The first major social response to generative A.I. unfolded like a detective story. Schools hunted for software that could flag model-written essays, publishers tested classifiers for synthetic prose and platforms floated watermarks and “A.I.-generated” labels, as if the next era of trust would be won by progressively sophisticated forensic tools.
A.I. detection systems, however, are currently straining in two ways: they still produce both false positives and false negatives. Human-written work is regularly flagged as synthetic, while lightly edited A.I.-generated text often passes unnoticed. What’s worse is that tools that work reasonably well in one context can fail in another, whether because of differences in language, style, model architecture or prompting techniques. In practice, detection already resembles an arms race, and arms races rarely end with one side declaring permanent victory.
That sounds alarming, but historically it is also familiar. When detection proves elusive, societies rarely solve the crisis by endlessly refining forensic tools. Instead, they build systems of accountability through authorship conventions, editorial oversight, provenance records, professional standards and legal liability, so that trust shifts from identifying origins to establishing responsibility.
To put it more visually, it’s a shift from fingerprints to passports. Fingerprints imagine authenticity as something embedded within the object that can be uncovered through closer inspection. Passports establish legitimacy through institutions, records and verification systems.
Why detection keeps failing
Detection fails not because experts are foolish or expert systems are flawed, but because imitation improves faster than exposure and because authenticity is a social judgment as much as a technical one.
Art history has many examples for us to consider. Han van Meegeren’s forged Vermeers did not merely fool casual viewers. They convinced serious authorities, only for connoisseurship-endorsed works to later be revealed as fakes. The issue isn’t merely that experts were fooled; it was that looking was never enough. Systems based on authenticity, inferred from style, story and desire, become guidance for forgers to manipulate masterfully.
The pattern recurs in Thomas Chatterton’s so-called “Rowley” poems, which were presented as medieval texts and welcomed as discoveries. Their reception depended on evidence as well as longing: a lost voice recovered, a past made present. The lesson here is that authenticity is not simply embedded in the work but is upheld by institutions and incentives.
Generative A.I. scales that instability. If a model can draft a plausible essay in seconds and a human can polish it in minutes, the text becomes an unreliable witness to its origin: the “tells” mutate, and the cost of imitation collapses.
When trust moves: provenance as the new center
Counterfeit currency offers a useful comparison. Modern economies do not rely on citizens to spot fake notes. They rely on central banks, controlled issuance, security features, serial numbers, audit trails, regulated intermediaries and enforcement mechanisms. A banknote is trusted not because every person can detect a forgery, but because a system makes accountability possible. This is passport logic at scale, in which verification moves away from individual perception and into institutional infrastructure.
The art world learned the same lesson. After forgeries and contested attributions, markets shifted emphasis away from “better eyes” and toward provenance: ownership histories, archival documentation, conservation records and chain-of-custody verification. Over time, paper trails often became more valuable than connoisseurship.
A.I. is pushing trust in the same direction. Durable trust will come less from stylistic residue than from provenance systems, including signed metadata, cryptographic attestations and standards that let platforms and publishers confirm where content came from and what happened to it. Initiatives like C2PA and content credentials are early media passports. Their goal is less “detect the output” than “verify the chain.”
The question, therefore, changes from “Can this be proven human-made?” to “Who issued this, who handled it, and who is accountable for it?”
Authorship is an accountability technology
Long before generative A.I., societies learned to live with fuzzy origins because authorship has never been only a claim about who produced every word. It has always operated as a practical mechanism for responsibility.
Consider ghostwriting. Political memoirs, executive books and celebrity publications often involve uncredited writers and heavy editorial shaping. Yet when a book contains an inaccuracy, the public rarely treats the ghostwriter as the accountable party. Responsibility attaches to the name on the cover: the byline functions less as a forensic statement than as a liability anchor.
The Shakespeare authorship debates make the same point at a higher altitude. We may never settle how collaborative the plays were or how much can be attributed to Shakespeare alone. But that’s not the point. The works remain meaningful because societies can function despite imperfect knowledge of origins, so long as institutions stabilize responsibility and value.
Michel Foucault gave this mechanism a name: the author-function. Authorship, in his account, is a social role that classifies texts, governs their circulation and enables systems of ownership and punishment; it tells societies whom to praise, whom to pay and whom to sue. It’s a point also made by Roland Barthes, who challenged the author as the sovereign source of meaning when he wrote that “The birth of the reader must be at the cost of the death of the Author.”
A.I. applies pressure elsewhere, destabilizing the author as the node of accountability. If words can be generated, remixed and distributed at scale, the urgent question is governance: who is responsible for synthetic persuasion, defamation and expertise? In the age of synthetic generation, bylines, editorial practices and publication standards are not ceremony but rather infrastructure for trust.
(Re)production didn’t end trust. It changed its scaffolding.
Authenticity anxiety is not uniquely digital, and the history of production (and reproduction) shows that trust usually survives by changing its scaffolding. When we return to the examples of art, works from medieval and Renaissance workshops complicate our modern obsession with singular originality. In studios associated with Raphael, Rubens or Titian, for example, multiple hands could contribute to a painting. Patrons bought a workshop’s reputation and a master’s supervision as much as a specific brushstroke. Attribution persisted because accountability was organized through commissions, contracts, standards and reputational guarantees.
It’s a lesson we had to address with the rise of mechanical reproduction through photography. When the first daguerreotypes appeared, the images were treated as mechanical truth. But this truth was short-lived as manipulation quickly arrived through retouching, staging and darkroom edits. Journalism with photography did not make every reader a forensic analyst; it built guardrails through editorial review, sourcing discipline, ethics codes, corrections practices and reputational penalties. We learned to trust a process, not an image.
Walter Benjamin captured the shift when he wrote that “that which withers in the age of mechanical reproduction is the aura of the work of art.” Reproduction changes what authenticity means and where value lives. Generative A.I. repositions that dynamic from reproduction to production itself. When plausible creation becomes cheap, “aura” becomes harder to locate inside the artifact, so trust reattaches to what remains legible: who published, under what standards, with what disclosures and with what liability.
Borges and the collapse of origin as a detectable property
Jorge Luis Borges’s short story Pierre Menard, Author of the Quixote, has become a fitting parable for authenticity and meaning in this A.I. era. The story recounts how Menard reproduces portions of Cervantes word-for-word, while the narrator nevertheless insists that the second text is “almost infinitely richer” because identical words mean differently when authored in another era by another person. Borges’s point is devastating for the authenticity-detector fantasy because the text itself cannot reliably reveal its origin; origin is external, supplied by attribution, context and institutional framing.
That is the A.I. problem in miniature. Two passages can be identical: one drafted by a student, one generated by a model and edited by a human. If the artifact cannot carry a stable internal signal of provenance, staring harder at it will not solve the trust crisis; the solution has to be built around the artifact.
The institutional future after detection
If detection remains unreliable, the most consequential change will not simply be that “fakes” flood the zone, though they will. The deeper change will be institutional, a migration toward accountability systems that make trust operable even when origins are ambiguous.
We should therefore expect more emphasis on provenance infrastructure, editorial accountability and liability regimes that clarify who bears risk when synthetic content harms. This is not utopian, since passports can be forged and institutions can fail. Rather, it should be recognized that systematic approaches scale trust better than improving individual perception.
The future of trust in the age of generative A.I. may depend less on identifying machine-generated content than on identifying who stands behind it: because when authenticity becomes hard to see, accountability becomes more valuable than detection.
