ReCAP’s approach to on-screen disclaimer footnotes in ads

Advertisers often place decisive qualifications in characters that are almost impossible to read during a normal broadcast: a price followed by an asterisk, a funding acknowledgement, a wagering condition, or a long legal sentence running along the lower edge. These on-screen disclaimer footnotes can alter the meaning of the prominent claim, yet they are easily lost in video archives, social reposts and fast live playback. ReCAP’s aim is to make that contextual layer as visible and searchable as the advertisement itself.

Manual logging usually depends on a reviewer noticing the notice, pausing the footage and copying its wording into a spreadsheet or asset-management system. The resulting record may omit the exact timecode, background frame, spelling variants or campaign version that produced it. That weakness becomes significant when teams need to compare a national broadcast with a metropolitan insertion or trace reused footage across several channels.

The EU-funded Real-time Content Analysis and Processing project takes a computational approach. Video analysis can locate text-like regions, recognise their wording, associate them with a particular advertisement and preserve the evidence as structured metadata. The output is intended to accelerate review rather than replace legal judgement, giving media teams a reliable starting point for checking whether a qualification was present and properly connected to its claim.

For Australian organisations, the value of this approach extends beyond simple text capture. Broadcast material may contain state-specific campaign edits, Australian pricing and consumer-law language, while public-service content can carry acknowledgements that differ from commercial advertising. ReCAP helps surface these details consistently without assuming that every visible line of small print has the same legal purpose.

Why small-print notices need structured records

A disclaimer is more than an extra line of text. It may qualify an offer, identify eligibility rules, distinguish an estimate from a guaranteed price or connect a promotional claim to conditions published elsewhere. In Australia, the Australian Consumer Law gives the Australian Competition and Consumer Commission enforcement powers over misleading or deceptive conduct, while most advertising does not require ACCC pre-clearance. A late-running qualification can therefore be significant even if the advertisement was never formally approved in advance.

Broadcasting adds another layer of oversight through the Australian Communications and Media Authority and sector advertising codes. The legal significance of a notice depends on its prominence, duration, placement and connection to the main message. Recording only its wording is insufficient: a reviewer may also need to know whether it occupied the screen for half a second, competed with a lower-third caption or appeared in a version aired only in Queensland.

Set out in ReCAP’s technical objectives, the project combines automatic metadata extraction, quality monitoring, face and logo recognition, and duplicate-content detection. Disclaimer logging fits this architecture because fine print is a compact metadata source with real temporal and spatial coordinates. It can be indexed, compared with other assets and attached to the advertisement from which it was extracted.

The resulting record should be treated as evidence, not a verdict. An automated system can report that a phrase was detected, where it appeared and how confident the reading is. It cannot conclusively determine whether the notice was legally adequate or whether viewers could reasonably read it. Those judgements remain with trained reviewers, supported by reproducible frame references rather than undocumented memory.

Detecting the fine print in moving video

A practical detection pipeline begins with temporal analysis. ReCAP can identify shot changes and advertisement boundaries, then examine frames for regions that behave like superimposed text. High-contrast edges, connected character shapes, stable placement and limited movement help separate legal notices from captions, channel bugs, sponsor logos and ordinary background signage. When a small disclaimer animates or scrolls, tracking across consecutive frames provides stronger evidence than any single image.

Each candidate region can then undergo optical character recognition. The recogniser benefits from rectification, contrast enhancement and enlargement, especially when broadcast compression has softened the edges of narrow lettering. Recognising the same notice across several frames allows the system to select the clearest version for transcription rather than averaging unreadable images together. This is particularly useful for dense terms, URLs and pricing qualifications that change only slightly between frames.

The difficult cases require context. A subtitle may sit in the same lower portion of the screen as a disclaimer, while a ticker contains large amounts of readable text that is unrelated to the advertisement. ReCAP’s broader analysis capabilities help separate these elements by combining visual characteristics with timing, scene structure and the relationship between on-screen objects and the surrounding programme. A repeated campaign can also be linked through duplicate-content analysis, revealing that one fine-print version travelled with several edits of the same commercial.

Speech recognition provides a useful secondary check, but it should not replace visual analysis. A voice-over might mention conditions that never appear on screen, while an on-screen notice may remain silent throughout the spot. The authoritative record for a visual disclaimer is therefore the text, its frame and its duration. Audio, logos and detected faces can enrich the record, helping reviewers understand who or what the notice concerns without becoming the primary source.

Turning visual evidence into useful metadata

A useful disclaimer record contains more than a transcription. It can include an asset identifier, campaign name, channel, programme, timecode, language, exact wording, alternative OCR readings, confidence scores, screen coordinates and start and finish times. A representative frame or short visual excerpt preserves the context in which the notice appeared. Hashes and duplicate-asset identifiers can then connect that observation to related masters, regional versions and online copies.

This structure makes the result useful during both live and archival operations. In a near-real-time workflow, a newly detected qualification can be attached to the current advert and placed in a review queue while the broadcast continues. A batch workflow can reconcile an entire day of captured media, identify every occurrence of a phrase and flag discrepancies between the live feed and the stored master. Media-assignment teams can search by wording or linked claim instead of relying on filenames created before transmission.

Reliable alerts also require a clear separation between analysis and delivery. An external guide to scheduled message delivery illustrates the same operational principle from outside the project: notifications need a defined trigger, destination, timing and delivery status. ReCAP’s role is to create the review event and preserve its evidence; whatever messaging channel a broadcaster or agency uses can carry the resulting alert to the responsible team.

Validation for Australian broadcast workflows

An Australian validation set should reflect the material organisations actually receive. It can include national free-to-air advertising, metropolitan and state-based variants, ABC and SBS public-service material, commercial channel trailers and clips supplied for digital platforms. Campaigns may be adapted for Sydney, Melbourne, Brisbane or regional markets, with fine print changing between executions even when the main visual remains identical. Standard-definition archives and heavily compressed online files deserve equal attention because they expose different recognition failures.

Recognisers also need the vocabulary of the local market. This includes Australian English spelling, dollar-denominated prices, GST references, state abbreviations and recurring wording such as “terms and conditions” or “conditions apply”. Where an advertisement includes an Aboriginal or Torres Strait Islander language, the system should preserve the original language and script while storing any translation as a separate field. Combining an interpretation with the original transcription could silently change what the issuer actually said.

Evaluation should measure several dimensions rather than treating recognition as a single pass-or-fail task. Precision and recall reveal whether notices are being found without flooding reviewers with ordinary captions. Character error rate measures the accuracy of the transcription, while spatial and temporal measures show whether the note is attached to the correct region and time interval. Confidence calibration matters just as much: a low score that sends everything to manual review offers little efficiency, whereas an overconfident score risks missing a meaningful qualification.

Human validation remains essential for prominence and interpretation. Reviewers need the source frame, surrounding sequence and relevant advertising rules, with their decisions stored as part of the audit trail. Access controls and retention periods should reflect the sensitivity of campaign information. Where face recognition is used to associate a notice with a presenter or sponsor, the presence of a person should not be treated as evidence of identity, consent or endorsement. The system records observable relationships; responsible teams determine their significance.

From automated detection to accountable decisions

Automation is most effective when it removes the search burden while preserving human accountability. A reviewer should receive a ranked set of detected disclaimers, each linked to the advertisement, frame evidence and extracted wording. Confirmed findings can be signed off, rejected findings can be labelled with a reason, and recurring errors can feed back into recogniser tuning. This creates a repeatable workflow instead of an opaque claim that the footage has been checked.

Three operating models show where automation and review fit:

Operating model Evidence produced Operational strength Control needed
Manual-only review Notes copied into a separate log Handles unusual wording and visual layouts Slow, inconsistent and difficult to reproduce
Automated detection only Text candidates, timestamps and confidence scores Fast triage across large volumes Susceptible to OCR errors and false matches
ReCAP-assisted review Detection linked to frames, assets and structured metadata Combines scalable analysis with documented human checks Requires review policy and ongoing quality monitoring

For an Australian broadcaster, the combined model can reveal that a mandatory qualification disappeared from a regional edit while remaining in the national master. For an agency, it can compare versions supplied to different television and digital publishers. In a media library, it turns an invisible overlay into a searchable object associated with the correct campaign, reducing the ambiguity that surrounds small-print records.

The durable point is that the smallest line on screen can carry decisive context. ReCAP’s value is to make that context findable, time-aligned and reviewable, so media teams can explain what was advertised and the qualifications that travelled with it.