FAQ

aiScreeningRoom predicts how real-world audiences will receive an unreleased film by watching the actual footage - the entire cut - with no human test audience in the room. You upload your film, declare how you are positioning it, and get a test-screening-style report: predicted audience reaction, craft assessment, where the film holds or loses people minute by minute, comparable titles, and (for feature narratives) an IMDb landing-zone projection.

It is built by the team behind iScreeningRoom (called "a revolutionary online test screening platform" by Forbes), using more than ten years of taste-matched online panel testing to tune the instrument.

A traditional in-theater test puts your cut in front of a recruited audience - one audience, one night, one theater - and creates a leak surface every time. aiScreeningRoom never shows your film to a human test audience. Testers never see the film, so they cannot leak it.

Because there is no recruitment or venue, you can test at any stage of post - rough assembly through picture lock - and re-test after editorial changes. Testing becomes part of cutting, not a single event after the fact.

iScreeningRoom is an online human audience test screening platform: nationwide audiences selected for interest in the film's genre, premise, and comparable titles, typically targeting about 300 completed surveys per film.

aiScreeningRoom is the AI instrument built on that panel data. It predicts the kind of reception those panels measure - without putting the cut in front of people. Human panel testing continues in parallel; new panels keep growing the calibration corpus.

aiScreeningRoom predicts how a taste-matched human panel would score the film. That 1–5 score is calibrated to more than 35,000 real iScreeningRoom survey responses, with nearly 43,000 moment-level reactions used to weight what audiences actually weight. No other AI, including ChatGPT, has that corpus.

Anyone can prompt a model at footage. The difference is a decade of panel data, a score you can check against those panels, and a published accuracy record.

• Studios and independent filmmakers and producers who need a reception read before release.
• Distributors, sales agents, and buyers who want a comparable, evidence-backed prediction before acquisition or positioning decisions.
• Editors refining the cut.

Your uploaded video file - not a synopsis alone, not a trailer, not a script. The system watches the full runtime in multiple independent viewing passes, then aggregates those reads into one report. Narrative synthesis and comparable-title research use additional language models and public film databases; the reception grades come from the footage evaluation.

After scoring, the report includes a positioning section that compares what you declared against what the footage actually plays as - advice on the sell, not a rewritten score for a different pitch.

For a feature-length narrative film, the full product:

• Predicted audience reaction on a 1–5 panel scale, with a range across independent viewing passes (pass-to-pass variation - not a full error bar on the film)
• Craft score out of 100 (reported separately from audience prediction)
• Minute-by-minute engagement timeline (where it holds or loses people), with cited moments
• Findings and synthesis grounded in timestamped evidence
• Positioning alignment (sold-as vs plays-as)
• Optional poster/key-art read if you uploaded art
• Comparable films with verified public scores and domestic grosses
• An IMDb landing-zone projection (center estimate with a range, not a single guaranteed number)
• Chat with Felix to dig into the results

For shorts, trailers, non-narrative, or other non-feature footage, you get a qualitative critique - synthesis, findings, and timeline - without calibrated audience, craft, or IMDb scores. Classification happens after the system watches; it is not something you pick from a menu to unlock grades.

Typically about 20–90 minutes after upload, depending on runtime. Longer features sit toward the high end.

Yes. Each screening evaluates the file you upload. Re-upload a new cut as a new screening when you want a fresh read after editorial changes. That is one of the main advantages over a one-night human test.

Felix is aiScreeningRoom's interactive cinephile analyst. You can ask what is working, what is broken, pacing and performance questions, and how to think about fixes.

The claim rests on two independently checked links. Neither peeks at the answer it is graded on.

1. Do taste-matched human panels predict public IMDb?
For corpus films that later earned an honest organic IMDb rating (screened for organized voting, verified title identity, adequate panel size), panels and public ratings usually moved together. Across 36 qualifying films, graded out-of-sample: when panels ranked films high, organic IMDb usually did too; average miss about half an IMDb point; 32 of 36 (89%) within 1.0 IMDb point; 75% within three-quarters; 64% within half a point.

2. Does the AI predict those panel results?
On 9 sealed holdout films - locked away from training and tuning - the AI was off by about a quarter of a star on average on the panel's 1–5 scale, with strong rank agreement with real panels.

A skeptic will ask: is this just recognizing famous movies? We tested that. We asked, with no footage, what the system already knew about a film’s public reception. On 44 films it could not recall, it was about as accurate as on films it could recall (off by 0.26 vs 0.28 on the 1 to 5 scale). So the score is coming from the picture, not from the title. Those 44 still include films used while we built the method. The accuracy number is the 9 films locked away from that work. That is why 0.285 on 9 films is the claim, not 0.26 on 44.

Full methodology, figures, and limitations: Validation summary

Unreleased films do not have public ratings - that is why you need a prediction. Public ratings also never enter the calibration that trains the instrument. The panel mean is the sole calibration target; IMDb is validation and reporting (the landing zone), not something the AI is fit to.

No. A typical in-theater test is one audience, in one theater, in one market, on one night - often recruited with a cash incentive, and sometimes forced into age/gender quotas that pull in people who would never choose the film on their own. That measures who showed up for the invite, not how the film's natural audience receives it. Sometimes the sample is also very small. Training a model on those scores teaches it to reproduce that construct's noise and bias.

That problem is why we built iScreeningRoom more than ten years ago: interest-based nationwide panels - people who opt in because the genre, premise, and comps match what they're interested in - sized for usable statistical inference (typically about 300 completed surveys). No cash bounty to fill seats. This is why iScreeningRoom panels tend to align with IMDb scores - in both cases, the film's natural audience is largely who is doing the rating. And that is why aiScreeningRoom is calibrated to the iScreeningRoom panel construct and data.

Stated plainly because a trust story that hides weaknesses is not trustworthy:

• The AI-to-panel record is still early (sealed holdout of 9 films). The beta exists in part to grow a prospective ledger: predictions logged before outcomes exist, scored only when panels or public ratings arrive later. If accuracy moves against us as n grows, claims will be repositioned accordingly.
• Films audiences love can be under-called. Top-band compression is a known failure mode.
• On weaker films, predictions can skew high. Bottom-band overprediction is holdout-confirmed.
• Public ratings for small or polarizing films are noisy ground truth. Paid vote-lifting and brigading exist; we screen and weight for visible cases. The projection assumes an organically rated release. Films marketed beyond their natural audience tend to land below the projected center.
• Faith dramas and documentaries: public IMDb is often distorted by culture-war voting or too thin to use. We report the panel-scale prediction and say why, rather than manufacture a false-precision IMDb forecast.
• Some films succeed or fail on taste a craft-and-engagement read underweights. Those perception inversions define part of the uncertainty band; reports disclose uncertainty rather than promising an oracle of taste.

No. Reports are predictions and craft assessments with measured accuracy and stated limitations - not guarantees of commercial outcome. Comparable titles with verified scores and grosses help you reason about commercial context; they do not forecast your P&L.

No. Analysis is automated. Your footage is not put in front of a human test panel through this product.

Customer film files are not viewed in the ordinary course. During private beta, authorized admins may read reports and chat transcripts to diagnose support issues. That access is read-only.

No. Your film file is not used to train, tune, evaluate, or benchmark any model. Reports, transcripts, and predictions are treated the same way, except that prediction and outcome data may feed the accuracy ledger described on the Validation page.

We use paid AI APIs under contract. Those providers do not use your film to train their general models. Who they are, and what each one receives, is in the Privacy Policy.

After analysis completes successfully, the source upload is deleted from storage. If that first delete fails, we retry in the same run, then about once an hour until the object is gone. Seven days after upload is the outer limit: leftover source files are removed automatically. The report shows the removal date once storage confirms the delete.

A failed analysis keeps the upload so you can retry without uploading again.

We delete the copy sent to the viewing API when processing finishes.

Reports and chat transcripts remain so you can return to your results. They do not include a downloadable video file.

For the self-serve product: we minimize copies and dwell time, store uploads in private cloud storage, use paid AI APIs under contract that do not use your film to train their general models, and delete source files after a successful analysis as described above. That is a strong practical posture for unreleased indies. It is not a TPN-assessed studio walled garden. Details: Security and Footage.

Not in the self-serve product. Studio tenancy is a separate engagement, in development. Self-serve aiScreeningRoom is the indie path. A studio deployment would be dedicated tenancy in the studio's environment, not customer API keys pasted into the shared app.

• Private cloud: the evaluation would run in the studio's own Google Cloud organization, under that studio's existing content security program. aiScreeningRoom is not currently TPN assessed.
• Roadmap - on-premises: true on-prem as open-weight video models mature to feature-length evaluation quality.

Studio pricing is not the self-serve beta or catalog rate.

Call for pricing

No. We do not sell or license your film. Paid AI APIs process a downscaled copy to produce your report. Those providers do not take ownership of your film. They do not get a license to sell, distribute, or reuse it as their own work. Who they are, and what each one receives, is in the Privacy Policy.

MP4, or MOV - H.264 delivery files preferred. Maximum 15 GB. Minimum about 2 minutes runtime (junk/empty upload floor). ProRes and camera masters are usually too large; export a compressed delivery screener.

Optional key art up to 1 MB.

Any stage of post where the picture is watchable as a film - rough assembly through locked picture. Incomplete assemblies will be judged as incomplete assemblies; the instrument reads what is on the timeline.

You still get a serious qualitative critique. You do not get the calibrated audience prediction, craft score, or IMDb landing zone. Those grades are for feature-length narrative classification.

If your film targets the faith-based market, say so at intake. That segment has different public-rating dynamics (especially on IMDb). We report segment context honestly rather than applying a hidden numeric fudge to the score.

Yes. You can create a read-only link to a completed report and send it to people you choose. Anyone with the link can open it. The link does not expire. Felix chat is not on the shared page.

Yes. After analysis, an admin can upload a real audience survey CSV. When that survey is released on the report, you see the panel scorecard next to the matching topline numbers, plus counted themes from the open-ended answers.

During private beta, screenings are complimentary for invited filmmakers.

Working filmmakers with feature films at any stage of post, by invitation / allowlist during private beta. Apply at aiscreeningroom.com. Beta participants help grow the prospective accuracy record (predictions logged before outcomes) and receive priority access when the commercial platform launches.

Contact support. If the failure is a platform error, that is a support case. Prediction accuracy itself is not a refund matter - the product is an instrument with a published record, not a guarantee.

As a decision input before acquisition, ask-setting, or positioning: predicted reception with uncertainty, craft-vs-reception gap, comps with verified scores and grosses, and an IMDb landing zone you can line up against known titles. It is faster and more iterable than a one-night recruited screening, and it does not require putting the seller's cut in front of strangers.

It does not replace taste, negotiation, or legal diligence. It replaces "we have no measured reception signal until we screen it."

Not through a human test audience - there isn't one. Processing still involves cloud storage and paid AI APIs under the retention rules above. For acquisition diligence at indie scale, that is usually the right tradeoff versus recruiting viewers. For major-studio security requirements, ask about the private-tenancy path rather than assuming self-serve meets that bar.

Incumbents assert authority and typically publish no auditable accuracy record against real-world outcomes. aiScreeningRoom publishes its methodology, figures, and limitations. Measured accuracy versus asserted authority is the comparison we invite.

Traditional tests also measure a different construct when they use paid convenience samples in a single venue. Our calibration target is taste-matched panel reception of the kind iScreeningRoom runs - closer to natural audience than "whoever showed up that night."