Five voices, five journeys
Each voice takes its own hypothetical route through three real messaging apps and ends with a different receipt. The apps’ roles are illustrative, not test results. Listen first, then send it. On a phone they all look the same; underneath, three are AI and two are real people.
The journey
Relay log (public)
The person whose voice it is
Trust list
Public drop register
Counts come from your runs of this simulation, not from real app behaviour. Our tool has no power of its own. It names who could act. Each time an app is caught dropping a record silently, the consequence steps up (Ostrom’s graduated sanctions).
| App | Based in | Carried | Passed on | Receipts | Caught | Consequence |
|---|
Where it fails
A prototype that cannot fail proves nothing. These three cases are not fixed by our rule.
A generator outside the EU
It signs nothing. The rule makes the missing record meaningful, but cannot create a record that was never made.
Two apps cheat together
Both write nothing. The next app sees the gap but cannot say which one dropped the record, so nobody can be sanctioned.
A signer nobody trusts
The record survives, but the phone can only say “present”. Whoever controls the trust list decides what counts as proof.
Limits we state openly
What it cannot do
- Prove who is really speaking.
- Force real apps to join.
- Decide who pays for receipts.
- Decide who is on the trust list.
What it could break
- The log shows which apps a message passed through: a privacy cost.
- Whoever runs the trust list gains quiet power.
- People may still read “not observed” as “fake”.
Who gains, who pays
- Listeners gain a reason behind every missing record.
- Apps pay to keep records and sign receipts.
- Honest apps gain: cheaters are named, not them.
- AI providers already mark under Art. 50; this keeps the mark alive.
Our real-world test
The journeys above are hypothetical: they show what each app could do under our rule, not what it does today. To find out what really happens, we sign a clip with C2PA’s open-source tools, send it, and check what arrives. Until a cell is tested, it says so.
| App | Sent as voice note | Sent as file | Shown to the listener? |
|---|
Never the word “fake”. The receipt reports what happened to the record.
A missing record says “not observed”, never a blank that looks fine.
The next app checks the last one; the monitor never depends on the monitored.
The tool has no authority. It names who could act on the evidence.
A compression default and a trust list are both governance choices.
The voice owner gets standing: they can dispute what is said in their voice.