We sell synthetic voices — and label every one. The deepfake economy is why

We are a company that sells synthetic voices. Blab, our voice studio, clones a voice only as a deliberate, consented act and ships disclosure with the output; VidiVidi runs an openly synthetic broadcast cast that is labelled synthetic on every screen, at the door and on every character card; Harbi reads its news bulletins in a machine voice and says so. We built it that way before it was fashionable, and this month the news cycle wrote the argument for us in other people’s losses. identifAI’s inaugural fraud report, published today, prices deepfake-enabled scams at some $40 billion a year, with video synthesis and voice cloning together accounting for roughly two-thirds of observed cases. The question of 2026 is no longer whether a voice can be faked — anyone’s can, cheaply. It is who you can still believe. Our answer is structural rather than hopeful: synthesis you can trace — consent at the clone, a label at the source, evidence under the claim.

One week of the deepfake economy

Take the first week of October as a core sample. On the 7th, a California mother — a Navy veteran — told a US Senate committee how fraudsters used an AI clone of her daughter’s voice, pitched mid-panic, to stage a kidnapping: five and a half hours on the phone, four wire transfers, $5,400 gone. On the 5th, Italy’s prime minister filed an EU trademark application for her own voice — a head of government forced to treat her larynx as intellectual property. The same day, Sony Music disclosed it had filed more than 260,000 deepfake takedown requests, nearly double its March total. On the 6th, a Paris hospital group and Sorbonne University filed a criminal complaint over deepfakes of a physician and a news anchor pushing fraudulent remedies, while an Australian party ran an AI-generated political ad with no viewer-level disclosure. Behind the anecdotes, the identifAI telemetry: video synthesis at 42.7% of synthetic-fraud incidents, voice cloning at 21.9%, synthetic media implicated in one in every hundred identity-verification failures globally — and the canonical enterprise case still the Arup video call, $25.6 million wired to a meeting of deepfakes. The response in the identity industry is telling: verification is moving into the live conversation itself, because the recorded artefact can no longer be trusted on its own.

The liar’s dividend is the real bill

The direct fraud is the visible cost. The deeper one is what legal scholars named the liar’s dividend: once everyone knows any recording can be faked, every real recording becomes deniable — the politician’s genuine hot-mic moment, the CEO’s actual promise, the authentic evidence, all waved away as “probably AI”. That is a tax on truth itself, and it compounds in a public square where, as we have written before, around four in ten people already avoid the news and trust sits near its lows. The instinctive answer — better detectors — loses the arms race, for reasons we laid out in our post on detector limits: detection is probabilistic, adversarial and always a generation behind, and the one survey-proof result is that human ears do not reliably improve. What survives scrutiny is the boring, structural stuff: provenance attached at creation rather than inferred at consumption; consent required at cloning rather than apologised for after; claims bound to evidence rather than to vibes. Labelling what is synthetic does not catch criminals — it keeps the honest layer of the information economy legible, so that the liar’s dividend has somewhere firm to stop.

Where we stand: synthesis you can trace

Our position is unusual and worth stating plainly: we are on the synthesis side of this economy, and everything we ship is built so that our output cannot feed the fraud column. Blab treats voice cloning as a deliberate opt-in — a consented act by the voice’s owner, not a checkbox — and its voices ship with disclosure built in; that norm stopped being merely ours in August 2025, when Article 50 of the EU’s AI Act made labelling synthetic content law in Europe. VidiVidi goes further than any law asks: a permanent cast that is synthetic and says so on every surface, a rule that a prepared broadcast is never sold as “live”, and retraction that closes everywhere — audio, transcript, clips. And Harbi attacks the dividend from the other side: when a labelled machine voice reads the bulletin, every factual sentence underneath it is bound to its sources, because the ultimate answer to “anything could be fake” is a claim that carries its receipt. Same stack, same conviction, three surfaces: synthesis is legitimate exactly as far as it is traceable.

• Consent at the clone: a voice is copied as a deliberate, owner-approved act — Blab
• A label at the source: synthetic audio ships disclosed, not discovered — Blab, Harbi’s bulletins
• A cast that says so: synthetic hosts labelled on every screen, “live” never faked — VidiVidi
• Evidence under the claim: sentences bound to sources, corrections versioned in the open — Harbi
• Alignment with the EU AI Act’s Article 50 labelling regime — in force since August 2025

What we will not claim

Three honest limits. First, labels bind the honest. The Matryoshka networks of the world will not disclose their fakes, and a watermark does not stop a wire transfer; what stops the kidnapping scam is slower, human stuff — callback rituals, code words, live verification, prosecution. Disclosure protects audiences from being misled by us and keeps the legitimate market legible; it is a floor, not a shield. Second, the $40 billion figure deserves its caveat in the same breath: it comes from a single vendor’s inaugural report, with no independent verification cited — read it as the scale of the problem, not a census of it. Third, we are not neutral observers. We sell voice synthesis; our commercial interest is a world in which labelled, consented synthesis stays legal and trusted while the fraud column gets squeezed — we think that interest aligns with everyone who still wants to believe their own ears, and we state it openly rather than pretending to referee.

Detection loses the arms race; denial collects the dividend. What survives is provenance: consent at the clone, a label at the source, evidence under the claim.

For readers in Türkiye the ground is familiar: voice-clone scam calls have long since arrived here too, and a person’s voice and likeness sit at the very heart of what the KVKK protects — an unconsented clone is a legal violation before it is an ethical debate. This month showed the ceiling: a prime minister trademarking her own voice because the law of identity is losing a footrace with the economics of imitation. Between floor and ceiling sits product design, which is where we work. Blab clones with consent and discloses by default, VidiVidi’s cast never pretends, Harbi’s bulletins carry receipts — all on our own models through our own gateway, from Arpanet Bilişim A.Ş., because a provenance promise is only as good as the stack that enforces it. The deepfake economy runs on voices nobody agreed to give. Ours are spoken for.