Evidence

What AI can and cannot tell you

By the neuro-viral team · 20 Sept 2026 · 8 min read

The category is full of confident claims. This is an attempt at the honest version, including the parts that are inconvenient for a company that sells a prediction.

What the model actually computes

A predictive model of the kind we build takes a video's visual, audio and language streams and estimates the average viewer's second-by-second response across the cerebral cortex. It is trained on brain-imaging recordings of people watching video, so it has learned a relationship between what is on screen and how the cortex responded.

That is the claim we are comfortable defending: predicted cortical response, for an average viewer, from the video alone.

What that is good for

• Locating moments. It points at specific seconds rather than giving a verdict on the whole video, which is what makes it actionable.
• Comparing versions. Two cuts, same model, same basis — you can see where they diverge instead of arguing about which feels better.
• Arriving earlier. It works before publication, which is the only window where changing the video is still cheap.
• Removing the author's blind spot. You cannot watch your own edit cold. A model can.

What it is not good for, stated plainly

It does not predict performance. Views, watch time, click-through, conversions and revenue depend on distribution, targeting, timing, budget, offer and audience. A model that sees only the video cannot see any of those. Anyone claiming otherwise is selling a stronger story than the evidence supports — and we tested this ourselves: in our own evaluation, predicted response did not stand up as a forecast of engagement outcomes.

It does not read an individual viewer. There is no scanning of your audience, no personal data, no per-viewer measurement. It is one prediction of a typical response.

It does not cover the whole brain. Reward, emotion and memory systems sit below the cortex and are not modelled. Claims about what a video makes people feel or remember do not follow from cortical prediction.

It degrades with distance. Like any model, accuracy is highest on material resembling its training data and falls off as content moves further away from it. That is a general property of predictive models, and we would rather state it than let you discover it.

Why the honest version is still worth paying for

Because the alternative to an imperfect early signal is usually no early signal at all. The realistic comparison is not “prediction versus certainty” — it is “prediction versus one person's opinion in a review meeting”, or “prediction versus finding out after the media budget is spent”.

Used as a filter, a prediction earns its place: it narrows what you test, it makes creative debate specific, and it costs a fraction of a live test.

How to use it without overtrusting it

Treat every model reading as a hypothesis with a timestamp attached. Look at the moment, decide whether the model found something real, make a change if it did, and then let a real audience settle the outcome. Keep a record of prediction, change and result. Over enough videos, that record tells you how much to trust the model on your content — which is the only calibration that actually matters.