Catch-22

Jeffrey Wernick ScheerPost

Wikipedia states that its articles follow a Neutral Point of View. The policy is the platform’s central promise, the thing that separates an encyclopedia from a brief. The fair way to examine any article is not to ask whether it is generous to its subject, but whether it honors the standard Wikipedia sets for itself. The article on BitChute is a useful case, because it can be measured against that standard, and the measurement does not require trusting the subject about anything. It requires only reading Wikipedia against Wikipedia.

Begin with the kind of failure this is, because it is not the kind most readers expect. A false statement can usually be corrected by another true statement, which is why factual error, for all its mischief, is the easy case. Selection is different. When every sentence is individually true, the resulting narrative acquires the appearance of objectivity while quietly steering the reader toward a conclusion the underlying record does not compel. Nothing in it can be flagged as false, because nothing in it is. That is why neutrality depends as much on what is omitted as on what is included, and why the only way to test it is to set the article beside the record it was drawn from and read for the gaps.

Begin with the sources. The great majority of the article’s citations fall between 2018 and 2021, and a small set of advocacy organizations supplies most of the characterizations. An entry on a living platform, built almost entirely from one narrow window years after that window closed, does not describe what the platform is. It preserves what a particular set of sources said during a particular period. Anyone can confirm this by scrolling the reference list, and it frames everything that follows. A time capsule is being presented as a current account.

Now the events the article recounts. It reports that a charity referred BitChute to the United Kingdom’s media regulator. It reports that Police Scotland opened an investigation into content targeting Scotland’s First Minister. It notes a German law requiring platforms to report hate speech. Read carefully and notice what these have in common. None produced an adverse finding against the platform, and the article does not say so.

The regulator’s matter is the clearest. The regulator’s own assessment, found in the same reporting the article draws upon, was that BitChute had engaged constructively and had increased the number of its moderators. The regulator stated that under the governing regime there was no requirement to proactively monitor content, and that the mere existence of harmful content would not mean BitChute was failing to comply. It further stated that the requirements applied to video content and not to the text comments attached to videos. That last point matters greatly, because the article relies on a description of racist comments in comment sections to undercut the platform’s stated policies. The authority with jurisdiction placed those comments outside the platform’s compliance obligation. There is a second reason the comment material cannot bear the weight the article places on it. For much of the period in question the comments were not BitChute’s system at all. They were served by a third party, Disqus, whose label was visible to every user, and which BitChute could not moderate at the level it later could when it built its own system. The article assigns to the platform the contents of a tool the platform did not yet control and the regulator had already placed outside its duty. The very material elevated to a defining trait was, by jurisdiction and by architecture, the part for which the platform was least answerable.

The German example works by arrangement rather than omission. The article notes that the law required platforms to report hate speech and that BitChute was not subject to it, leaving a reader to infer evasion. The actual reason is a size threshold. The law was aimed at the largest platforms, and BitChute fell below the line, as many smaller services did. The same reporting the article relies on notes that Steam was likewise unaffected. Steam’s article carries no such mention. To raise a law a subject is not governed by, in a way that implies the subject slipped through it, and to raise it for one platform and not the other named in the same breath, is insinuation by selection. It tells a reader nothing about conduct and everything about posture.

The Scottish investigation completes the pattern. Police Scotland opened it after a complaint about specific videos aimed at the First Minister, and the article frames it as a matter concerning the platform. Read what it actually was. The complaint was about content and the people who posted it, not a finding against BitChute, and the same videos were cross-posted to far larger platforms, where they drew more views and in places remain. An investigation into a creator’s post becomes, on the page, a fact about the host, and only this host, though the post lived in more places than one. As with the regulator and the German law, no adverse finding against the platform followed, and the article does not say so.

This is the place to concede what should be conceded, because the argument does not require BitChute to have been faultless and it was not. Illegal content did appear. Counter Terrorism Policing told a parliamentary committee that material from a proscribed organization had been hosted and that action had been taken to remove it. The platform’s founder did, on his personal account on another service, promote conspiracy material, and that is fairly attributed to him. A neutral article would record these things. It would also record what happened next, and this is where the article’s method shows. Take the anti-extremism report the article cites for the presence of terrorist content. The figures belong to the report and can be checked against it, and they should be before anything is published. The report identified a set of videos in the low hundreds, many with only tens of views. When BitChute learned of it, it asked the group for the list so the content could be removed. The group declined to provide it. BitChute then informed the police, the police obtained the list, and the videos came down. None of that sequence appears. Nor does the composition. The article foregrounds the neo-Nazi examples, while the larger share of the flagged material, by the report’s own count, came from other sources entirely. The accusation is printed. The removal is not. The emphasis inverts the report it claims to summarize. That is the pattern in miniature, the referral without the finding, the report without the response, the investigation without the result, so that the accumulated impression is of guilt established, when what the record shows is guilt alleged and, in the formal venues, never found.

Consider the causal claims, which sit at the center. The article quotes an advocacy group’s assertion that the platform chooses to almost exclusively promote hateful content, and that the content is the result of deliberate decisions by its founder and team. Read the verbs. Chooses. Deliberate. Promote. These are claims about purpose, and purpose is not what the evidence presented underneath them measures.

Here is the strongest form of the other side, and it deserves to be stated plainly before it is answered. An encyclopedia, a defender would say, is entitled to report what a platform contains. If a great deal of extreme material found a home here, that is a true and relevant fact, and the no-mechanism defense answers a charge of intent the article need not be making. Report the concentration and let the reader judge. Grant all of it. If the article confined itself to composition, stated its method, and carried the resolutions, there would be little here to answer. That is not what the article does. It does not say the platform contained. It says the platform chooses, is dedicated, promotes. It adopts sources that assert design. The objection is not that composition is off limits. It is that the article crosses from the empirical claim it might support to the intent claim it cannot, and drops the resolutions that would temper even the empirical one.

Two operations are being collapsed into one. The first is to observe and categorize a platform’s content, a task that, done with a stated method, can tell you what a platform contains. The second is to attribute intent, to say a platform is dedicated to something or chooses to spread it. The first does not by itself establish the second, and the reason is exact. The same content distribution is consistent with two entirely different platforms, one designed to favor a viewpoint, and one that is content-neutral and was simply migrated to by people removed from elsewhere. The composition looks identical in both cases. An observation of composition cannot distinguish a dedicated platform from a neutral one, which means the cited evidence does not by itself support the word dedicated. To read purpose off content is to claim the thermometer has measured weight.

What would be required to establish intent is evidence of mechanism, of the choices a platform actually makes, screening at the door, ranking by ideology, the promotion of one post over another. Here the architecture answers plainly. The platform screened no one by viewpoint, ran no ranking algorithm, and gave no post priority over any other. There was no instrument through which a preference could be enacted, because there was no instrument of preference at all. A platform that does not algorithmically rank or preferentially distribute content lacks the principal mechanism through which modern platforms promote one viewpoint over another, and a platform that does not screen at the door lacks the means to exclude one. And no systematic pattern of viewpoint suppression has been demonstrated, because no mechanism capable of producing such a pattern has been identified. If the platform suppressed a viewpoint, there would be users of the disfavored kind with throttled posts, removed videos, denied accounts, and a documented pattern among them. No such pattern has been shown, and no such mechanism has been found that could produce one. The same defect sinks the strongest single phrase in the article, the advocacy claim that the platform was a recruiting ground. Recruiting is an action, and it requires a means. The alleged recruiting the claim rests on did not occur on BitChute. It occurred on another service, on Telegram, where a video hosted here was said to have been posted. The platform with the tools for organized contact is given the neutral treatment. The platform that served the file is assigned the act. The accusation describes a capability the architecture does not contain.

So the content observation, even performed perfectly, does not establish purpose, and the purpose claim has no other evidence presented behind it. This is the difference between a policy and a design. The sources have asserted the design while documenting, at most, the effect of a policy. To adopt the stronger claim when the evidence reaches only the weaker one, and when the mechanism the stronger claim requires is demonstrably absent, is not a finding. It is a characterization given the grammar of a finding. The same grammar shapes the origin story. The article says that since launching the platform accommodated far-right groups, a sentence that is technically a claim about what was permitted and rhetorically a claim about what the platform was. The archived launch content shows a different early mix, weighted toward libertarian, non-political, and gaming channels. What a neutral host allows is read backward as what a dedicated host intended.

The list of creators makes the selection visible. The article names figures, banned or demonetized elsewhere, who posted to BitChute, and offers them as evidence of what the platform is. For nearly all of them BitChute was a secondary platform, a mirror or a backup, not the place they built or held their audiences. They are characterized by their primary platforms and used to characterize this one. The same person, posting across a dozen services, is treated as evidence about the smallest of them and not the largest. A trait shared across an industry is assigned to a single member of it.

Notice the opening word. The article sorts BitChute into a category it calls alt-tech, a term whose construction implies a lesser, off-brand alternative to legitimate platforms. The label is presented as plain description. It is a contested framing, drawn from one set of observers and set down as definition. To see that the choice was a choice, read the article on Rumble, a platform with much the same self-description and a heavily overlapping roster of creators. There, the term is attributed. Some describe Rumble as alt-tech, that article says, marking the word as a characterization others have made. On BitChute’s page the same word is stated as fact, in the first breath, unattributed. The identical term is handled two ways on two pages, and the difference tracks nothing but which platform was already suspect. The reader meets the verdict before the first fact.

The same asymmetry runs through the grammar of the article. What favors the platform is rendered as assertion. What cuts against it is rendered as fact. The platform describes itself as offering free speech, the article says, holding the claim at arm’s length, while the platform is known for hosting the rest, stated flat. The pattern repeats on a question with a checkable answer. The article says BitChute claimed to use peer-to-peer distribution, and that this was disputed, though the implementation is visible to this day in archived source code. A documented technical feature is cast as a doubtful boast. A characterization by critics is cast as settled record. The positive is always the platform’s word. The negative is always simply true. That distribution of doubt is not neutral. It is a thumb on the scale, applied one grammatical clause at a time.

Selection also decides what stays on the page long after it has stopped being true. The article makes a point of the registrar BitChute once used, a company it is careful to call known for hosting far-right content, and it leans on the Southern Poverty Law Center to do so. Ask why a domain registrar belongs in an encyclopedia at all. The platform left that registrar for Cloudflare years ago, and used others before it, and none of those appears. What survives on the page is the single supplier whose name could be made to carry an adjective. The current arrangement is ordinary, so it goes unmentioned. The retired one is named, because the naming was never about the registrar. It was about the company the word kept.

The article goes further and cites a study in the journal Science that sorts these platforms into two groups, the ones it calls dedicated to right-wing communities, where it places BitChute alongside 4chan, 8chan, Parler, and Gab, and the ones it calls ideologically neutral, where it places Discord and Telegram. The sorting collapses under a single question. By what criterion is BitChute dedicated to the right while Telegram is neutral? In the seditious-conspiracy prosecutions arising from January 6, evidence introduced at trial showed that channels on Signal and Telegram were among the platforms where the Oath Keepers and the Proud Boys conducted much of their operational coordination. If the classification tracked the use of a platform by violent political actors, Telegram could not be the neutral example. If it tracked content, the same would follow. What places Telegram in the neutral group and BitChute in the other is not a measured property. It is prior reputation and size. The large and familiar platform receives the benign label, the small and already-suspect one receives the damning label, and a sorting that claims to describe the platforms in fact describes their reputations. That is precisely the judgment an encyclopedia is meant to test rather than inherit.

There is a deeper failure in the study, and it belongs to the word dedicated. A journal of science could honestly report a measurement of content, this platform contains a high proportion of such material, a claim that is empirical and, done properly, testable. To say a platform is dedicated to a community is not that claim. It is an attribution of purpose, and purpose cannot be measured by sorting content, for the reason already given. A content sort published as an intent finding is not a more rigorous claim than an op-ed. It is the same claim wearing the authority of a method that does not reach it. That a journal named Science set its name to a classification of institutional purpose, derived from no criterion that would let another researcher reproduce the sort, is the failure. Not that the conclusion offends, but that the one standard distinguishing the journal from an opinion page, a stated and reproducible method, is the standard the classification lacks.

None of these is, in isolation, an error of fact. Each sentence can likely point to a source. That is what makes the article a good illustration rather than a simple case of falsehood. The failure is not in any one citation. It lives in the selection and the arrangement, in what is included and what is left out, which accusation is foregrounded and which resolution goes missing, which window of time stands in for the whole. Selection is where neutrality lives or dies, because an article can be built almost entirely of true sentences and still, by the choice of which true sentences to print, deliver a verdict the sources do not support.

Here the deeper mechanism comes into view, and it is the reason this piece carries the title it does. At one point an editor added a single sentence drawn from a Pew Research Center study, a source the article already relies on for other propositions, and the sentence happened to cut toward the platform’s neutrality. It was removed. A neutral research organization’s finding was excluded, while the characterizations of advocacy organizations remained as the article’s baseline. The asymmetry is not incidental, and it is not even hidden, because Wikipedia has written its own test and failed it in plain view. Wikipedia’s own guidance treats two of the article’s anchor sources, the Southern Poverty Law Center and Hope not Hate, as biased and opinionated advocacy groups whose statements should be attributed rather than stated in Wikipedia’s voice, and in the case of the first, kept out of the lead absent a case-by-case decision. The article uses them anyway, in its voice, near the top, the SPLC supplying even the registrar claim already discussed. So set the two together. The peer-reviewed journal and the independent research body are about as authoritative as sources get under Wikipedia’s own reliability standard. One was admitted and featured. The other was removed. The advocacy sources the policy says to attribute were stated as fact. The difference was not rigor, which would have kept the strong sources and qualified the weak ones. The difference was direction. A standard that names reliability as its criterion and applies the conclusion as its criterion is no longer a reliability standard. It is a filter wearing the name of one.

Sit with what that requires. Wikipedia’s neutrality policy does not ask editors to determine truth. It asks them to reflect reliable sources. In ordinary cases that is a sound rule, the means by which an encyclopedia keeps out cranks. But consider what the rule produces when the universe of sources deemed reliable has, on a given subject, settled into a single posture. The advocacy organizations are admitted. The disinterested research body is turned away. And the editor enforcing that exclusion can say, accurately and within the rules, that he is only reflecting reliable sources, while the reliable sources available to him are precisely the parties who have already reached the verdict.

That is the trap, and it closes on itself. To be represented fairly, a subject needs coverage in the sources the policy treats as reliable. When those sources will not engage the subject except to condemn it, the only path to correction runs through the very outlets that have declined to give a fair hearing, and their refusal is itself what makes the record uncorrectable. You cannot get in because you are out, and you are out because you cannot get in. The rule meant to ensure quality becomes the rule that guarantees the subject can never answer.

What that mechanism denies is older than any of the parties to it. It is the right of anyone characterized by others to engage those who characterize them. The law enshrines this in its most serious setting, the right of an accused to confront an accuser, not because the principle is confined to courtrooms but because a courtroom is where we are most careful to honor what the principle reflects everywhere. Let me be exact about what this argument rests on, because the honest thing is to name the value doing the work. The case here does not depend on free expression being the correct philosophy of moderation. That is a separate question, and a reader can hold any view of it and still follow every step above. The case depends on a narrower commitment, that an account assembled to characterize a living subject owes that subject the chance to answer, and that an encyclopedia which assembles such an account entirely from a subject’s critics, with the subject’s evidence ruled inadmissible, has rendered a judgment without confrontation. That is a value, not a finding, and I am stating it as a value. It happens to be the value Wikipedia printed at the top of its own page.

So I will honor the principle from my side of it, since the structure will not honor it from the other. To anyone who has studied this platform and reached a conclusion about it, the door is open, and my name is on it. Come and make your case directly, in an unedited conversation, with the floor shared and the time yours as much as mine. State your findings to me and to an audience, and let me answer, and let the exchange stand as the engagement that a record built without it has lacked. I am not hiding. I am inviting. The right to engage those who characterize you is a principle of justice, and the way to vindicate a principle is to live by it, especially toward those who have not.

This is the place to widen the lens, because the article on one platform was never the real subject. It is a worked example. If a single entry, read closely against Wikipedia’s own standard by someone holding the primary record, reveals this much, the referral printed without the finding, the report stripped of its response, the advocacy source stated as fact while the research body is turned away, then the honest question is not whether this article failed. It is how many others have, unread by anyone in a position to notice. The defect is invisible by construction. It is found only when a subject with the documents and the standing to answer happens to sit down and check, and almost no subject ever does. For every article audited this way there are thousands that never will be, and nothing on their faces would tell a reader which ones carry the same quiet thumb on the same scale.

So this was never about BitChute, and it was never really about free expression as a philosophy of moderation. It is about what we agree to treat as a trusted source, and what happens to a subject when the sources a rule deems reliable have closed ranks against it. The damage falls on the BitChutes of the world first, the marginal, the already-suspect, the ones the reliable sources will engage only to condemn. But the mechanism that damages them is indifferent to its target. It will characterize anyone the approved sources have decided about in advance, and it will do it in the institution’s calm voice, and it will call the result neutral.

And the stakes have just risen by an order of magnitude, because the captured page no longer ends at the page. Wikipedia is among the most heavily weighted sources in the material on which large language models are trained, and among the first an AI system reaches for when asked what something is. A characterization assembled by selection launders, through that training, into the voice of every system that learned from it. The model does not see the talk page. It does not see the removed Pew sentence or the stripped resolutions or the registrar that was named for the company it kept. It sees the conclusion, absorbs it as fact, and repeats it to a user who has even less ability to confront the source than a reader of the original article had. There is no editor to petition, no diff to inspect, no primary record in the loop. The judgment rendered without confrontation is now rendered again, automatically, at scale, by systems that cannot themselves be confronted either. The concern is not hypothetical, and it is not only mine. As these systems come to mediate what people know, the worry that a captured record poisons everything trained on it has begun to be raised by others watching the same convergence.

This is the Catch-22 raised to a higher power. The original trap was that a subject could not correct the record without the cooperation of the sources that refused to engage it. The new trap is that the uncorrected record now trains the machines that the next generation will treat as the arbiters of what is true, and those machines inherit the verdict without inheriting the means to question it. The cost of an abuse of trust at the source was once bounded by the reach of the encyclopedia. It is now bounded by the reach of artificial intelligence, which is to say it is, for practical purposes, unbounded. Every distortion that enters here is copied forward into systems that will repeat it with a confidence the original never claimed, and that confidence will be received as authority.

So the value at stake was never the reputation of one website. It is whether the infrastructure of shared knowledge, the encyclopedia and now the models built upon it, can be trusted to characterize a subject it has reason to dislike without abandoning, in that very case, the standard it advertises to everyone else. On the evidence of a single page, read against its own rule, it cannot yet be. And the machines are already learning from the page.

A rule remains at the top of it. It is cited, invoked, obeyed in form. What has gone missing is the thing the rule was written to produce, an account a reader could trust to have weighed both sides rather than assembled one. The form persists. The function is gone. That was a containable failure when the only thing that inherited it was the next reader. It is a different kind of failure now that the next thing to inherit it is every system we are building to think for us. There is only the standard, printed at the top of every page, and an article, and soon a machine, that can no longer see it.

You can also make a donation to our PayPal or subscribe to our Patreon.

Please share this story and help us grow our network!