Detecting AI in Client Stems: A Mix Engineer’s Practical Checklist

What You Are Actually Detecting

Clients may send: fully generative songs, AI-sung toplines over human beats, stem-separated rips from commercial tracks, voice-converted takes, or ordinary human recordings processed with AI assistants. Each case carries different legal and quality risks. Your job as mixer is not to play internet detective for clout—it is to protect delivery quality, your reputation, and contract scope.

No public detector is 100% accurate. False positives happen on heavy tuning, narrow codecs, and aggressive restoration. Treat tools as signals, not verdicts, and document what you observed.^[1]

As of July 2026, platform rules and label policies around AI disclosure continue to evolve—verify current distributor and DSP terms for any release you touch.

Intake Questions Before You Open the Session

• Origin disclosure Ask in writing: which tracks are human-recorded, which are AI-generated or AI-converted, and whether any commercial song was stem-split for “reference.”
• Rights confirmation Who owns the composition and master? If AI tools were used, does the client’s license allow commercial release? Get a short warranty in the invoice/contract.
• Source files Prefer multitracks at 24-bit with session sample rate. A single stereo “master” plus mystery stems is a red flag for rip+split workflows.
• Tuning / editing history List of pitch tools, vocal riders, and restoration already applied so you do not double-process artifacts.

If the client refuses any disclosure on a commercial release job, price in risk or decline. Mix engineers are not unpaid copyright insurers.

Audio Tells (Useful, Not Infallible)

Stem-separation tools (Demucs-class, commercial splitters, editor features) leave characteristic warble in cymbals, bass sidebands, and vocal fricatives. If a client claims “live band multitrack” but every instrument is surgically isolated with separation haze, ask how it was recorded.^[2]

Tooling Stack for Screening

Keep a written studio policy: what you screen, what you do not claim, and how you handle suspected infringing stems (e.g., known hit song separations). Do not post client audio to random public detectors that claim ownership of uploads.

Contracts, Ethics, and Delivery

Your mix agreement should state that the client warrants they own or license all material, including AI-generated parts, and will indemnify you against infringement claims. If you only mix and do not clear rights, say so explicitly.

If you detect a likely rip of a commercial master, stop and escalate—do not “make it less obvious.” If the client discloses AI vocals for an original song, you can still mix; document disclosure needs for their distributor and any guest features.

Quality bar: AI vocals often need different de-essing, less reverb wash (which exposes artifacts), and careful stereo. Budget time for artifact control the same way you budget for noisy location dialogue.

Recommended Studio Workflow

• Tier A — full disclosure + clean multitracks Normal mix pricing and schedule.
• Tier B — partial AI / heavy separation Add artifact-cleanup hours; limited revision rounds.
• Tier C — suspected unauthorized stems Pause project until paperwork clarifies; retain kill fee if work already started.

This is risk management, not anti-technology ideology. Generative tools are part of 2026 production; unlicensed stems and silent substitutions are the actual problem. Stay factual with clients—show clips of artifacts rather than moral speeches.