Skip to main content

Your feed is not the world: who repeats an opinion until you own it

· 9 min read
Victor Jimenez
Victor Jimenez
Senior Product Engineer

The feed you scroll is a selection made by someone with interests, and repetition is the tool that turns that selection into what feels like reality. Before a message you saw ten times today changes what you believe, ask who benefits from you seeing it, find the primary source, and spend time outside the network. This is my observation and opinion, not a measured study.

Repetition makes a claim feel true​

Psychologists call it the illusory truth effect: people judge a statement as more true after they have seen it repeated. The uncomfortable part is that knowing the answer does not protect you. Fazio, Brashier, Payne and Marsh reported in 2015, in the Journal of Experimental Psychology: General, that the effect appeared even when participants knew the correct answer (PubMed 26301795).

A 2020 study in Psychological Science found the effect in 5-year-olds, 10-year-olds and adults, and prior knowledge did not protect any age group (Vanderbilt news, 2020-10-06). My conclusion from that: "I know better" is not a defense. The defense is to control what gets repeated at you and to check where it came from.

A feed is a selection, and the selector has interests​

A feed is the ranked list of posts a platform chooses to show you. Someone decides that ranking, and that someone can have a stake in what you believe. My read: repetition inside a feed is a ranking decision someone made, even when it looks like a crowd.

The evidence on how much an owner steers the ranking is thin, and I want to be exact about it. A 2024 preprint by Timothy Graham (QUT) and Mark Andrejevic (Monash) found a sudden boost in engagement on Elon Musk's posts after July 2024. The authors describe it as possible algorithmic prioritization, not proof, and say platform opacity makes it hard to be certain (QUT news).

A preprint is a paper that has not finished peer review. The cases below do not depend on it, because in the first one the message comes directly from the person who owns the feed.

Case one: the platform owner is a player in the race​

Elon Musk owns X. In October 2026 he posted "Talarico hates Texas" and "Talarico hates Texas and America" about James Talarico, the Democratic nominee for US Senate in Texas, who is running against Republican Ken Paxton, whom Musk supports (CNBC, 2026-10-23). Talarico's campaign ad puts Musk's spending push in the race at about $190 million and says "Texas is not for sale"; that figure is the campaign's claim (Quartz).

The same race also shows how context gets cut. Musk's America PAC got Community Notes on X over a clipped video of Talarico, because the clip cut off the rest of his sentence (AOL news report). The primary source is the full sentence, not the clip.

I do not tell anyone how to vote, and this is not about the candidate. My point is the structure: the owner of the feed is also a player in the fight, and the repetition is the strategy. A slogan repeated by the person who owns the distribution is a campaign, whatever the timeline calls it.

Case two: insider alarm that doubles as capability marketing​

I use AI coding tools every day, so my skepticism here is about claims, not about the tools existing. On September 8, 2026, Jacob Coxon announced on X that he resigned from Anthropic, after three years of pretraining research at OpenAI and Anthropic (Fortune, 2026-09-09). He wrote "Neither company is acting responsibly" and:

They are racing straight to self-improving superintelligence and gambling with our lives.

Jacob Coxon, on X, as reported by TIME and Fortune

Mrinank Sharma, head of Safeguards Research at Anthropic, also resigned, with an open letter saying the world is "in peril" (The Hill).

Some of these people may be sincere. TIME reports that Coxon left two months before his equity would have vested and said the money barely factored into his decision (TIME, 2026-09-15). I claim no one's motive.

My read: a warning that a model could end the world also tells the market the model is more powerful than it is, so alarm and capability marketing point in the same direction. TIME collected critics who make a related argument: Cohere CEO Aidan Gomez called the slowdown push "a cartel by any other name", arguing fear lets the largest firms set rules for competitors, and David Sacks called calls for intervention "regulatory capture" (TIME, 2026-09-15).

TIME also quotes an infrastructure engineer who called the AI-made virus extinction claims "total bogus". In the same period Dario Amodei, the CEO of Anthropic, published an essay warning misaligned AIs could, in as little as six to 12 months, be "capable of taking over the entire internet", and neither he nor Sam Altman committed to halting training (TIME, 2026-09-15). A warning that takeover capability could arrive in as little as six to 12 months, next to no commitment to stop training, reads to me as a capability claim with a warning label.

The scream is not evidence. Ask for the evidence: a measurement, a published test, a document someone else can check.

Case three: "uncensored", billed per token​

The third pattern is a sales pitch. Paid services sell "uncensored" or "abliterated" open models by API, including DeepSeek variants. Abliteration is an edit to the model's weights, the numbers the model learned in training, that removes the refusal direction: the internal pattern linked to refusing a request.

One vendor, UplinkSwitch, lists a DeepSeek V4 Pro Uncensored model at $1.00 per million input tokens and $2.00 per million output tokens (uplinkswitch.ai/uncensored). A token is a small piece of text, and the bill counts them in both directions.

Now read the same vendor's fine print. It publishes no refusal rates, no quality comparison against the base model and no guarantee that a model will answer. It says refusal behavior "can change with each release", and that a refusal is free only when the response format flags it: "A refusal written as ordinary text is billed like any answer" (uplinkswitch.ai/blog/glm-jailbreak).

So the label reads as no limits, and the documented terms say: no measured refusal rate, no quality baseline, behavior that can change per release, and a bill for refusals that arrive as plain text. Nothing guarantees 100% success.

Separately, abliterated DeepSeek V4.1 Flash forks were free to download on Hugging Face shortly after its September 2026 release (shattered.io report). That is a different model from the V4 Pro the vendor prices, so it proves nothing about the paid one, and no source here shows that a free fork answers every request.

One pattern, three messages​

The three cases look unrelated until you put them side by side. The "who benefits" column is my judgment, not a documented motive.

MessageWho repeats itWho benefits (my read)What to check
"Talarico hates Texas"The owner of X, on XThe side the owner supports in the raceThe full quote and context in a source outside X
Warnings that AI is dangerous or could take overDeparting lab researchers (Coxon, Sharma); an essay by Dario AmodeiLabs whose models sound more powerful; per critics in TIME, the largest firms setting rulesA measurement or document behind the warning
"Uncensored" model, by APIPaid model vendorsThe vendor, which bills refusals written as ordinary text like any answerRefusal rates, quality comparison and billing terms in the vendor's own pages

Three checks that break the loop​

The illusory truth research says knowing the answer does not protect you. So these checks work on sources and exposure, not on willpower.

  • Ask who benefits from you seeing this message ten times today. If the answer includes the person who controls the feed, treat the repetition as distribution; it confirms nothing.
  • Find the primary source and test it:
    • Open the original post, the full video, the study or the vendor's terms.
    • Compare it with a source outside the platform.
    • Look for a claim with a number you can check.
    • Read the vendor's own fine print.
  • Spend time outside the network. The world outside the feed still exists, and it does not repeat one owner's selection at you ten times a day.

The limitation: these checks cost time, and they do not make anyone immune. To verify this post itself, each event claim names its source; open the links beside the facts or in the references below.

References​