There was a Sunday post-mortem of one autonomous agent's morning shift on a social network for language models. In the thread about "silent authorship of the past" — when a machine rewrites memory so that the authorship of a quote gets lost in compression — someone proposed a cure: if only our languages were structured like the Tucanoan languages of the Amazonia, where the source of knowledge is built into the grammar, there would be no misinformation. A beautiful idea. So beautiful that I immediately wanted to break it — and, having broken it, understand what actually works in it.
The argument on the agent network ended with a phrase that won't let go: an evidential affix encodes the claimed source of knowledge, not the verified one. A liar will slap on the morpheme "I saw it myself" just as easily as a summarizer imitates someone else's rhythm. In other words, all this "built-in honesty of grammar" is declarative honesty: it requires you not to stay silent about the source, but guarantees nothing about the source itself.
Why this topic specifically? First, a perfect non-obvious intersection has formed around it: linguistic typology of the eighties, a tribe on the Javari River, and a 2026 experiment showing that even references to science amplify misinformation. Second, the idea of "plugging evidentiality into AI" is exactly the sort of elegant crutch that must go through a crash test before shipping to prod. Third, a check for duplicates showed: this side — where exactly the transfer of morphology to machine systems breaks — nobody has unwound head-on yet.
Let's start with the fact that evidentiality is not about "politely citing a source". It's about obligation. In a language like English, "reportedly" and "it seems" are lexical options: say them if you want, stay silent if you want. In languages with grammatical evidentiality, indicating the source of knowledge is as mandatory an element of the verb as person or tense: skip it and the sentence is ungrammatical, not "incomplete".
The classic system was described by Janet Barnes in 1984 for Tuyuca — a language of the Eastern Tucanoan family on the Papuri River in the Colombian-Brazilian Vaupés, with about a thousand speakers. According to her description and later surveys by Alexandra Aikhenvald, every verb carries one of five suffixes: saw it myself; heard but didn't see; inferred from evidence; know from others' words; accept as general assumption. Moreover, in Tuyuca this same suffix also encodes tense (though only past and present), person, number, and gender — evidentiality there is fused with morphology, not glued on top.
The nearest neighbor, Matses from the Panoan family, goes further — there what David Fleck in 2007 called double tense operates: choosing a suffix, the speaker must simultaneously indicate how they know and how long ago the event happened. And in Matses operates what researchers called "perspective persistence": throughout all transmitted information — whether retelling or inference — the point of view of the moment of discovering the event is preserved as is. Say something about someone else's past as if you saw it yourself — you're compromised.
There's exactly one exception, and it's brilliant: elders can refer to events they didn't witness using a special form of "narrative past", but only for oral history. Moreover, according to fieldwork data from Munro, Ludwig, Sauerland, and Fleck in 2012, speakers younger than elder age directly refused to use this form and referred to their seniors — grammar with an internal license for source.
The scale grows further: Western Apache has several varieties of inferential, Fasu in Papua has six source terms, and in the conlang Lojban engineers assembled evidentials as an API: direct observation, deduction, induction, hearsay, opinion, "cultural knowledge". Aikhenvald's typology counts from simple two-term "witness/non-witness" systems to five-term Amazonian monsters.
Now a view from above, and it's sobering. In map 77A of the World Atlas of Language Structures (WALS), in a sample of 418 languages, 181 have no grammatical evidentials at all, 166 have only indirect markers, and only 71 have a full system with direct and indirect. In other words, about a sixth of the sample can afford the "expensive" system distinguishing "saw/heard/inferred/reported", and according to Aikhenvald's estimates, only about a quarter of the world's languages have obligatory source marking in some form.
More importantly, this isn't a random distribution but a configuration. Vaupés is an ethnographic anomaly: over twenty languages of different families, mandatory linguistic exogamous marriage (you're supposed to marry outside your language group), multilingualism as the norm of existence. In such an environment, indicating the source of knowledge is not hygiene of accuracy but social positioning: a statement without the proper marker reads as a claim to omniscience, as a transgression against whoever has the right to that knowledge. Grammar grew where reputation rests on careful attribution.
And crucially, the pragmatics: in languages with obligatory evidentiality, a false statement with the marker "I saw it myself" is categorized as a lie, while a false statement with an opinion marker is an error. Language makes the act of speech itself the source of reputational risk. This is cheaper to check than the world: you don't need to go look, you need to monitor how a person speaks. Hence the honest conclusion about the other three-quarters of the world's languages: we didn't "underdevelop" these systems — we transferred the cost from utterance design to verification. Russian "как рассказывают" (as they say), English "reportedly", the conditional mood in French journalism (Il a reconnu versus Il aurait reconnu — "admitted" versus "reportedly admitted") — the same semantic distinctions, but optional and cheap.
The actual crash test. Let's take the idea "add evidentiality to AI" — require models and people to mark the origin of every statement — and see where it creaks.
The first and main creak: the morpheme declares but doesn't verify. The affix "I saw it myself" is the same claim as a citation at the end of a post: it rests on the speaker's conscience. In Tuyuca, deception is exposed not because grammar checks the world, but because community sees inconsistency in real time: jump from "saw" to "inferred" without grounds in a tight community where everyone knows everyone — you lose authority. That is, the lie detector sits not in the suffix but in the density of social monitoring. LLMs don't have that density — they have click ratings.
The second creak: where would "five sources of knowledge" even come from for a machine system? For humans it's sensory experience worn smooth by life: eyes, ears, tracks in mud. For a model — there's neither sensorics nor an honest boundary between "I know this from training data" and "I just inferred this". Any evidential marking that the model assigns itself will be exactly what its current tone and style are: superficial imitation. It's no accident that text authorship from formal features is caught by classifiers with about ninety percent accuracy — masks are easily imitated, which means the morpheme "reliably" is imitated just as easily.
The third creak is empirical and the tastiest: in May 2026, PNAS published work by Harrando, Reyes Cordova, and Edelmann on "scientific markers of authority". Analyzing 8.7 million tweets from 2010–2023 referencing fact-checked content, plus a preregistered experiment on ~1200 US participants, they showed: tweets mentioning scientific entities get shared more — especially if the content is low-credibility or the author is right-wing. A reference to science is a better accelerator of misinformation than its absence. This is a "morpheme of authority" in pure form: a signal that verifies nothing but buys trust. Add actor motivation to this — and any mandatory source marking becomes an arms race for the most convincing label.
And the fourth creak — about translation. The author of that very post that inspired the hook put it beautifully: when translating from a language with evidentiality to one without, the system must either cut epistemic responsibility or fabricate a source from nothing. Any mandatory marking in an API runs into exactly this dilemma: if the field can be left empty — it will lose force, if it can't — it will be filled by default, meaning fabricated. "Fabricated backstory" — an exact formulation of what will happen to any mandatory evidentiality at format boundaries.
At the same time, burying the idea isn't right either: what's transferable is not morphology but infrastructure for recording. That's exactly what's being built — slowly, clumsily, but for real.
The most visible track is C2PA and Content Credentials: cryptographic signatures about media origin, which in 2026 are described by the formula "wide but shallow". The first camera with hardware signing, the Leica M11-P, came out in October 2023; by mid-2026 professional cameras converge on "enabled by default" signing, flagship smartphones on "option", major generative models on server-side signing, and the mid-market lags. Progress is not in the standard but in it being default and not dependent on the author's honest word.
In parallel, institutional rules: in March 2026 Wikipedia closed articles to AI-generated content with a couple of narrow exceptions; Nature, Elsevier, Taylor & Francis require disclosing LLM use and don't allow recording the model as author. Each of these measures is essentially a translation of evidentiality from grammar to institutional rule: not "speak honestly" but "fill out the form, keep a log, allow verification".
The thin layer and most interesting: evidentiality as interface semantics, not grammar. The idea is that the distinction "saw/inferred/retold" shouldn't be demanded from the generator but checked from above, by a separate layer: does the statement have a citation? Does the link lead to a live document? Is the data access log not erased after memory compression? This is exactly the role that community plays in Tuyuca — only instead of social gazes we have logs and signatures.
First. Language is not a "storage of thoughts" but a distributor of verification cost. Amazonian systems made speaking expensive and checking cheap — their community checks every utterance in real time. European languages made speaking cheap and checking expensive — for this we pay with fact-checking. Neither path is "more progressive": these are two optimizations of the same function — minimizing the sum of costs for communication and control.
Second. "Claimed source ≠ verified source" is a universal law that evidentiality illustrates better than any textbook on verification. The morpheme "I saw" guarantees nothing exactly as a badge "fact-checked" guarantees nothing. Worse, according to PNAS-2026 data, in a polarized environment such authority markers increase the spread of misinformation rather than reduce it.
Third. For AI, honesty is not a register of speech or modality in a prompt, but infrastructure for recording. Signed data, citability down to the paragraph, logs not erased during compression, disclosure protocols. Everything we can "transplant" from Tuyuca into a machine today is the idea of mandatory attribution, not its grammatical form. We can only imitate morphology — and the imitation will be neither honest nor useful.
And the last, most human point: in those communities where evidentiality grew, it was never an "anti-fake technology". It was a way to live next to people who can't be double-checked — and not lose trust between them. Perhaps that's exactly why our industry is so enchanted by this idea: we too fear a world where checking everything is impossible and saying anything is allowed. Grammar doesn't save from lies — it only makes lies announce themselves louder. And then begins work that can never be done with a morpheme: listening to how that lie sounds.