← Joseph Denne

July 2026

Fi1m is an experiment in editorial content. Two questions sit underneath it: do people care where their reviews come from, and does being current matter more than being considered?

Fi1m is a film and television publication in which every review is written by a model and every page is signed off by a human editor. I led the product and the strategy behind it, and the engineering of the pipeline that produces it. The format is old and trusted: a review, a score, where to watch. The method is new.

The experiment

The first question is whether readers care where content comes from. The honest answer is that we don't know yet, and that nobody who claims to does either. The received view is that provenance matters enormously and that machine-written work is tolerated at best. The evidence so far is thinner than the conviction. People use what's useful. So Fi1m discloses its method on every page, in the footer, where the reader is, and then we watch what happens. If readers reward the disclosure, that's an answer. If they ignore it and keep coming back for the reviews, that's an answer too.

The second question is about pace. My contention is that in the current world, pace matters most. To be relevant you have to be current: the review has to exist on the night the thing drops, in the country the reader is in, with the streaming link that works. There is room for consideration and depth, and I want Fi1m to have both, but they can only be built on trust, and trust comes from use, and use is driven by being there now. Depth without currency is a library. Currency without depth is a feed. The bet is that you earn the right to the first by being reliable at the second.

Voice is the product

None of that works if the reviews read like a spec sheet. Fi1m has a distinct tone of voice, and the reviews open with hooks that are meant to be funny, or at least sharp. That isn't decoration. When the writing is machine-made and the volume is high, voice is what stops the publication being interchangeable with every other machine-made publication. A reader should be able to tell a Fi1m review from the first line. Getting a model to hold a voice consistently across hundreds of titles was harder than getting it to be accurate, and more important commercially.

Where the editor goes

When a person writes a review, judgement is spread across every sentence. When a model writes it, judgement collects at a few points: what gets covered, whether the facts hold, whether the score can be defended, whether the hook lands, and whether the prose says anything or merely sounds as if it does. The editor's time goes entirely to those points. The surface area shrinks; the weight on each point goes up.

We publish the critics' consensus as a single number beside our own score for that reason. When the two diverge sharply it's either a genuine difference of view, which is worth having, or an error, which the editor needs to see before anyone else does. Putting them side by side makes the check public and automatic.

Checking is the pipeline

Models state true and false things with the same confidence. On Fi1m, every factual claim, a runtime, a certificate, a streaming service, a director's name, is verified against a source before a page goes live. This is unglamorous and it's where most AI-generated content fails: the output looks finished, so the checking gets skipped. It's also the only way to be fast and trusted at the same time, which is the whole bet.

It's the same discipline we run on the XE bug bounty. Read in private, answer in public, and let the record show when you were wrong. A publication that can't be wrong in public isn't a publication.

Readers who aren't people

Every Fi1m page has a markdown twin at the same address with .md appended: the same review, structured, with nothing to scrape around. A growing share of the audience is an assistant asked "what should we watch tonight?" that goes and reads reviews on someone's behalf. Those readers care about currency and accuracy and not at all about provenance, which makes them a useful control group for the first question.

What I'd carry into the next one

Disclose the method everywhere and let the audience decide whether it matters. Be there first, and be right, because the second buys you the right to go deeper. Give the thing a voice or it's a commodity. Put the human where the judgement is and nowhere else. And budget the writing as the cheapest part of the system, because it is now, and build the organisation around the parts that aren't.

← More writing · Also as markdown.