Legal · AI Policy

AI Policy

Pre-incorporation draft. The status note below explains what that means.

Talisman is a platform for human creative work. AI is a real tool, present in real workflows, and not the enemy. The point of this policy is to keep the bargain that artists upload to a platform that does not feed their work into model-training pipelines, that does not allow others to do so, that does not host imitation of their voice and likeness without their permission, and that says all of this clearly enough to be enforceable.

Status. Pre-incorporation draft. The AI Policy is operational from the start and binds the Talisman Project today; it will be reissued under Talisman Music CIC on incorporation. Detection mechanisms named below describe the production pipeline as it exists; see What is Talisman? for the technical context.

1. The two commitments

1.1 We do not use Artist Content to train AI models that produce sound recordings

Talisman does not, and will not permit any third party to, use any Artist's Content to train, fine-tune, or evaluate machine-learning models intended to generate sound recordings, voices, lyrics, music compositions, or other creative outputs, except with that Artist's express, separately-given written permission.

This commitment binds:

  • Talisman's own staff, contractors, and infrastructure
  • Any vendor or processor we engage (we contractually require them to honour it)
  • Any successor entity to Talisman following incorporation, reorganisation, or sale

The narrow exceptions are:

  • Detection and integrity work that runs inside the Service itself. Audio fingerprints, content fingerprints, and short feature vectors are used to detect AI-generated content, copyright matches, fraud, and platform-policy violations. These pipelines do not produce generative outputs and the model artefacts are not made available outside the Service.
  • An Artist's own voluntary, opt-in participation. An Artist who chooses to license their work for training (for example, to train a model on their own catalogue and retain the model output) does so under a separate, explicit, written grant. Default is opt-out for everyone; participation is opt-in for those who want it.

1.2 Third-party scraping for AI training is prohibited

Crawling, scraping, indexing, or other automated extraction of Service Content for the purpose of training, fine-tuning, or evaluating AI models is prohibited under the Terms of Use and the Acceptable Use Policy. This is enforced through:

  • A machine-readable opt-out at the network layer (robots.txt, including specific blocks on AI crawlers; reservation of rights under Article 4(3) of Directive (EU) 2019/790 (the DSM Directive); future support for ai.txt, C2PA do_not_train assertions, and the W3C TDM Reservation Protocol as those standards mature).
  • Clear contractual prohibition in the Terms of Use and the Acceptable Use Policy.
  • Detection and rate limiting on platform crawl behaviour.
  • Civil action where automated extraction at scale is detected.

We treat this as an active commitment. We notice when crawlers ignore our opt-out, and we act on it.

2. AI-generated and AI-assisted Content on the platform

The line between "AI-assisted" and "AI-generated" is real and worth being precise about.

  • AI-assisted Content uses AI tools as part of a human-authored production process: a synth voice as one layer in a track that is otherwise human-performed and human-arranged; AI-generated stem material treated as a sample inside a human-led work; AI mastering or noise-reduction on a human-recorded source. This is permitted, and is the dominant category in modern production.
  • AI-generated Content is wholly or substantially produced by an AI model: a track whose primary instrumental and vocal lines are model output; a piece of music produced from a single text prompt; an album generated end-to-end by a generative service.

2.1 Disclosure

Wholly or substantially AI-generated music must be disclosed at upload via the AI-content disclosure flag in the upload flow. Where Talisman delivers Content to downstream services that require AI-content disclosure (Spotify, Apple Music, Deezer, others as their requirements stabilise), we relay the disclosure using the relevant DDEX-standard fields. Failure to disclose is a breach of this Policy and may result in removal of the Content and other action under the Notice & Takedown Policy and section 5 below.

2.2 Acceptability of AI-generated Content

AI-generated Content is permitted on the platform only where:

  • The disclosure under section 2.1 is in place.
  • The Content does not infringe the rights of any third party (this includes copyright in training material and any rights in the prompts or model outputs).
  • The Content does not imitate an existing artist's voice, name, likeness, or distinctive style without that artist's permission. This is covered in detail in section 3.
  • The Content is not designed to inflate streaming counts, generate fraudulent transactions, or otherwise manipulate Service metrics.

Where these conditions are not met, the Content will be removed.

We may, at our discretion, exclude AI-generated Content from editorial and algorithmic surfaces (for example, the discovery feed), and we may apply per-track tags that make the AI status visible to listeners. Whether and how we do so is reviewed against the Principles and the user-facing experience commitments; we will publish the operating rules at the point any such measure is introduced.

3. Voice, likeness, and identity

Generation, replication, or imitation of an existing person's voice, name, likeness, or distinctive performance style without that person's express permission is prohibited on Talisman. This applies whether the AI output appears as music, dialogue, vocal sample, image, video, or any other media on the Service.

The reasoning is jurisdictional and forward-looking. As of the date of this Policy:

  • In Tennessee, USA, the Ensuring Likeness, Voice, and Image Security Act (the ELVIS Act, in force from 1 July 2024) makes the unauthorised commercial use of a person's voice (including AI-simulated voices) actionable under state right-of-publicity law.
  • In the United States federally, the NO FAKES Act has been reintroduced (2025) with broad cross-industry support but is not yet enacted as of the date of this Policy. Our commitment under this section is structured to comply with the Act's notice-and-takedown framework on enactment, without waiting for the law to land.
  • In the European Union, the AI Act includes synthetic-content disclosure obligations (Article 50) phased in through 2 August 2026; our disclosure pipeline under section 2.1 is designed to meet those obligations as they apply to relevant content.
  • In the United Kingdom, no specific statute on voice/likeness AI exists as of the date of this Policy, but rights of publicity, passing off, defamation, and data-protection law (where the impersonation involves processing of personal data) provide enforceable causes of action. Our commitment is that we honour them as if NO FAKES-equivalent UK provisions were already in force.

Anyone whose voice, name, likeness, or distinctive performance style has been used in AI-generated Content on Talisman without permission may submit a takedown request under section 4 of the Notice & Takedown Policy. We will act on credible requests within the same timelines that apply to copyright takedowns.

4. Detection and verification

Talisman runs an in-house AI-content verification pipeline (audio, image, and video). The operational properties are:

  • Probabilistic. Detection produces a confidence score, not a verdict. Even high-performing detectors can be defeated by adversarial post-processing, codec re-encoding, or model evolution.
  • Advisory. Detection results inform human moderation. They do not, on their own, remove Content or close accounts. Final decisions are made by humans, and decisions that affect publication or account standing are appealable.
  • Auditable. The detection pipeline is logged, with the same audit-and-review commitments that apply to other moderation processes under the Transparency commitment.

Where Content carries cryptographic provenance signals (C2PA Content Credentials, watermark schemes such as SynthID), we read and surface them. Where Content carries Talisman's own Temporal Fingerprint (a quantum-resistant cryptographic timestamp generated at upload, signed with ML-DSA-44, anchored to consensus state), we use it to support provenance claims and to make ownership and timing of an upload verifiable independently of any detector.

We do not treat absence of provenance signals as proof of AI generation. We do not treat presence of provenance signals as conclusive proof of human origin.

5. Enforcement

A breach of this Policy may result in any combination of:

  • Removal of the Content
  • Tagging the Content with an AI-status indicator
  • Exclusion from editorial and algorithmic surfaces
  • A formal warning logged against the account
  • Account suspension or termination under the Terms of Use section 17
  • Reversal of any payouts attributable to the breach
  • Referral to law enforcement or to a regulator where there is a credible legal infringement
  • Civil action where appropriate

Repeated breaches will be treated under the repeat-infringer policy in the Notice & Takedown Policy.

We commit to reasoned decisions. Where we remove Content, suspend an account, or apply a tag, we will state the rule that was breached and the basis for our finding, and we will provide an appeal channel.

6. Coalition posture

Talisman publishes detection signals and provenance assertions in a form that other platforms can verify and act on. Our intention is to operate as part of a federated, cross-platform protective layer for the music industry rather than a closed silo. We will publish the public-facing API and the participation conditions when the federation interface is launched.

This is the practical sense in which Talisman is more than a destination platform: a digital safeguard, not a label, that helps the broader ecosystem hold the line against unauthorised training, unauthorised cloning, and undisclosed AI-generated content.

7. Updates

This Policy is a living document. We update it when:

  • New legal duties arrive (Online Safety Act guidance, EU AI Act implementing acts, NO FAKES Act enactment, ICO codes of practice, court decisions in the United Kingdom or the European Union).
  • New industry standards stabilise (DDEX AI-content disclosure, C2PA assertions, ai.txt, TDMRep).
  • A real near-miss in our own operations exposes a clause that needs to be tightened.

We do not update this Policy to relax existing protections under commercial pressure. The protections here are the reason artists upload.

The "updated" date in the document header reflects the date of the most recent change. Material changes will be notified in-app or by email at least 14 days before they take effect, except where the change is required to address a security or legal risk.

Drafted in pre-incorporation form, 28 April 2026, for review and finalisation by qualified UK counsel. Statutory references current to the date of drafting; counsel to confirm the status of NO FAKES Act, EU AI Act implementing acts, and any UK-specific developments at sign-off.

Last updated 28 April 2026