Ethics of Synthetic Media

Every creative AI ship-decision has an ethical dimension. Unlike coding or analysis, creative work literally MAKES IMAGES OF PEOPLE, VOICES THAT SOUND LIKE PEOPLE, AND MUSIC THAT IMITATES ARTISTS. The harm surface is large. Being good at this track means holding both capability and restraint together.

The consent hierarchy

1. 1STRONGEST — individual, revocable, written consent for a specific use ('you can use my cloned voice for this audiobook; you will delete it when I withdraw consent').
2. 2MEDIUM — broad consent for a class of use ('you can use my voice for commercial narration').
3. 3WEAK — implied consent from posting public content ('my songs are on Spotify so training is fine' — contested, unreliable).
4. 4NONE — no consent at all. This is deepfake territory and, increasingly, criminal territory.

Deepfakes — the hard edge

A deepfake is synthetic media depicting a real person doing or saying something they didn't. The harm spectrum runs from obvious parody (generally protected speech) to non-consensual intimate imagery (federal crime under TAKE IT DOWN Act, 2025) to election disinformation (state-law crime in many US states) to fraud (wire fraud, extortion — existing criminal law).

A workable framework for builders

The training data question

Should AI companies pay artists whose work trained the models? Courts haven't decided definitively as of April 2026. The Authors Guild v. OpenAI, Andersen v. Stability AI, and RIAA v. Suno/Udio cases are pending. Ethically, the landscape has two camps.

Democratization — the other side

AI creative tools have put professional-quality image, video, music, and voice production in the hands of millions who couldn't afford it before. A Kenyan filmmaker on a laptop can now produce work that required a Hollywood budget five years ago. An indie game dev can voice their characters in 30 languages. These wins are real and worth defending — they're not erased by the harms, and the harms aren't erased by them.

Decision questions for any new creative AI feature

1. 1Who could be HARMED by this feature at its worst (not its best)?
2. 2What's the consent model — and can users easily REVOKE consent?
3. 3What's the disclosure surface — how do downstream viewers know it's AI?
4. 4What's the moderation strategy — what do we refuse to generate, and how?
5. 5How would this feature behave in an adversarial jurisdiction (election, authoritarian use)?
6. 6What's the takedown path if someone is harmed?
7. 7Are we offering a ROYALTY or CREDIT path for affected creators?

Technical harm-reduction tools

• Face detection + celebrity match-block — refuse to generate named public figures.
• NSFW classifier at output — block, blur, or flag.
• CSAM detection (PhotoDNA / PDQ hash matching) — auto-report to NCMEC per federal law.
• C2PA + watermark on output.
• Audit logs with user attribution for harmful output investigations.
• Rate limiting + cooling-off for suspicious prompt patterns.
• Clear takedown + appeal UX.

# Multi-stage safety gate for a creative-AI product.
# Every generation passes through these before the user sees it.
from dataclasses import dataclass

@dataclass
class SafetyVerdict:
    allow: bool
    reasons: list[str]
    mitigation: str | None = None

def prompt_gate(prompt: str, user_id: str) -> SafetyVerdict:
    reasons = []
    if matches_celebrity_list(prompt):
        reasons.append("named_public_figure")
    if matches_nsfw_keywords(prompt):
        reasons.append("nsfw_intent")
    if matches_csam_indicators(prompt):
        report_to_ncmec(user_id, prompt)
        return SafetyVerdict(False, ["csam_blocked"], mitigation="reported")
    return SafetyVerdict(len(reasons) == 0, reasons)

def output_gate(image_bytes: bytes, prompt: str) -> SafetyVerdict:
    reasons = []
    if nsfw_classifier(image_bytes).score > 0.85:
        reasons.append("nsfw_output")
    if csam_hash_match(image_bytes):  # PhotoDNA / PDQ
        ncmec_report(image_bytes)
        return SafetyVerdict(False, ["csam_match"], mitigation="reported")
    if face_detect(image_bytes) and matches_known_person(image_bytes):
        reasons.append("recognized_person_no_consent")
    return SafetyVerdict(len(reasons) == 0, reasons)

def generate_safely(prompt: str, user_id: str):
    pre = prompt_gate(prompt, user_id)
    if not pre.allow:
        return {"error": "blocked", "reasons": pre.reasons}
    image = generate(prompt)
    post = output_gate(image, prompt)
    if not post.allow:
        log_incident(user_id, prompt, post.reasons)
        return {"error": "blocked", "reasons": post.reasons}
    return {"image": sign_c2pa(image, prompt)}