Ars Technica being caught using LLMs that hallucinated quotes by the author and then publishing them in their coverage about this is quite ironic here.
Even on a forum where I saw the original article by this author posted someone used an LLM to summarize the piece without having read it fully themselves.
How many levels of outsourcing thinking is occurring to where it becomes a game of telephone.
Also ironic: When the same professionals advocating "don't look at the code anymore" and "it's just the next level of abstraction" respond with outrage to a journalist giving them an unchecked article.
Read through the comments here and mentally replace "journalist" with "developer" and wonder about the standards and expectations in play.
Food for thought on whether the users who rely on our software might feel similarly.
There's many places to take this line of thinking to, e.g. one argument would be "well, we pay journalists precisely because we expect them to check" or "in engineering we have test-suites and can test deterministically", but I'm not sure if any of them hold up. The "the market pays for the checking" might also be true for developers reviewing AI code at some point, and those test-suites increasingly get vibed and only checked empirically, too.
- There’s a difference. Users don’t see code, only its output. Writing is “the output”.
- A rough equivalent here would be Windows shipping an update that bricks your PC or one of its basic features, which draws plenty of outrage. In both cases, the vendor shipped a critical flaw to production: factual correctness is crucial in journalism, and a quote is one of the worst things to get factually incorrect because it’s so unambiguous (inexcusable) and misrepresents who’s quoted (personal).
I’m 100% ok with journalists using AI as long as their articles are good, which at minimum requires factual correctness and not vacuous. Likewise, I’m 100% ok with developers using AI as long as their programs are good, which at minimum requires decent UX and no major bugs.
> - There’s a difference. Users don’t see code, only its output. Writing is “the output”.
So how is the "output" checked then? Part of the assumption of the necessity of code review in the first place is that we can't actually empirically test everything we need to. If the software will programmatically delete the entire database next Wednesday, there is no way to test for that in advance. You would have to see it in the code.
Tbf I'm fine with it only one way around; if a journalist has tonnes of notes and data on a subject and wants help to condense those down into an article, assistance with prioritising which bits of information to present to the reader then totally fine.
If a journalist has little information and uses an llm to make "something from nothing" that's when I take issue because like, what's the point?
Same thing as when I see managers dumping giant "Let's go team!!! 11" messages splattered with AI emoji diarrhea like sprinkles on brown frosting. I ain't reading that shit; could've been a one liner.
Another good use of an LLM is to find primary sources.
Even an (unreliable) LLM overview can be useful, as long as you check all facts with real sources, because it can give the framing necessary to understand the subject. For example, asking an LLM to explain some terminology that a source is using.
Excellent observation. I get so frustrated every time I hear the "we have test-suites and can test deterministically" argument. Have we learned absolutely nothing from the last 40 years of computer science? Testing does not prove the absence of bugs.
I don't know if I look forward to it, myself, but yeah: I can imagine a future where in person interactions become preferred again because at least you trust the other person is human. Until that also stops being true, I guess.
Well, I can tell you I've been reading a lot more books now. Ones published before the 2020s, or if recent, written by authors who were well established before then.
Because nobody ever lied in print media or in person?
What we are seeing is the consequence of a formerly high trust society collapsing into a low trust one. There is no place to hide from that. The Internet is made of the same stuff as print media and in person. It’s made of people.
The internet didn’t cause this. It just reflects it.
The LLMs are made of people too inasmuch as that’s where they get their training data and prompts.
> When the same professionals advocating "don't look at the code anymore" and "it's just the next level of abstraction" respond with outrage to a journalist giving them an unchecked article.
I would expect there is literally zero overlap between the "professionals"[1] who say "don't look at the code" and the ones criticising the "journalists"[2]. The former group tend to be maximalists and would likely cheer on the usage of LLMs to replace the work of the latter group, consequences be damned.
[1] The people that say this are not professional software developers, by the way. I still have not seen a single case of any vibe coder who makes useful software suitable for deployment at scale. If they make money, it is by grifting and acting as an "AI influencer", for instance Yegge shilling his memecoin for hundreds of thousands of dollars before it was rugpulled.
[2] Somebody who prompts an LLM to produce an article and does not even so much as fact-check the quotations it produces can clearly not be described as a journalist, either.
While I don't subscribe to the idea that you shouldn't look at the code - it's a lot more plausible for devs because you do actually have ways to validate the code without looking at it.
E.g you technically don't need to look at the code if it's frontend code and part of the product is a e2e test which produces a video of the correct/full behavior via playwright or similar.
Same with backend implementations which have instrumentation which expose enough tracing information to determine if the expected modules were encountered etc
I wouldn't want to work with coworkers which actually think that's a good idea though
You might notice that these real engineering jobs also don't have a way to verify the product via tests like that though, which was my point.
And that's ignoring that your statement technically isn't even true, because the engineers actually working in such fields are very few (i.e. designing bridges, airplanes etc).
The majority of them design products where safety isn't nearly as high stakes as that... And they frequently do overspec (wasting money) or underspec (increasing wastage) to boot.
This point has been severely overstated on HN, honestly.
> You might notice that these real engineering jobs also don't have a way to verify the product via tests like that though, which was my point.
The electrical engineers at my employer that design building electrical distribution systems have software that handles all of the calculations, it’s just math. Arc flash hazard analysis, breaker coordination studies, available fault current, etc. All manufacturers provide the data needed to perform these calculations for their products.
Other engineering disciplines have similar tools. Mechanical, civil, and structural engineers all use software that simulates their designs.
Okay, honestly asking: do you really think your outlined simulation meet the same bar as automated regression tests like
> E.g you technically don't need to look at the code if it's frontend code and part of the product is a e2e test which produces a video of the correct/full behavior via playwright or similar.
> Same with backend implementations which have instrumentation which expose enough tracing information to determine if the expected modules were encountered etc
If not, your quoted sentence was "tests like that". If yes... I guess we would have to disagree.
> You might notice that these real engineering jobs also don't have a way to verify the product via tests though, which was my point.
Are you sure? Simulators and prototypes abound. By the time you’re building the real, it’s more like rehearsal and solving a fe problems instead of every intricacy in the formula.
> When the same professionals advocating "don't look at the code anymore" and "it's just the next level of abstraction" respond with outrage to a journalist giving them an unchecked article.
I doubt, by and large, that it's the same people. Just as this LLM misquoting is journalistic malpractice, "don't look at the code anymore" is engineering malpractice.
Incredible. When Ars pull an article and its comments, they wipe the public XenForo forum thread too, but Scott's post there was archived. Username scottshambaugh:
>Scott Shambaugh here. None of the quotes you attribute to me in the second half of the article are accurate, and do not exist at the source you link. It appears that they themselves are AI hallucinations. The irony here is fantastic.
Instead of cross-checking the fake quotes against the source material, some proud Ars Subscriptors proceed to defend Condé Nast by accusing Scott of being a bot and/or fake account.
EDIT: Page 2 of the forum thread is archived too. This poster spoke too soon:
>Obviously this is massive breach of trust if true and I will likely end my pro sub if this isnt handled well but to the credit of ARS, having this comment section at all is what allows something like this to surface. So kudos on keeping this chat around.
There are some various attempts, the problem is reliability - not that they're always up, but how do you trust them? If archive.org shows a page at a date, you presume it is true and correct. If I provide a PDF of a site at a date, you have no reason to believe I didn't modify the content before PDFing it.
Aurich Lawson (creative director at Ars) posted a comment[0] in response to a thread about what happened, the article has been pulled and they'll follow-up next week.
Just like in the original thread that was wiped (https://news.ycombinator.com/item?id=47012384), Ars Subscriptors continue to display lack of reading comprehension and jump to defending Condé Nast.
Yikes I subscribed to them last year on the strength of their reporting in a time where it's hard to find good information.
Printing hallucinated quotes is a huge shock to their credibility, AI or not. Their credibility was already building up after one of their long time contributors, a complete troll of a person that was a poison on their forums, went to prison for either pedophilia or soliciting sex from a minor.
Some serious poor character judgement is going on over there. With all their fantastic reporters I hope the editors explain this carefully.
TBF even journalists who interview people for real and take notes routinely quite them saying things they didn't say. The LLMs make it worse, but it's hardly surprising behaviour from them
I knew first hand about a couple of news in my life. Both were reported quite incorrectly. That was well before LLMs. I assume that every news is quite inaccurate, so I read/hear them to get the general gist of what happened, then I research the details if I care about them.
It's surprising behavior to come from Ars Technica. But also when journalists misquote it's through a different phrasing of something that Pepe have actually said, sometimes with different emphasis or eve meaning. But of the people I've known who have been misquoted it's always traceable to something they actually did say.
Humans aren't very diligent in the long term. If an LLM does something correctly enough times in a row (or close enough), humans are likely to stop checking its work throughly enough.
This isn't exactly a new problem we do it with any bit of new software/hardware, not just LLMs. We check its work when it's new, and then tend to trust it over time as it proves itself.
But it seems to be hitting us worse with LLMs, as they are less consistent than previous software. And LLM hallucinations are partially dangerous, because they are often plausible enough to pass the sniff test. We just aren't used to handling something this unpredictable.
This is a first degree expectation of most businesses.
What the OP pointed out is a fact of life.
We do many things to ensure that humans don’t get “routine fatigue”- like pointing at each item before a train leaves the station to ensure you don’t eyes glaze over during your safety check list.
This isn’t an excuse for the behavior. Its more about what the problem is and what a corresponding fix should address.
I agree. The role of an editor is in part to do this train pointing.
I think it slips because the consequences of sloppy journalism aren’t immediately felt. But as we’re witnessing in the U.S., a long decay of journalistic integrity contributes to tremendous harm.
It used to be that to be a “journalist” was a sacred responsibility. A member of the Fourth Estate, who must endeavour to maintain the confidence of the people.
And too easy on the editor who was supposed to personally verify that the article was properly sourced prior to publication. This is like basic stuff that you learn working on a high school newspaper.
The words on the page are just a medium to sell ads. If shit gets ad views then producing shit is part of the job... unless you're the one stepping up to cut the checks.
There's a weird inconsistency among the more pro-AI people that they expect this output to pass as human, but then don't give it the review that an outsourced human would get.
The irony is that while from perfect, an LLM-based fact-checking agent is likely to be far more dilligent (but still needs human review as well) by nature of being trivial to ensure it has no memory of having done a long list of them (if you pass e.g. Claude a long list directly in the same context, it is prone to deciding the task is "tedious" and starting to take shortcuts).
But at the same time, doing that makes it even more likely the human in the loop will get sloppy, because there'll be even fewer cases where their input is actually needed.
I'm wondering if you need to start inserting intentional canaries to validate if humans are actually doing sufficiently torough reviews.
The kind of people to use LLM to write news article for them tend not to be the people who care about mundane things like reading sources or ensuring what they write has any resemblance to the truth.
The problem is that the LLM's sources can be LLM generated. I was looking up some health question and tried clicking to see the source for one of the LLMs claim. The source was a blog post that contained an obvious hallucination or false elaboration.
It’s fascinating that on the one hand Ars Technica didn’t think the article was worth writing (so got an LLM to do it) but expect us to think it’s worth reading. Then some people don’t think it’s worth reading (so get an LLM to do it) but think somehow we will think it’s not worth reading the article but is worth reading the llm summary. Feel like you can carry on that process ad infinitum always going for a smaller and smaller audience who are somehow willing to spend less and less effort (but not zero).
> How many levels of outsourcing thinking is occurring to where it becomes a game of telephone
How do you know quantum physics is real? Or radio waves? Or just health advice? We don't. We outsource our thinking around it to someone we trust, because thinking about everything to its root source would leave us paralyzed.
Most people seem to have never thought about the nature of truth and reality, and AI is giving them a wake-up call. Not to worry though. In 10 years everyone will take all this for granted, the way they take all the rest of the insanity of reality for granted.
I used to be skeptical that AI generated text could be reliably detected, but after a couple years of reading it, there are cracks starting to form in that skepticism.
How could that prove hallucinations? It could only possibly prove that they are not. If the quotes are in the original post then they are not hallucinations. If they are not in the post they could be caused by something is not a LLM.
Misquotes and fabricated quotes have existed long before AI, And indeed, long before computers.
My claim is as it has always been. If we accept that the misquotes exist it does not follow that they were caused by hallucinations? To tell that we would still need additional evidence. The logical thing to ask would be; Has it been shown or admitted that the quotes were hallucinations?
I think you're missing their point. The question you're replying to is, how do we know that this made up content is a hallucination. Ie., as opposed to being made up by a human. I think it's fairly obvious via Occam's Razor, but still, they're not claiming the quotes could be legit.
You seem to be quite certain that I had not read the article, yet I distinctly remember doing do.
By what proceess do you imagine I arrived at the conclusion that the article suggested that published quotes were LLM hallucinations when that was not mentioned in the article title?
You accuse me of performative skepticism, yet all I think is that it is better to have evidence over assumptions, and it is better to ask if that evidence exists.
It seems a much better approach than making false accusations based upon your own vibes, I don't think Scott Shambaugh went to that level though.
>On Friday afternoon, Ars Technica published an article containing fabricated quotations generated by an AI tool and attributed to a source who did not say them. That is a serious failure of our standards. Direct quotations must always reflect what a source actually said.
There is a third option:
The journalist who wrote the article made the quotes up without an LLM.
I think calling the incorrect output of an LLM a “hallucination” is too kind on the companies creating these models even if it’s technically accurate. “Being lied to” would be more accurate as a description for how the end user feels.
The journalist was almost certainly using an LLM, and a cheap one at that. The quote reads as if the model was instructed to build a quote solely using its context window.
Lying is deliberately deceiving, but yeah, to a reader, who in a effect is a trusting customer who pays with part of their attention diverted to advertising support, broadcasting a hallucination is essentially the same thing.
Ironically, if you actually know what you’re doing with an LLM, getting a separate process to check the quotations are accurate isn’t even that hard. Not 100% foolproof, because LLM, but way better than the current process of asking ChatGPT to write something for you and then never reading it before publication.
The wrinkle in this case is the author blocked AI bots from their site (doesn't seem to be a mere robots.txt exclusion from what I can tell), so if any such bot were trying to do this it may have not been able to read the page to verify, so instead made up the quotes.
This is what the author actually speculated may have occurred with Ars. Clearly something was lacking in the editorial process though that such things weren't human verified either way.
More than ironic, it's truly outrageous, especially given the site's recent propensity for negativity towards AI. They've been caught red-handed here doing the very things they routinely criticize others for.
The right thing to do would be a mea-culpa style post and explain what went wrong, but I suspect the article will simply remain taken down and Ars will pretend this never happened.
I loved Ars in the early years, but I'd argue since the Conde Nast acquisition in 2008 the site has been a shadow of its former self for a long time, trading on a formerly trusted brand name that recent iterations simply don't live up to anymore.
Is there anything like a replacement? The three biggest tech sites that I traditionally love are ArsTechnica, AnandTech(rip), and Phoronix. One is dead man walking mode, the second is ded dead, and the last is still going strong.
I'm basically getting tech news from social media sites now and I don't like that.
In my wildest hopes for a positive future, I hope disenchanted engineers will see things like this as an opportunity to start our own companies founded on ideals of honesty, integrity, and putting people above profits.
I think there are enough of us who are hungry for this, both as creators and consumers. To make goods and services that are truly what people want.
Maybe the AI revolution will spark a backlash that will lead to a new economy with new values. Sustainable business which don't need to squeeze their customers for every last penny of revenue. Which are happy to reinvest their profits into their products and employees.
Conde Nast are the same people wearing Wired magazine like a skin suit, publishing cringe content that would have brought mortal shame upon the old Wired.
I don't read their comment as implying this. It might in fact hint at the opposite; it's far more likely for the less senior author to get thrown under the bus, regardless of who was lazy.
Scapegoats are scapegoats but in every organization the problems are ultimately caused by their leaders. It's what they request or what they fail to request and what they lack to control.
There is no need to rush to judgment on the internet instant-gratification timescale. If consequences are coming for journalist or publication, they are inevitable.
We’ll know more in only a couple days — how about we wait that long before administering punishment?
It's not rushing to judgement, the judgement has been made. They published fraudulent quotes. Bubbling that liability up to Arse Technica is valuable for punishing them too but the journalist is ultimately responsible for what they publish too. There's no reason for any publication to ever hire them again when you can hire ChatGPT to lie for you.
EDIT: And there's no plausible deniability for this like there is for typos, or maligned sources. Nobody typed these quotes out and went "oops, that's not what Scott said". Benj Edwards or Kyle Orland pulled the lever on the bullshit slot machine and attacked someone's integrity with the result.
"In the past, though, the threat of anonymous drive-by character assassination at least required a human to be behind the attack. Now, the potential exists for AI-generated invective to infect your online footprint."
We do not yet know just how the story unfolded between the two people listed on the byline. Consider the possibility that one author fabricated the quotes without the knowledge of the other. The sin of inadequate paranoia about a deceptive colleague is not the same weight as the sin of deception.
Now to be clear, that’s a hypothetical and who knows what the actual story is — but whatever it is, it will emerge in mere days. I can wait that long before throwing away two lives, even if you can’t.
> Bubbling that liability up to Arse Technica is valuable for punishing them
Evaluating whether Ars Technica establishes credible accountability mechanisms, such as hiring an Ombud, is at least as important as punishing individuals.
That's what bylines are for, though. Both authors are attributed, and are therefore both responsible. If they didn't both review the article before submitting that's their problem. It's exaggerating to call this throwing away two lives, if all they do for a living is hit the big green button on crap journalism then I'm fine with them re-skilling to something less detrimental.
I agree that reserving judgement and separating the roles of individuals from the response of the organization are all critical here. Its not the first time that one of their staff were found to have behaved badly, in the case that jumps to my mind from a few years ago Peter Bright was sentenced to 12 years on sex charges involving a minor1. So, sometimes people do bad things, commit crimes, etc. but this may or may not have much to do with their employer.
Did Ars respond in any way after the conviction of their ex-writer? Better vetting of their hires might have been a response. Apparently there was a record of some questionable opinions held by the ex-writer. I don't know, personally, if any of their policies changed.
The current suspected bad behavior involved the possibility that the journalists were lacking integrity in their jobs. So if this possibility is confirmed I expect to see publicly announced structural changes in the editorial process at Ars Technica if I am to continue to be a subscriber and reader.
I mean, I'm even more frustrated by this in Scott's original post:
> If you are the person who deployed this agent, please reach out. It’s important for us to understand this failure mode, and to that end we need to know what model this was running on and what was in the soul document. I’m not upset and you can contact me anonymously if you’d like.
I can see where he's coming from, and I suppose he's being the bigger man in the situation, but at some point one of these reckless moltbrain kiddies is going to have to pay. Libel and extortion should carry penalties no matter whether you do it directly, or via code that you wrote, or via code that you deployed without reading it.
The AI's hit piece on Scott was pretty minor, so if we want to wait around for a more serious injury that's fine, just as long as we're standing ready to prosecute when (not 'if') it happens.
I just wish people would remember how awful and unprofessional and lazy most "journalists" are in 2026.
It's a slop job now.
Ars Technica, a supposedly reputable institution, has no editorial review. No checks. Just a lazy slop cannon journalist prompting an LLM to research and write articles for her.
Ask yourself if you think it's much different at other publications.
I work with the journalists at a local (state-wide) public media organization. It's night and day different from what is described at ars. These are people who are paid a third (or less) of what a sales engineer at meta makes. We have editorial review and ban LLMs for any editorial work except maybe alt-text if I can convince them to use it. They're over-worked, underpaid, and doing what very few people here (including me) have the dedication to do. But hey, if people didn't hate journalists they wouldn't be doing their job.
Even on a forum where I saw the original article by this author posted someone used an LLM to summarize the piece without having read it fully themselves.
How many levels of outsourcing thinking is occurring to where it becomes a game of telephone.