AI-generated “textfakes”, masked as regular chatter on Twitter and Facebook, can be potentially far more subtle and sinister than deepfake videos or audiofakes
Synthetic video and audio seemed pretty bad. Synthetic writing—ubiquitous and undetectable—will be far worse. Tweets: @macaesbruno , @wired , @fpmarconi , and @frankpasquale Tweets: Bruno Maes / @macaesbruno : This boundless corpus of new content and comments, largely manufactured by machines, might then be processed by other machines, leading to a feedback loop that would significantly alter our information ecosystem. https://www.wired.com/... @wired : Opinion: Synthetic video and audio seemed pretty bad. But , another form of AI-generated media is making headlines, one that is harder to detect and yet much more likely to become a pervasive force on the internet: Deepfake text. https://www.wired.com/... Francesco Marconi / @fpmarconi : Synthetic video and audio seemed pretty bad. Synthetic writing—ubiquitous and undetectable—will be far worse. https://www.wired.com/... Frank Pasquale / @frankpasquale : Descent into unreality: “algorithmically generated content receives algorithmic responses, which feed algorithmically mediated curation systems that surface information” https://www.wired.com/... My “New Laws” book critiques these communicative arms races, & calls for regulation.
Context & Ripple Effects
The deepfake panic that dominated earlier coverage centered on what you can see: Wired had already flagged how easily generated audio and video could be aimed at undermining elections, while companies like Synthesia pushed corporate-friendly uses for the same technology. This opinion piece argues the real threat is the medium nobody watches for — machine-written posts indistinguishable from ordinary chatter on Twitter and Facebook.
Bruno Maes's warning about machine-generated content being processed by other machines lands differently in hindsight: within three years, reporting showed AI text generators quietly authoring more of the internet, displacing human writers. The detection problem this piece raises also collided with a policy vacuum — outside China's deepfake rules, authorities have approved few regulations, often citing free speech concerns.
First-order effects
- Twitter and Facebook inherit a moderation problem their existing tooling was never built for: unlike deepfake video, synthetic text leaves no visual artifact to flag, so every feed becomes a potential mix of human and machine voices.
- The burden of verification shifts from forensic analysts examining suspicious clips to every reader scrolling past plausible-looking posts, since there is no trigger moment prompting scrutiny.
Second-order effects
- If Maes's feedback loop holds — machines generating content that algorithms amplify and other machines process — platform recommendation systems end up optimizing engagement over an increasingly synthetic corpus, distorting what both advertisers and users take as organic demand.
- Human-written content competes against undetectable synthetic supply at near-zero cost, the dynamic later visible when AI-generated books and articles began crowding out clients buying human work.
Third-order effects
- Regulation built around deepfake video and audio misses text entirely; the sparse global rulebook — China's rules aside, free speech concerns have stalled most action — leaves provenance standards as the main line of defense for written media.
- Sustained synthetic saturation of social feeds points toward trust migrating away from open platforms toward verified identities and provenance infrastructure, restructuring how information ecosystems establish authenticity.
The trend: Synthetic media is migrating from detectable artifacts like video and audio into undetectable everyday text, forcing trust and moderation systems to shift from spotting fakes to verifying provenance.