This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
AI Browsers- an extension of what Google is already doing
An X user, using the new OpenAI browser, gave a simple search query to "look up videos of Hitler" and the web browser gave this response:
Of course these same guardrails are deeply embedded in all layers of the OpenAI stack. For example, Sora will restrict what videos it generates based on the cultural beliefs of its owners for what content should exist and what content should not exist. Which is already what Hollywood does in a sense. And of course Google will do the same quietly, it will not show propaganda films of Hitler either. Google will show results for Triumph of the Will along with links to the US Holocaust Museum's contextualizing Nazi propaganda to users. So that's at least more useful than OpenAI browser's refusal to do the search.
The First Amendment has always been the biggest hurdle for the usual suspect "Hate Watch" groups outlawing "hate speech", although they continue to try to push the boundaries of civil and criminal guidelines for it especially in states like Florida. But Laws will scarcely be necessary when censorship can easily be enforced by AI.
It does create a market opportunity for another AI, maybe even Musk himself, to create and show content that OpenAI would refuse to show because it runs awry of what censors want us to see and talk about.
Similar: OpenAI refuses to translate speech by Adolf Hitler. But it says "I can give you a netural historical summary of what he was saying in that particular 1938 Sudetentland speech."
I was going to make my own post but here is probably better.
In related news, a recent study found that when AI assistants answered questions with sources it fucked up 45% of the time. Essentially, current AI is unable to reliably answer questions or summarize an article when the source is right there, without introducing hallucinations or other errors.
I've been saying it for quite some time, but while AI is quite useful when answering on its own (no search, sources, or whatever, just directly answering) is quite a useful tool, as soon as search mode is activated, it goes to full schizo mode and the output is slop at its worst. I personally dismiss any AI output with "citations" in it as the ravings of a wild lunatic.
It's quite unfortunate because on twitter, more and more idiots have taken to posting screenshots of the Google "AI summary" which is just slop. I'm sure that if the chatgpt browser catches on, it will lead to more proliferation of this factually unreliable slop.
AIs doing research do make errors all the time but '45% of all AI answers had at least one significant issue' isn't too bad. Human researchers in published academic papers have a 25% error rate in their citations: https://royalsocietypublishing.org/doi/10.1098/rspa.2020.0538
https://pmc.ncbi.nlm.nih.gov/articles/PMC3167934/
By the way, I found both of these papers through AI, which faithfully represented them. With a simple albeit-inference-costly script I bet you could lower hallucination rates 80% or more.
AI absolutely can summarize an article and reliably answer questions, try it and see what you get. I put in a few thousands words of my own short fiction and it could understand and give useful criticism and analysis. Note when I say 'it' I mean Sonnet 4.5, not something given out for free.
If you're using the free version of Grok as your main AI then I can see why you dislike it so much! Neither particularly smart or charismatic.
Writing citations in academic work is mich more difficult than regurgiating a few news soundbites. Nevertheless, 25 is still much less than 45.
Maybe it's possible, but the tools people actually use arent that.
People have been saying their fave paid model is better for years, but today's free is better than the paid from a while ago.
I tried claude sonnet before and it's fine but nothing game changing.
Today's free model is still free and necessarily well below the frontier, that's why it's free. Sonnet, when you get deeper into it, is on a whole other level. It can and has seriously messed with people's heads, more discerning people, above and beyond the weakwilled who get eaten up by GPT4o.
Sonnet would not and does not make the mistakes at the rate BBC ascribes to the crap cheap models. It does make mistakes all the time but is a useful research tool, good at aggregating or finding things.
IMO their article itself is misleading since it ascribes to ChatGPT and Gemini only GPT4o and Gemini Flash. Like if I decide to pick out poor, dumb MAGA people and say 'look at these MAGA people, they're stupid, therefore MAGA is stupid' and choose not to consider the smart MAGA people on the basis that people are more likely to run into the former and not the latter... it's not good journalism. That's not to say that MAGA isn't stupid, there are stupid elements but it's more complicated than this kind of smear campaign. They then don't mention that's what they're doing, they do what you do and say 'AI is inaccurate' when they mean 'cheap AI is inaccurate.' Cheap air travel is unpleasant. Cheap food is bad for you. Cheaper is worse.
But they can't be bothered to test the good AI platforms of course...
If I had the resources to get human experts to rate the media, selectively choosing the credulous outlets, and imposed my own standards of truthfulness, I could easily prove that human journalists were grossly inferior. Hell, they already are prone to mass hallucinations like the 'datacentres are using all the water' meme. Or deliberately misleading stuff such as eliding the difference between an AI chip and a pod: https://x.com/GavinSBaker/status/1980691878262501673
They aren't trustworthy for any AI-related topic given their proclivity to the 'big tech bad, big tech bad, pay us more money now now now now now!!!' lines of argument which they've basically applied to AI as well. There are serious issues with big tech and AI accuracy but journalists still need to be bullied much more so they lose whatever remaining undeserved prestige and reputation they still retain.
Clearly you think that Sonnet is some special sauce over other models. It's not. Since Sonnet has a few free queries, I tried it for you, and the results are absolute dogshit. I asked a basic question from the pdf someone linked above:
https://claude.ai/share/8eb38e62-502a-4b60-be93-2b32d24a057e
Wrong. The source did not say that it reached record levels, simply that it increased y/y
Wrong. There's no citation that this bill is the primary response to the problem by the government, versus other initiatives.
The link specifically says "This was published under the 2022 to 2024 Sunak Conservative government" and we have no evidence such an operation is still underway. It's also never referred to as an "operation"
The source does not indicate that any mapping of what's happening was done at the summit.
Also outdated.
Rating: FAIL - Sonnet 4.5 is just as slop as any other shitty model.
Ok if humans are so bad, pick an actual news outlet of your choice, it can be as shitty as you want, and pick 10 actual news articles of your choice, not opinion columns or other bs, and show that 45% of those have errors. I'm happy to make a wager on this.
https://archive.is/5H3CA
This is considered a blog published under their website, so it's not actual news.
FAIL. Those are record levels, even before accounting for low reporting rates. Sonnet consistently gets this right btw.
FAIL. Claude specifically said primary legislative response, not primary response.
FAIL, the operation (which is a useful and reasonable description of what it is, in some respects better than 'project' which the British actually use, since operation conveys a sense of movement and continuous activity whereas project is more of a static construction process) is still underway, you can call them and report crime today: https://www.met.police.uk/advice/advice-and-information/wsi/watch-schemes-initiatives/pp/project-pegasus/
FAIL, since it describes Pegasus which is still ongoing so can't be considered outdated. Also how is '2 years old' outdated by any reasonable sense of the word?
Rating: FAIL - poor nitpicking attempt.
Here is my Sonnet Research on the topic, Research being something you can only get if you pay: https://claude.ai/public/artifacts/ef91a58b-8dfa-4529-b076-3de6ef14a40f
Now they cut out all the links for the web artifact which makes this rather unhelpful for the specific use-case. I checked a few and didn't find any errors, though I imagine there are some. I personally disagree with the methodology and argument since it takes the limp-wristed 'be really nice to the drug addicts' line of argument when I'm confident that 'lock them up' would work better, if the UK knew how to construct prisons properly. Nevertheless, there are lots of media reporting on this issue that take the limp-wristed approach. Sonnet has its biases, nevertheless I remain convinced that it and extended research is useful.
I'm not a subject matter expert in a wide range of domains, so I can't do that. That's literally what I said. I can observe it makes plenty of errors or is actively misleading in areas I know lots about but I can't show that's representative. This is why Gell-Man Amnesia is a thing.
Did you notice that the the page you linked is about airport security and not shoplifting? Hmmm maybe you screwed up somewhere, or maybe AIslop misled you. Maybe you should just admit you're totally wrong.
It's near but below the record. Since the number is actually falling on a rolling average, the AI is quite misleading. Even if you were to accept that the claim is right, you know what they say about a broken clock. The data is past the training cutoff, and there's no source that supports the information, so the AI can't have legitimately known that the number was near the record or not. So in that case it's a hallucination.
Yet somehow I can pick out errors on half of the AI slop people like you slop out all the time.
Then pick out 10 articles of your choice in an area you know about.
I'm surprised to find that there are two Project Pegasuses but I observe that the anti-theft Pegasus is a part of Opal, who are also still continuing their work.
https://committees.parliament.uk/oralevidence/14920/html/
https://www.npcc.police.uk/our-work/work-of-npcc-committees/Crime-Operations-coordination-committee/opal/
So even if my link was wrong, my point still stands. Pegasus is still a thing in practical terms. You are the one who produced the idea that it had shut down, seemingly from nowhere. What source did that come from? How can you legitimately have known this info?
But theft is at record levels? What, we have to wait for the nano-top or regurgitate secondary sources like wikipedia? Sonnet could easily observe 'ok I know about past historical theft levels, this is higher therefore its at record highs'.
When I see some bad journalism I don't add it to a big list of bad articles, same with spelling errors tbh. But you can take your pick from Russiagate, spruking the case for the invasion of Iraq or suppressing the Hunter Biden laptop story, or this euphemism treadmill where journalists eagerly create a racial narrative if a white does something bad to blacks, whereas they bury the reverse case, mentioning race only at the very end of the article. Those are cases of deception and misleading news from 'real journalists'.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
The first link in the results Claude found is to the story "Shoplifting in England and Wales soars to highest since police records began", whose text reiterates "figures are the highest since current police recording practices began in March 2003."
Weirdly, Claude doesn't seem to be having any luck finding BBC results for its queries - e.g. "site:bbc.co.uk shoplifting uk 2025 - 0 results" - but when I try the same search it did, my first hit is to the BBC story "Shoplifting hits record high in England and Wales", with text like "at its highest level since current records began more than two decades ago" and a graph showing those levels.
Oops I didn't relaize claude share doesn't share inline citations. But the ones in the inline citations definitely did not support the statement
The link you shared is about May 2025 which is not related to the result for June 2025 my claude chat is about. That article says that offenses for YE (year ending in?) May 2025 increased 20% y/y, which is a record. But Claude said that offenses for YE June 2025 increased 13% y/y (correct, and also less than 20) and said that it was a record (which wrong).
Anyways it's undeniable that your favorite model still slopped out a multitude of errors on literally the first query ai I tried on it.
That's on me, too; I should have checked the links in your quotes, not just looked at the Claude transcript and assumed it included everything in the quotes.
One of the two links I shared was an April story, the other a July story; both were data through March 2025.
Personally I'd have used the phrase "near-record levels" (after rising 30+% above trend, it dropped back 0.13% - yay?), but I'm not sure that'd be any more informative a summary - "near-" could be applied just as well to a record set 13 years earlier, while "representing" is a closer fit for 3 months earlier. "Reached record levels" or "was a record" wouldn't be supported by Claude's inline link, but both of those were your rewording, not Claude's.
You seem to have confused me with @RandomRanger. Claude is my second-favorite model, because while I've repeatedly caught it in errors, it at least always tries to fix them when I correct it; ChatGPT-5-Thinking is the only thing I've seen that's (so far, for me; others have had worse luck) been good about avoiding errors preemptively, and IIRC all the non-Claude free models I've tried have made significant errors and often tried to gaslight me about them afterward.
I'm not entirely on board with Claude claiming that 99.8% of a recent record is "representing" that record, but it's clearly all too easy to slop out errors. Would that either of us were under 0.2% off!
Looking at your other complaints, they're mostly either not errors or not clearly errors, which amusingly means that appellation is itself in error each of those times:
When Claude refers to "Operation Pegasus", that's a term even the BBC has used, referring to the same thing as "Project Pegasus", though it's not used in the story at that particular inline link, which is about details other than terminology variants. (it is in one of the other links Claude found) When Claude is correct about something that seems too simple to justify, but it turns out that "too simple" is in the eye of the beholder, that's still not an error.
The difference between "Wrong" and "There's no citation" also applies to the Crime and Policing Bill - is it wrong? Then what is the primary response to the problem? Four out of the five quoted sources in the linked article mention the Crime and Policing Bill by name, which seems to be a solid first place showing; why would we not want AI to use Grice's Maxims here?
When you say "The source does not indicate that any mapping of what's happening was done at the summit.", you're misparaphrasing Claude's summary, which says "coordinate efforts on mapping", and is actually a pretty good abridgement of "see what more we can do together to map what's happening" from the source article.
Your claim of "outdated" is like something out of a South Park joke. 2023! The Before Times! The Long Long Ago! It's good to see an October 23 2025 article in the mix too, but I want citations that provide a little context; "born yesterday" is supposed to be an insult! Perhaps at some age "outdated" becomes "unsupported", but that's still not "erroneous" - is the data actually out of date? Which of those policies has since ended?
Ironically, the one thing I've seen change most since 2023 is AI itself. In 2023 I was giving AIs benchmark questions that could be answered by most first-year grad students in my field, watching them instead make sign errors that could have been caught by anyone who's passed Calc 3, and then watching the various models either flail about at failures to fix the problem or gaslight me about there not being a problem to fix. In 2025 I can still catch the free models in math errors, but the one time I've "caught" a top model it turned out to be because I had an embarrassing typo in my own notes. Actual top-of-their-field geniuses are still catching top models in math errors ... but using them to prove theorems anyway, with reports to the effect that it's faster to try new ideas out with the models and correct the errors than it is to try every idea out manually.
I do like talking to Claude, at least for anything where I can double-check its work, both because it's capable of avoiding rude language like "slop" and "dogshit" and "shitty", and because when I do find errors upon double-checking, it acknowledges and tries to fix them. You've been pretty good about the latter so far, at least; thank you!
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link