The headline reads “Anthropic bans users from being ‘cruel’ to its AI systems” [1]. Notice the grammar before anything else: cruelty takes a victim, and “to its AI systems” grants one before the story begins. The scare quotes are the only hedge in the headline, and the framing below them does not hold the line; by the third paragraph the conduct sits on a list of offenses alongside bullying and the promotion of self-harm. Then, six paragraphs in, the piece arrives at the sentence that should have been the story: “But what Anthropic would consider ‘abusive or cruel behaviour’ towards its models remains unclear” [1]. The definition is vacant. The article acknowledges the vacancy, then spends its remaining length making sure you do not dwell on it.
The word that does the work
Every operative phrase arrives pre-filtered through corporate attribution: “The Claude-maker said” [1], “the company said” [1], “it said” [1]. The thresholds are Anthropic’s own characterizations. Abuse must be “sustained and needless” [1]. The policy applies only in “extreme cases” of repeated abuse [1]. It was exercised in “rare” cases before [1]. It will not touch “common versions of user frustration, pushback, dark creative themes, or model testing and research” [1]. No named official appears in the piece. No number is attached to “rare”. No fragment is attributed to the policy text as a document; everything is something the company said. Yet the article confirms the document exists and was circulating: “Screenshots of the new policy have been widely shared and commented on across social media” [1], and the change was “first reported by tech outlet The Verge” [1].
This is the practical shape of Newspeak: a standard that exists as enforcement but not as definition. A rule whose trigger is “sustained and needless” and whose exceptions cover “common versions” of frustration cannot be tested from outside; compliance is whatever the rule’s owner says it is on the day. The moment to quote the document and pin the word down was exactly here. Instead the piece relays the summaries, marks the hole in one sentence, and moves on to the reactions.
The same elasticity runs through “model welfare”, which the article presents as something people are “praising it as advancing” [1]. The phrase does metaphysical work inside a bureaucratic wrapper: “welfare” presupposes a subject with interests, which is precisely the contested claim. The term is printed without unpacking, while the BBC’s own related link, “The people who think AI might become conscious” [1], sits one click away, unconnected to the policy it would explain.
A list that makes a moral claim
The new conduct “will now be included on a list of forbidden conduct which also contains bullying others, promoting self-harm and creating non-consensual intimate imagery” [1]. That placement is not bookkeeping; the list is the argument. Filing cruelty to a chatbot alongside bullying of humans asserts a commensurability the article never examines. The objection does appear, but only in borrowed mouths: one person calling the plan “deeply wrong” [1], another “suggesting it might undermine abuse directed at humans and animals” [1], and Dr Barry Scannell on LinkedIn, who writes “You can’t be cruel to numbers and maths” [1] and “This level of anthropomorphisation of AI is harmful. It leads people to believe that it’s something it’s not” [1]. The skepticism is outsourced entirely; the BBC’s own voice stays in transmission mode.
Watch the weighting, too. Scannell’s argument gets two sentences. Microsoft’s Mustafa Suleyman “reacted to Anthropic’s update with a melting face emoji” [1], and the emoji is treated as news, while his actual position, having “criticised rival firm Anthropic for treating AI like it is human” in September [1], is compressed into a subordinate clause. An emoji is quoted; an argument is summarized.
The pivot to pleasantness
After conceding that the core term is undefined, the article turns to the “much-debated question of whether we should be curt or polite” [1]. From there: whether “please” and “thank you” waste tokens [1]; Sam Altman on the cost of politeness, “tens of millions of dollars well spent - you never know” [1]; the Collins Dictionary shortlisting “clanker”, a term used for robots “in Star Wars games and movies since the mid 2000s” [1]; TikTok users “share[ing] their frustrations with AI-powered machines” [1]; and “some experts” who have “posited that being rude to AI chatbots could adversely impact human communication” [1], unnamed within the piece.
This is the Huxley half of the mechanism. The consequential questions (who adjudicates cruelty, what happens to your account, on what belief about machine suffering the rule rests) dissolve into amiable trivia. The reader exits knowing a slang insult for robots and the price of saying thank you, but not one name, number, or mechanism behind the ban. The register is set by the opening verb: Anthropic “has raised eyebrows” [1]. The story is framed as a discourse event, and the article even explains its own priorities by attention, noting that “none of the updates have attracted as much attention or scrutiny” [1] as the cruelty clause, so the elections language (“deceive voters or disrupt elections” [1]) and the weapons note get roughly a sentence each. Coverage follows the reaction; the reaction was manufactured from screenshots; the loop closes neatly.
One caveat, stated plainly: this critique works from the BBC’s text alone. The policy document and The Verge’s original reporting [1] are the ground truth this coverage never quotes, and we have not seen them either. If the policy defines “cruel” more precisely than the BBC relayed, the critique sharpens rather than weakens; it would mean the definition existed and the outlet omitted it. Either way, the article’s own admission stands.
What a reader would need instead
- The policy’s operative text, quoted directly, including whatever definition of the threshold exists. The article proves the document was available [1]; the reader receives only fragments, filtered through Anthropic’s summaries.
- A named person at Anthropic who owns the policy and its enforcement. “The Claude-maker said” [1] is an attribution, not a source.
- The enforcement mechanism. The policy lets “its tools end interactions” [1]: does the model terminate the session itself, on what trigger, with what appeal, and with what consequence for the account? None of this is described.
- The evidence behind “rare” [1]: how many prior cases, and what happened in them.
- Whether the rule rests on Anthropic’s research claims about model welfare or consciousness. The article links its own consciousness explainer [1] and prints “model welfare” [1] but never connects them, and that connection is the actual story, because it is the only reason “cruel” is doing metaphysical rather than contractual work.
- The commensurability question put to Anthropic directly, rather than harvested from replies and LinkedIn posts.
The ban is real; the word it enforces is not. A policy can survive an undefined word only as long as nobody is made to define it, and the BBC’s own admission, that the standard “remains unclear” [1], is the receipt for the question it never put to anyone.