Why we do it this way
A useful film site needs a review of everything you might actually watch, kept current as things arrive on and leave the streaming services, and it needs to do this without covering every page in advertising to pay for it. No human team of a sensible size can write and maintain 20,000 considered reviews. AI models can, if they are held to an editorial standard rather than let loose. We built the standard first and the site second.
What the models write
Everything on a film or series page that reads as prose: the review, the synopsis, the summary of the critical consensus, the one-line reasons attached to recommendations. Person pages: the biography and, where there is one, the Fi1m perspective. Longer features. The weekly email. And Discover's answers, which are composed from our own published reviews and say which ones they drew on.
Facts on those pages — cast, running time, certificate, release dates, where a title is streaming and at what price — are not written by a model. They come from licensed data sources and are checked against them. The model's job is to write about the film, not to remember things about it.
What they never do
- Invent a quotation. Every quoted critic is linked to the review the words came from.
- Invent a fact. Any claim about a film's making, release or reception must trace to a source the pipeline recorded. If it cannot, the sentence goes.
- Pretend to be a person. There is no fictional critic with a name and a photograph. The byline is Fi1m.
- Copy a review. We summarise and link the press; we do not reproduce it.
- Reveal the plot. Nothing past the first act, in reviews or in Discover, unless you ask Discover directly and it warns you first.
- Write members' reviews, or touch them. See below.
How work is checked
Nothing is published straight from a model. A draft passes through a second, independent model whose only job is to check it: every factual claim against the recorded sources, every quotation against its origin, the spoiler rule, the house style, and a list of words we do not use. A draft that fails any check is sent back or discarded, not patched. What passes is queued for the editor.
A human editor reads a sample of everything published each week, all of the features and the weekly email in full, and every review of a title in the UK top twenty. The editor also reads ten reviews a week blind, mixed with pieces from a reference set written and edited by hand, and has to pick the reference pieces out. We publish the result: in August 2026 the editor identified them correctly 4 times out of 10, which is what chance would give.
The voice
Fi1m's reviews are written in a single house style: unhurried, dry, specific about craft, allergic to hype, ending on a recommendation or a caveat rather than a summary. The style is defined by a set of thirty reviews the editor wrote and edited by hand, which the models are shown as examples, and by a set of rules they are held to. When a review reads like a person, that is the point. It is still a model's work, and the disclosure line says so.
What the people do
Set the editorial standard and the rules. Write and maintain the reference set. Read, blind, every week. Read everything that goes out about the films most people are choosing between. Decide which sources we trust and which we don't. Handle every correction and every complaint personally. Decide what the site is for. A named editor is accountable for everything published; their name is on the contact page.
Members' reviews
Ratings and written reviews from Fi1m members are written by those members, appear under their names, and sit in their own clearly labelled section beneath our review. A model reads each one before it is published, for spoilers outside a marked section, abuse and spam, and nothing else; it does not edit the text, and members' words are never used to train or steer the house voice. If a member's review is held back, they can appeal to the editor.
Disclosure
Every page carrying model-written prose says so in one consistent line, and links here. Discover says so in its answers. The weekly email says so in its footer. We think a reader should never have to wonder, so we never make them.
When we get it wrong
We will. When a reader tells us, or the editor finds it, the page is corrected and the correction is logged on the corrections page with the date, what was wrong and what changed. Corrections of fact are made within a working day. Disputes about opinion are not corrections, and we say so politely. If a correction changes the argument of a review, the score can change with it, and the page shows the old one.
Models in use
We use commercial models from more than one provider and we say which. As of September 2026, reviews and other prose are drafted by Anthropic's Claude family; independent checking is done by a model from OpenAI running in Microsoft's UK data centres; research and retrieval use smaller models from both. Versions are fixed, not floating; when we change one, the whole reference set is re-run and the editor signs off before a word is published. Data processing is described in the privacy policy. No reader data is used to train any model.