Design a Fair Use / Excerpt Handling System

Medium45 min
1 / 30
System design answer

Design a Fair Use / Excerpt Handling System

Medium45 min30 sections
understanding•9 min read

Problem Statement: Matching Excerpts Without Punishing Legitimate Commentary

Frames the system as a segment-level content matching and policy problem, not a simple duplicate detector.

Problem statement

Design a Fair Use / Excerpt Handling System for a user-generated streaming platform. Creators upload reaction videos, reviews, critiques, educational breakdowns, and parodies that intentionally contain short portions of copyrighted reference material. The platform must automatically detect how much of a registered reference asset appears in each upload, classify the usage as a short excerpt versus a substantial reproduction, apply rights-holder and jurisdiction policy, and route genuinely ambiguous cases to human review instead of silently blocking or silently allowing them.

This is not a whole-file duplicate detector. A naive fingerprint match that says "this upload contains reference asset R" is insufficient, because the legal and product difference lives in the segment: three seconds of a music video inside a ten-minute critique is a different situation from nine minutes of the same video reuploaded with a one-second intro. The system therefore has to answer four questions for every upload: what matched, where it matched, how much of both the upload and the reference it covers, and what policy applies to that specific shape of usage.

Why the problem is distinctive

A recommendation backend can be wrong and recover. A copyright decision is a legal act with real liability. Over-claiming chills legitimate speech, invites counter-notifications, and damages creator trust; under-claiming exposes the platform to takedown obligations and rights-holder relationships. Fair use in the United States is a four-factor legal defense, and no automated system can genuinely adjudicate it. YouTube has repeatedly stated that Content ID cannot determine fair use and that disputes are where that question is surfaced. Our system must therefore be honest about its role: it measures the objective, computable factor of amount and substantiality, applies explicit policy, and escalates the rest to humans with rich evidence.

The four architectural planes

  1. Ingest and fingerprinting plane: decode uploads and reference assets, extract robust perceptual fingerprints, and persist them for matching.
  2. Matching plane: find segment-level alignments between an upload and the reference catalog, and quantify matched duration and ratios.
  3. Policy plane: evaluate rights-holder rules, excerpt thresholds, jurisdiction law, and attribution requirements to produce a disposition.
  4. Governance plane: disputes, counter-notifications, escalation queues, audit evidence, and rights-holder transparency reporting.

A strong answer keeps these planes separate. Matching produces evidence; policy produces a disposition; humans resolve disputes. Measurement errors, policy errors, and legal judgment errors are different failure classes and must be observable separately.

Public operating baseline versus design assumptions

Public figures establish that the category is real and enormous. YouTube has reported that more than 500 hours of video are uploaded every minute, and its Copyright Transparency reporting describes Content ID claims as overwhelmingly automatic. Audible Magic and Zefr describe fingerprint-based identification used across multiple platforms. These are cited company signals, not requirements for our fictional system.

For capacity planning, this answer explicitly assumes a mid-size UGC platform with 3,000,000 uploads per day, average upload length 90 seconds, a reference catalog of 20,000,000 fingerprints, and a five-times event peak. Unless a number is tied to a citation, it is a stated design assumption, target, or budget, not a claim about any company's private architecture.

Key Highlights

  • •The unit of decision is the matched segment, not the whole file, so matching must quantify amount and placement.
  • •The system measures the computable 'amount' factor and applies policy; it never claims to adjudicate fair use itself.
  • •Four planes: ingest/fingerprinting, matching, policy, and governance/disputes.
  • •Public company figures provide context; every uncited scale or SLO in this answer is an explicit design assumption.
  • •Over-claiming and under-claiming are distinct failure classes with distinct owners and metrics.
Lead With the Segment
State in the first two minutes that the decision unit is the matched segment and its ratio, not a whole-file match. This instantly separates an excerpt system from a duplicate detector.
Do Not Claim to Decide Fair Use
A design in which an algorithm declares 'this is fair use' misunderstands the law and will fail a serious interview. The system measures amount, applies policy, and escalates judgment.

Section Rescue Kit

Buzzwords to use:

Amount and SubstantialityDisposition

Safe statements:

  • "I will separate what the system can measure from what only a human or court can judge."
  • "Before selecting services, let me define which outcomes are automatic, which need attribution, and which must escalate."
Design a Fair Use / Excerpt Handling System - System Design | WinJob | WinJob