revise its own answers against a written set of principles, and a reinforcement-learning stage then uses AI-generated feedback based on those principles. The result, from every maker, is a remarkable generalist that drafts, summarises, translates and explains in seconds, at its best where the reader can check the work.

Photo Credit
Image courtesy of © NASAThe weakness is the confident mistake, which the industry calls hallucination. Anthropic’s own help pages warn that Claude can produce responses that are incorrect or misleading, including quotes that look authoritative but are not grounded in fact, and that its data may be out of date. Google’s course is blunter still: “LLMs hallucinate.” Fluency has a price: a system built to write plausible sentences writes plausible wrong ones too.
Treat a chatbot like a brilliant, well-read intern who has never once said “I don’t know”.
A September 2025 paper by OpenAI and Georgia Tech researchers argues that training and testing reward guessing over admitting uncertainty: on a multiple-choice exam with no penalty for wrong answers, the bluffer beats the student who says “I don’t know”. Asked one author’s birthday three times, a model gave three different wrong dates, all with poise. Treat a chatbot like a brilliant, well-read intern who has never once said “I don’t know”.
Bias is the second fault line. Google’s course states plainly that, like all machine-learning models, LLMs “can exhibit all sorts of bias”, inherited from the data they learned from; it matters most where a system helps decide something about a person, such as a CV sift or a loan. Cost is the third: the computing power and electricity behind these systems are enormous, every time a question is answered as well as in training.
Who is writing the rules
The European Union’s AI Act entered into force on 1 August 2024 and sorts AI by risk. Social scoring has been banned since 2 February 2025; rules for general-purpose models, the kind behind chatbots, have applied since 2 August 2025; and since August 2026 a chatbot must tell people they are dealing with a machine. After a 2026 amendment, high-risk rules, for uses such as CV sorting and credit scoring, apply from 2 December 2027.
Britain set up an AI Safety Institute in November 2023 to evaluate advanced AI systems, as a research body, not a regulator. On 14 February 2025 it was renamed the AI Security Institute, to focus on serious risks with security implications, from cyber-attacks and fraud to AI misused to help build chemical or biological weapons. It describes its mission as equipping governments with a scientific understanding of the risks posed by advanced AI.

Photo Credit
Image courtesy of © GoogleFour questions worth asking
Our advice: before you trust an answer, ask four questions. Where did this come from? Ask for sources, then open them. How current is it? Ask whether the model searched the web. What happens to what you type? Read the privacy settings before pasting in a contract or a medical letter. Would you act on it without a second opinion? For health, law, money and safety, Anthropic’s own help pages advise checking high-stakes advice carefully.
Closer to home: Brighton
Sussex AI, the University of Sussex’s Centre of Excellence for artificial intelligence, is directed by Professor Thomas Nowotny and Professor Julie Weeds. Its AI Research Group was formed by merging the university’s Evolutionary and Adaptive Systems group with its Data Science group, which explains a research list that runs from robotics to the workings of biological neurons.
On Tuesday 24 November 2026 the university and Brighton Dome stage “AI: hope, hesitation or hype?”, an evening of short talks, a panel discussion and audience questions in the Dome’s Studio Theatre, 6.30–8.30pm, with Sussex researchers from computer science, business, education and the arts and humanities.
Why it's on our list
- From a parlour game to autocomplete
- How it learns, and where it slips
- Who is writing the rules
- Four questions worth asking
Location
The map comes from Google Maps, which sets its own cookies once it loads.
More in depth
The ObservatoryEarthset, and the long road back to the MoonArtemis II’s record flight, a rehearsal in Earth orbit, and the many roads to the lunar south pole.
The ObservatoryPearls on a thread: the satellites that wired the skyStarlink, OneWeb, Amazon Leo and China’s networks: internet from orbit, and the night sky’s cost.
The ObservatoryLondon to New York in two hours? The long, quiet road back to supersonicConcorde’s legacy, NASA’s quiet X-59, Boom’s Overture and the hypersonic dreamers, by status.