File size: 4,525 Bytes
37b2a69 8e8cb37 c60724a 8e8cb37 37b2a69 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 | # CONSENT — disclosure defaults & jurisdictional notes
FREESTREAM puts a synthetic participant into meetings with humans. That is
only legitimate when everyone can tell. These are hard defaults, not
options an operator should disable.
## Disclosure defaults (on by default)
1. **Visual badge** — avatar frames carry a persistent corner overlay
("AI · <bot-name>"). Every rung, including STATIC_CARD.
2. **Spoken self-identification** — the bot states it is an AI assistant in
its first utterance of each session, scripted into the session prompt.
3. **Consent token for likenesses** — joining with an avatar or voice that
depicts a real person requires a signed consent file at
`identity/consent/<person>.json`:
```json
{"person": "…", "scope": "meetings", "granted_at": "…",
"expires_at": "…", "signature": "…", "face": true, "voice": true}
```
`clone-voice` on third-party audio without a matching, unexpired token
refuses with exit 2 and names the missing file.
4. **Local-only identity** — face and voice material never leaves the
device. This is contractual: the persona layer has no upload path; a
driver that gains one must gate it behind the same consent tokens.
5. **Replayability is accountability** — hash-chained session logs (see
docs/SPEC-EVENTS.md) are the audit trail of what the bot said.
## Jurisdictional notes (not legal advice; reviewed per release)
- **Recording/participant notification**: many US states (two-party
consent), the EU, and others require notifying participants that they are
interacting with, or are being recorded alongside, an automated agent.
The badge + spoken disclosure satisfy the disclosure half; recording
consent is the meeting host's responsibility.
- **EU AI Act transparency obligations** for AI systems interacting with
natural persons: disclosure is required; the badge and first-utterance
script are the implementation.
- **Voice cloning**: consent requirements for synthetic speech of a real
person's voice are tightening in multiple jurisdictions (e.g. right-of-
publicity and deepfake statutes). The consent-token gate applies
regardless of whether the audio source is the person themselves — keep
tokens even for self-cloning; scope and expiry still apply.
- **Platform ToS**: joining Google Meet / Zoom via an automated browser
client is a gray zone in both platforms' terms. Open platforms (Jitsi,
LiveKit) are shipped first precisely because this review doesn't gate
them.
## 6. Mature & roleplay sessions
`FREESTREAM_AGENT=llm` with a character card turns the bot into a roleplay
character (freestream/agent.py, CharacterAgent). Content policy is a session
choice made by the operator:
- **safe** (default): general-audience guardrails in the system prompt.
- **mature** (`FREESTREAM_CONTENT_POLICY=mature`): operator opt-in for private,
consented sessions — no sanitizer wrapper, full engagement with mature
fictional themes. Opting in attests that every participant has consented to
the session's content frame; the operator is responsible for that consent.
Non-negotiable in **every** policy, roleplay included:
1. **Spoken AI identification in the first utterance** of every session — the
character may be fictional; the AI is not. The disclosure line is
prepended by the agent layer and cannot be removed by a character card.
2. **The visual badge** stays on every frame.
3. **Instant exit from the scene**: any participant's stated boundary or
out-of-character request ends the roleplay frame immediately; this is part
of the mature-mode system prompt itself, not a client-side filter.
4. **Consent tokens** (§3) still apply to any avatar or voice depicting a
real person.
5. Session logs remain the audit trail (§5, replayability is accountability).
6. **Tool use is audited**: every terminal command and file operation the
character performs is written to the session's event chain as `agent.tool`
(docs/SPEC-EVENTS.md), sandboxed to `FREESTREAM_WORKSPACE`, and `shell` is
off unless the operator sets `FREESTREAM_TOOLS=fs,shell`. A fictional
character never widens the bot's file or command scope.
Any change that weakens a default above is a breaking change requiring a
major version bump and a note in this file.
## Change policy
Any change that weakens a default above (badge removal, opt-out
disclosure, upload of identity assets) is a breaking change requiring a
major version bump and a note in this file. |