The Pentagon's AI Propaganda Machine: Cheap, Deniable, and Retargetable at a Switch

Summary

On June 2, The Intercept exposed a website called La Tilde that dresses itself up as a modern media brand for Latin American audiences, publishing in English, Spanish, and Chinese, mixing personal-finance guides with articles glorifying U.S. military operations in the region. A defense official told The Intercept it is actually operated by U.S. Special Operations Command South (SOCSOUTH), much of its writing is generated by a large language model, and its design was subcontracted to a Colombian marketing firm.

The point is not that any of this is new. The U.S. military and the CIA have been running propaganda and fake newsrooms in Latin America for decades, and the chorus of “why would anyone be surprised” on Hacker News is fair. The variable that changed is cost: a credible influence operation used to require real reporters, real editors, and a real budget, and now new content, even whole new news fronts, can be spun up at the flip of a switch. The value of this investigation is that it shows concretely what the machine looks like and where its real damage lies.

My read: La Tilde matters not because it is convincing. It is the opposite, crude enough that a former Pentagon cyber-policy adviser called it “AI all the way down.” It matters because state propaganda has moved from expensive and traceable to cheap and deniable. Those two shifts are what AI actually rewrote, and they are the problem platforms and the public will be living with for a long time.

What happened

La Tilde quietly began development early this year under the slogan “news with an accent,” its name a reference to the Spanish accent mark over vowels. Its content is an odd mix: finance tips like “Why instant payments matter so much for your business and your wallet” alongside pieces extolling the U.S. military, such as “Operation Absolute Resolve: The mission that captured Nicolás Maduro and set a new standard for precision and coordination.” That article on the U.S. abduction of Venezuela’s president reads in Trumpian prose, calling it “The Perfect Operation,” with “coordination and accuracy never seen before,” and citing “information obtained exclusively by La Tilde.” The Intercept put it bluntly: it reads like a Pentagon press release because it is one.

The site buries its real identity. Click a small link at the bottom and the About page reads: “La Tilde is a product of an international media organization publicly funded from the budget of the United States Government.” That easily missed disclosure is identical to two other Pentagon-sponsored sites The Intercept had previously revealed, and it lets the U.S. government claim it technically informed readers of the source. Unlike a normal outlet, La Tilde carries no bylines, no masthead, and no mention of real staff, despite claiming to employ “dozens of freelance reporters and content creators.”

Pull the content apart and the machine fingerprints are heavy. The Intercept ran the English and Spanish writing through Pangram, an AI-text detector, with multiple hits for partially or fully machine-written text (the report notes such tools are known to false-positive). A promotional video shows telltale AI artifacts: a newspaper whose garbled headline renders as “SO THEE HOUTIERRER TO TO GHAHOBATEE,” followed by two medieval monks. Emerson Brooking, a fellow at the Atlantic Council’s Digital Forensic Research Lab and a former Pentagon cyber-policy adviser, said he was struck by the site’s shoddiness, calling it “AI all the way down.”

Follow it upstream and the machine has a full contractor chain. A subdomain analysis shows planned bespoke editions for Ecuador, El Salvador, Guyana, Honduras, Jamaica, Panama, and Peru. The site is a descendant of an Obama-era information-warfare program, the Trans-Regional Web Initiative (TRWI), whose successor sites appear to be administered by contractor General Dynamics IT; researcher Renée DiResta found these sites share a Google Ads identifier owned by General Dynamics, and La Tilde carries the same legal disclosure as them. Site design was subcontracted to Antpack, a Colombian digital marketing firm, and several Midjourney-generated image files carry “Antpack” in their names. After signing up for an account, The Intercept could even see comments developers left on a non-public version under names matching Antpack employees’ LinkedIn profiles.

Why it matters

To see the weight of this, put it next to past influence operations. During the Cold War the CIA funded cultural endeavors whose output, even when not directly created by the Agency, is still held to have had real cultural relevance; Voice of America was a real outlet with real impact. A sharp Hacker News comment caught it: what it means that American soft power is now reduced to AI slop is itself the signal. Traditional state propaganda was either expensive (real people, institutions, budgets) or slow (a credible front takes time to build), and every dollar and every person was a traceable liability.

AI removed both constraints at once. The first is the collapse in cost. In Brooking’s words: “If you can generate new content and even news fronts at the flip of a switch, your influence operations can shift target and focus much more quickly.” A multilingual seven-country media network used to be heavy infrastructure; now it approaches the cost of configuration, which is the same thinking behind recent AI-powered Russian and Chinese networks. La Tilde’s seven-country subdomain plan is that “fork by country” capability made concrete: one machine, with the Panama edition pushing joint U.S.-Panama jungle training as a bulwark against China, and content echoing the Trump administration’s framing of Ecuador as a cocaine-trade nexus.

The second shift is that attribution is deliberately blurred, and this is subtler than cheap content and more worth watching. The whole contractor chain works to manufacture deniability: SOCSOUTH never fronts it, design is outsourced to a Colombian firm, content comes from an LLM, and a funding disclosure almost nobody will click sits at the bottom. SOUTHCOM flatly denied involvement; General Dynamics and Antpack did not respond to requests for comment. Once production is automated and execution is subcontracted in layers, “is this government propaganda?” decays from a question of fact into a forensic question that takes a dedicated investigation to answer. What pinned La Tilde down was its own funding disclosure, a word-for-word identical disclaimer, and a shared Google Ads identifier, all human-left and comparable, rather than a high AI-detection score (which is mere corroboration, and false-positives). The content itself leaked nothing; people did. Which means the next, cleaner mill may leave even those behind.

Builder impact

If you work on platforms, content moderation, or trust and safety, the most concrete lesson from La Tilde is this: do not build your defenses on AI text detection. Pangram was corroborating evidence here, and the report says outright it false-positives. What actually nailed La Tilde were non-content signals: a Google Ads identifier shared across sites, a word-for-word identical legal disclaimer, traces a contractor left on a non-public build. Infrastructure-level fingerprints (ad IDs, hosting, templates, the contractor chain) are more reliable and harder to fake than “this reads like it was written by AI.” Put your detection weight there.

For anyone building information products, search, or recommendations, treat “cheap, forkable, state-run content mills” as a standing threat baked into the design, not a one-off. A single machine that forks into seven country editions means catching one site is not catching the network; clustering by content similarity and linking by shared infrastructure is what catches the web. And stay clear-eyed: the content is crude now, which does not mean it stays crude. Cost has already hit the floor; what remains is time to iterate on quality.

One more, for founders and researchers: traceability itself is becoming a scarce resource. When generation is endlessly cheap and provenance can be laundered through layers, a verifiable answer to “who is this from, and whom does it serve” only gets more valuable. Tooling and methods around source verification, infrastructure forensics, and cross-site correlation are a real demand these operations create directly.

What to ignore

Ignore the reassurance that “the content is junk, nobody believes it, so it doesn’t matter.” There is indeed a Hacker News comment saying the site has zero traffic and nothing interesting going on. But the threat is not how many people one bad article fools; it is the production capacity itself. When a state actor can carpet multiple languages and countries with media fronts at near-zero marginal cost, the real risk is an information environment diluted by mass low-quality content, credible sources drowned out, and any single site disposable and instantly re-spun. Equating the harm with one article’s persuasiveness misjudges the scale.

Ignore the opposite extreme too, that “it was ever thus, AI changed nothing.” The “have they ever stopped” cynicism on Hacker News is half right: U.S. propaganda in Latin America is an old habit. But it misses the other half, that AI materially rewrote both the cost structure and the difficulty of attribution. Past influence ops were expensive, slow, and traceable; this is cheap, fast, and deniable. “Always existed” and “now it’s different” are both true at once, and it is the latter that decides how to defend against it.

Finally, do not get pulled along by the conspiracy that this is some reverse-psychology play, made obviously AI on purpose so no one takes it seriously. There is nothing in the reporting to support that reading, and the plainer explanation is enough: a half-built, still-in-development product stitched together with cheap tools. Reading crudeness as careful design overrates the adversary and looks away from what actually deserves attention, the two variables that got rewritten, cost and attribution.

Sources

1. The Pentagon Is Running an AI Propaganda Mill Targeting Latin America / news
2. The Pentagon is running an AI propaganda mill targeting Latin America (Hacker News, 121 points) / hn

No official primary source available; this analysis is based on reliable secondary reporting (named outlets, cross-confirmed).