AI Disclosure

Last updated: May 24, 2026

Safi uses artificial intelligence to analyze the music you upload. This page explains what AI does, what it doesn't do, and how your data is handled.

What Safi's AI does

Safi's analysis pipeline uses one large language model plus multiple commercial music databases to produce a safety report:

Language model (the analysis brain):

• Anthropic Claude — the AI that reads your lyrics, weighs results from the databases below, applies copyright doctrine (substantial similarity, idea/expression, de minimis, scènes à faire), and produces the scored report.

Music databases Safi queries during an analysis:

• AudD — commercial audio fingerprinting against released recordings (strongest signal when you upload audio).
• AcoustID — open-source audio fingerprint matching backed by MusicBrainz.
• MusicBrainz — canonical artist/song metadata to verify cited matches are real registered works.
• Spotify — track search across the global commercial catalog by title and key lyric phrases.
• Pandora — additional commercial catalog search.
• Genius — lyrics database used to find songs with overlapping wording.
• MusixMatch — licensed lyrics database; stronger coverage for Latin and Brazilian-Portuguese music than Genius alone.
• Hooktheory — chord progression and music-theory pattern reference.
• Discogs — the world's largest crowd-sourced music release database (~17M releases); catches obscure, independent, and older works.
• Deezer — alternative commercial streaming catalog; catches regional and pre-release tracks not on Spotify.
• Last.fm — track search plus listener / tag data; surfaces adjacent artists and stylistically similar works.
• WikiData — structured music-entity database used to verify that cited matches are real registered works.
• WhoSampled — sample / cover / remix relationship database (the most directly relevant source for sample-clearance risk). Active when a partnership API key is configured; otherwise we surface a search-URL hint for human follow-up.
• Bandcamp — independent and pre-release uploads outside the major streaming catalogs. Active when API access is configured; otherwise a search-URL hint.
• OpenAI Moderation — safety classifier checking submitted content against abuse policies (separate from copyright analysis).

Not every database fires on every analysis. Some are gated behind credentials that may or may not be configured at a given time, and the orchestrator skips ones that are unavailable or return no signal. Each analysis report includes a list of which sources were consulted.

The output is a safety report with risk scores, cited matching works (if any), recommended fixes, and a side-by-side comparison panel when a specific lyric match is found.

What Safi's AI does NOT do

• Safi does not make legal determinations. A clean report is not a guarantee that your work is free from infringement claims.
• Safi does not register, file, or transmit your work to copyright registries.
• Safi does not store your audio long-term by default (see Data Handling below).
• Safi does not use your songs to train AI models.

Data handling

Lyrics — text you submit is sent to Anthropic's API for analysis. Anthropic processes this data per their privacy policy and does not use it to train their models. Your lyrics are stored in your Safi account so you can revisit reports.

Audio files — uploaded audio is sent to AudD for fingerprinting and deleted immediately after analysis. The AudD fingerprint match data is retained in your report so you can revisit results; the original audio file is not retained.

Reports — your generated safety reports are stored in your account so you can access them later. You can delete individual reports at any time from your dashboard.

Known limitations — what Safi sometimes gets wrong

Safi is a fast first-pass safety filter, not a copyright lawyer. Like every AI system, it can be wrong. These are the specific failure modes we have observed and are actively working to reduce. We surface them here so you can read every report with the right context.

Song titles are not copyrightable

US Copyright Office (37 CFR § 202.1) is explicit: “Names, titles, and short phrases or expressions are not subject to copyright protection.” Hundreds of songs legally share titles — there are 30+ songs called “Tonight,” many “Hold On”s, multiple “Free Fallin’”s. If Safi flags your song because the title matches a known release, but the melody, vocal, and lyrical substance are different, that is not infringement. We have explicit prompt rules to score these as low-risk; if you see a high score driven purely by a title match, please report it.

Audio fingerprint matches can fire on production texture

AudD’s commercial fingerprinter is excellent but it matches on acoustic features — tempo, key, instrumentation, production texture — not just on copied melody. A fingerprint hit on a song you’ve never heard can mean the engines share a chord progression, a Brazilian percussion groove, or a similar mix — none of which is protectable. If Safi reports an AudD match but you listen at the timecode and can’t hear your song in theirs, the match is most likely production-texture overlap, not melodic borrowing. We weight AudD matches by cross-database confirmation precisely because of this.

Database matches mean “recognizable,” not “protected”

When multiple databases (Spotify, Last.fm, Deezer, MusicBrainz, Discogs) all return the same matched title, it just means that song exists and is indexed in many places. It does not mean the matched song’s phrase or hook is legally protected against your reuse. Cross-database confirmation increases our confidence that the match is real; it does not change copyright law.

Genre conventions and cultural language

Brazilian gospel, sertanejo, MPB, K-pop, reggaeton, country, and many other genres have shared vocabularies — scripture references, common metaphors, standard chord progressions, characteristic instrumentation. Safi has rules to ignore these as scènes à faire (genre conventions), but the model occasionally over-flags. If your song uses common cultural or genre language and Safi scores it high, please listen to the cited match before changing anything.

The model has a knowledge cutoff

The underlying language model has a training cutoff, meaning very recent releases or extremely obscure works may not be cited accurately. Database lookups partially compensate for this, but Safi cannot guarantee coverage of every song ever released.

What to do when Safi seems wrong

• Listen to the cited song — we now include direct Spotify / Apple Music / Deezer links at the matched timecode. If you don’t hear your song in theirs, trust your ears.
• Re-run the analysis — if you checked “Keep my file for 30 days,” you can re-run the same audio against an updated pipeline as we refine the model.
• Consult an attorney for high-stakes releases — a sync deal, major-label release, or commercial campaign should always involve qualified entertainment counsel, regardless of what Safi says.
• Email us — support@akilitech.io if a score looks wrong to you. We use real cases to calibrate the model.

We will keep updating this list as we learn more from real reports. Safi’s scoring reflects what the AI detected against the databases it has access to — AI systems can produce false positives (flagging similarity where none exists in fact) and false negatives (missing real similarity). Always review the cited works manually before making release decisions on tracks that score in the higher-risk ranges.

Your rights

You can:

• Delete your reports at any time from your account dashboard
• Request full account deletion by emailing support@akilitech.io
• Receive a copy of the data Safi has about you (data export) on request

Changes to this disclosure

We may update this AI Disclosure when we add new AI features, change models or providers, or alter how we handle your data. The "Last updated" date at the top reflects the most recent change.

Contact

Questions about how Safi uses AI? Email support@akilitech.io.