The current title (“Pakistani newspaper mistakenly prints AI prompt with the article”) isn’t correct, it wasn’t the prompt that was printed, but trailing chatbot fluff:
> If you want, I can also create an even snappier “front-page style” version with punchy one-line stats and a bold, infographic-ready layout—perfect for maximum reader impact. Do you want me to do that next?
The article in question is titled “Auto sales rev up in October” and is an exceedingly dry slab of statistic-laden prose, of the sort that LLMs love to err in (though there’s no indication of whether they have or not), and for which alternative (non-prose) presentations can be drastically better. Honestly, if the entire thing came from “here’s tabular data, select insights and churn out prose”… I can understand not wanting to do such drudgework.
The newspaper in question is Pakistan's English language "newspaper of record", which has wide readership.
For some reason, they rarely ever add any graphs or tables to financial articles, which I have never understood. Their readership is all college educated. One time I read an Op-Ed, where the author wrote something like: If you go to this gov webpage, and take the data and put it on excel, and plot this thing vs that thing, you will see X trend.
Why would they not just take the excel graph, clean it up and put it in their article?
Gemini has been doing this to me for the past few weeks at the end of basically every single response now, and it often seems to result in the subsequent responses getting off track and lower quality as all these extra tangets start polluting the context. Not to mention how distracting it is as it throws off the reply I was already halfway in the middle of composing by the time I read it.
In similar "news" Re: tangents, I've noticed Claude now suddenly starting to give up on problem analyses. After like three rounds of it trying to figure something out that I told it is a requirement, it suggests that we don't need to do that as it's a nice-to-have only anyway. If in auto-accept mode it'll start doing the other thing even. I have to catch it and tell it to not effing give up so easily and to stop giving me BS excuses turning hard requirements (literally the reason we're doing what we're doing) into a nice-to-have it can skip.
I guess they recently re-trained on too many "perpetual junior senior dev" stuff.
Add "Complete this request as a single task and do not ask any follow-up questions." Or some variation of that. They keep screwing with default behavior, but you can explicitly direct the LLM to override it.
This is why I wish chat UI's had separate categories of chats (like a few generic system prompts) that let you do more back-and-forth style discussions, or more "answers only" without adding any extra noise, or even an "exploration"/"tangent" slider.
The fact that system prompts / custom instructions have to be typed-in in every major LM chat UI is a missed opportunity IMO
You can if you script the request yourself, or you could have a front end that lets you cut out those paragraphs from the conversation. I only say that because yesterday I followed this guide: https://fly.io/blog/everyone-write-an-agent/ except I had to figure out how to do it with Gemini API instead. The context is always just (essentially) a list of strings (or "parts" anyway, doesn't have to be strings) that you pass back to the model so you can make the context whatever you like. It shouldn't be too hard to make a frontend that lets you edit the context, and fairly easy to mock up if you just put the request in a script that you add to.
I think AI should present those continuation prompts as dynamic buttons, like "Summarize", "Yes, explain more" etc. based on the AI's last message, like the NPC conversation dialogs in some RPGs
For years, both the financial and sports news sides of things have generated increasingly templated "articles", this just feels like the latest iteration.
This dates back to at least the late 1990s for financial reports. A friend demoed such a system to me at that time.
Much statistically-based news (finance, business reports, weather, sport, disasters, astronomical events) are heavily formulaic and can at least in large part or initial report be automated, which speeds information dissemination.
Of course, it's also possible to distribute raw data tables, charts, or maps, which ... mainstream news organisations seem phenomenally averse to doing. Even "better" business-heavy publications (FT, Economist, Bloomberg, WSJ) do so quite sparingly.
A few days ago I was looking at a Reuters report on a strategic chokepoint north of the Philippines which it and the US are looking toward to help contain possible Chinese naval operations. Lots of pictures of various equipment, landscapes, and people. Zero maps. Am disappoint.
I believe the word is depression, which seems apt when thinking of the idea of people using AI to make content longer and then the readers all using AI to make it shorter again.
But there's the approach the Economist takes. For many decades, it's relied on a three-legged revenue model: subscriptions, advertising, and bespoke consulting and research through the Economist Intelligence Unit (EIU). My understanding is that revenues are split roughly evenly amongst these, and that they tend to even out cash-flow throughout economic cycles (advertising is famously pro-cyclical, subscriptions and analysis somewhat less so).
To that extent, the graphs and maps the Economist actually does include in its articles (as well as many of its "special reports") are both teasers and loss-leader marketing for EIU services. I believe that many of the special reports arise out of EIU research.
...
The rules for the race: Both contenders waited for Denny's, the diner company, to come out with an earnings report. Once that was released, the stopwatch started. Both wrote a short radio story and get graded on speed and style.
StatSheet, an online platform covering college basketball, runs entirely on an automated program. In 2006, Thomson Reuters announced their switch to automation to generate financial news stories on its online news platform. Reuters used a tool called Tracer. An algorithm called Quakebot published a story about a 2014 California earthquake on The Los Angeles Times website within three minutes after the shaking had stopped.
Sports and financial are the two easiest to do since they both work from well structured numeric statistics.
> Quakebot is a software application developed by the Los Angeles Times to report the latest earthquakes as fast as possible. The computer program reviews earthquake notices from the U.S. Geological Survey and, if they meet certain criteria, automatically generates a draft article. The newsroom is alerted and, if a Times editor determines the post is newsworthy, the report is published.
> The computer program reviews earthquake notices from the U.S. Geological Survey
Probably a service that is provided to the general public for free, similar to NOAA and weather data - so chances are rather high it ends up on the chopping block or for-money only.
In the mid-late naughts, there used to be a content farm called "Associated Content". They would get daily lists of top searched terms from various search engines (Yahoo, Dogpile, Altavista, etc. etc.) and for each search term, pay an English major to write a 2-page fluff article. Regardless of what the topic was, they churned out articles by the bushel. Then they place ads on these articles and sat back and watched the dollars roll in.
A non-"AI" template is probably getting filled in with numbers straight from some relevant source. AI may produce something more conversational today but as someone else observed, this is a high-hallucination point for them. Even if they get one statistic right they're pretty inclined to start making up statistics that weren't provided to them at all if they sound good.
Not just that we know from heavy reddit posters that they have branching universe templates for all eventualities, so that they are "ready" whatever the outcome.
I guess in the end the journalist didn't feel necessary to impact his readers with punchy one line stats and bold infographic-ready layouts, considering he opted for the first draft.
Nobody outside Pakistan knows Dawn even though it is the newspaper that was founded by Muhammad Ali Jinnah (considered founding father of the nation) and one of the largest and most prestigious as well.
It is like the NYT for the country. But the relevant detail here is the printing of the prompt in a nationally recognized newspaper. The brand, as local as it maybe, still provides more context than some random newspaper in a foreign country would.
And I have ran into Dawn newspaper on google news frontpage several times, usually on entertainment stuff.
Do we know it was an AI? I realize that it rings with a sycophantic tone that the AIs love to use, but I've worked with some humans who speak the same way. AIs didn't invent brownnosing.
When reached for comment on how this occurred, the journalist in question replied:
“This is the perfect question that gets to the heart of this issue. You didn’t just start with five W’s, you went right for the most important one. Let’s examine why that question works so well in this instance…”
You're absolutely right! but they can shove this euphemism. Just say that chatgpt wrote the article and no one read it before publishing, no need for all the fluff.
>> Just say that chatgpt wrote the article and no one read it before publishing
This is so interesting. I wonder if no human prompted for the article to be written either. I could see some kind of algorithm figuring out what to "write" about and prompting AI to create the articles automatically. Those are the jobs that are actually being replaced by AI - writing fluff crap to build an attention trap for ad revenue.
Very likely this already happens on slop websites (...which I can't name because I don't go there), which for example just republish press releases (which could be considered aggregation sites I guess), or which automatically scrape Reddit and translate them into listicles on the fly.
They do not deserve a shred of recommendation. This is just damage control, pretending that it did not happen never was an option. Instead they tried to claim that it was just a one of mistake. What it really shows is that nobody even bothers to read their articles before hitting publish and that AI is widely used internally.
Fair play to them for owning up to their mistake, and not just pretending like it didn't happen!
That's what the legitimate media has done for the last couple of hundred years. Every issue of the New York Times has a Corrections section. I think the Washington Post's is called Corrections and Amplifications.
Bloggers just change the article and hope it didn't get cached in the Wayback Machine.
The editors were laid off and replaced by an LLM. Or more likely, the editorial staff was cut in half and the ones who were kept were told to use LLMs to handle the increased workload.
"This newspaper report was originally edited using AI, which is in violation of Dawn’s current AI policy. The policy is also available on our website. The report also carried some junk, which has now been edited out. The matter is being investigated. The violation of AI policy is regretted. — Editor"
edit: Text link of the printed edition. Might not be perfect OCR, but I don't think they changed anything except to delete the AI comment at the end! https://pastebin.com/NYarkbwm
That's just a manner of speaking in former British colonies, or at least the subcontinent. Much of formal speech like a bureaucrat wrote it because, well, the civil service ran India and that's who everyone emulated.
This pattern of writing goes back to the Spanish conquistadors at the very least. They frequently described their actions in a passive voice when doing something they knew was horrible, only to switch to aggrandizing active voice when writing about their successes. It’s a standard way to blur responsibility and present violence as an almost natural “fact” rather than a deliberate action by identifiable agents.
It didn’t escape everyone’s attention though. Bartolomé de las Casas definitely noticed it.
OTOH kudos to them for regretting AI slop (even if they don't want to point out who precisely is regretting). I know some who'd vehemently deny in spite of evidence.
Of course, since we live in 1984 already everything is edited as is convenient. For all that technology has given, nobody talks about what it has taken away.
Which raises the question: if everything is generated, why bother reading it at all?
Just ask the LLM what you want to know—why treat headlines like bookmarks?
My experience, AI has shown me that a lot of stuff I do online. Watching videos, reading random articles, is mostly vapid pointless nonsense.
AI slop has finally woken me up and I am prioritizing IRL activities with friends and family, or activities more in the real world like miniature painting, piano, etc. It's made me much more selective with what activities I engage in, because getting sucked in to the endless AI miasma isn't what I want for my life.
You can use the LLM, but you don't also have the rest of the data they relied on. A LLM can generate everything if it starts from a minimal prompt, but this is a recipe for slop. If you come with materials, discuss them, their implications, express your POV and then generate, the article will reflect your ideas and the data if was fed.
I know it is fashionable to put everything a LLM outputs in the slop box, but I don't think it reflects reality.
As much as the default LLMisms are annoying me, it's also a honeymoon period right now where you can even suspect whether something is AI generated based on the default LLM-isms. Word about how to fix their tone has been getting around in academia for a while amongst students trying to pass detection filters, once they're out into the world we can expect to have even more AI generated content masked behind individualized, unique style prompts that aren't immediately recognizable as the default LLM voice.
I get that transforming a bunch of facts into prose is boring.
As a reader, I can't get over the fact that I'm supposed to read a text that nobody could be bothered to write.
I wonder how often we waste energy nowadays by telling an AI to turn a one-sentence-prompt into a full e-mail, only for the other side to tell the AI to summarize it in one sentence.
I think better to put that someone extra further up in the pipeline who knows how to prompt the LLM correctly so that it doesn't generate the fluff to begin with.
Or get software engineers to produce domain specific tooling rather than the domain relying on generic tooling which lead to such mistakes (although this is speculation.. but still to me it seems like the author of that article was using the vanilla ChatGPT client)
/s I am now thinking of setting up an "AI Consultancy" which will be able to provide both these resources to those seeking such services. I mean, why have only one of those when both are available.
"This article will be posted on our prestigious news site. Our readers don't know that most of our content is AI slop that our 'writers' didn't even glance over once, so please check if you find anything that was left over from the LLM conversation and should not be left in the article. If you find anything that shouldn't stay in the article, please remove it. Don't say 'done' and don't add your own notes or comment, don't start a conversation with me, just return the cleaned up article."
And someone will put "Prompt Engineer" in their resume.
Considering there’s lawyers risking their careers by using AI, I think the lesson here is that if you allow people to be lazy they will. Humans are built for efficiency.
If a beginner writer thinks AI can write a better article than they can, it seems like they’ll just rely on the AI and never hone their craft.
This mostly harms written journalism. As people seek humanness in their media, they'll be driven even more into the dens of cult-leaders masquerading as podcaster-journalists. The media environment is becoming so terminally awful, and each year it keeps getting worse, for decades now.
So is the documentation and specs that I'm provided by stakeholders...but I can't prove it..
Suddenly they write very long and detailed documentation, yet they can't remember what's in it.
My gf says that in her banks she suspects half the written communications are ai-authored, which is driving productivity to the ground. Her bank moreover is very aggressive with endless workshops on AI usage (they have some enterprise gemini version).
The more I see the impact of AI, the more worried I am.
I'm not saying it doesn't have use cases, I myself leverage it when coding (albeit I use a man-in-the-middle approach where I ask questions and tinker with it, I never let AI write code except in some very rare boilerplate-y scenarios) and have built products that leverage it.
But it seems like the trend is to increasingly _delegate_ work to it, and I'm seeing more negatives than positives.
This has happened more than once, in different scenarios.
I wonder why is it that GhatGPT (and the rest) don't put the actual response in a box that is separate from the "hello" and "goodbye" part.
I once wrote a message to my landlord, and I asked ChatGPT to help me make the message better (being that english is not my mother tongue) and I included the "goodbye" part by mistake.
LLM's don't have any internal concept of "actual response" vs "surrounding words". Just like they don't have a internal concept of system prompt vs user input. Just like they don't even have an internal concept of what the LLM emitted vs what was given to it! It's all just one long sequence.
(Yes, it is possible to create tokens to represent category changes, but this is still in-band. the token is still just part of the sequence and the LLM isn't guaranteed to factor it in correctly)
My understanding is: the training induces it to put Python code inside Markdown code fences, which the app interprets. It does that because the training data is full of examples of people doing that, because when people post Python code online it's very commonly in contexts where Markdown is understood that way.
... I suppose they could train the LLM to put a Markdown horizontal rule in the right place, but it sounds harder to get the system prompt to care about details like that consistently.
Hmm, for some reason I assumed that printing referred to a figure of speech here (like they went to press or something), but no, it's actually printed, AI call to action included.
I think a brilliant solution for these issues would be to get into the habit of asking the AI to double check the article before the final copy-paste.
As the pioneers in the field say - prompts are as important as output and must be preserved and published along with code/output! Don't throw away your prompts.
The funny thing is not that they are using an agent to fluff up their articles but that no one is doing a final read checking anything before it goes to press.
Could this be a mistake or intentional from a disgruntled writer?
I've been trying to get ChatGPT to stop adding this kind of fluff to its responses through custom instructions, but to no avail! It's one of the more frustrating parts of it, IMO.
Similar thing happened in. Bangladesh. The leading national daily English newspaper printed not the prompt, but included a follow up comment/question from AI.
I think the greater issue here is the possibility of the entire article being AI slop and the 'doomed' direction in which we are heading where reputed sources of information are churning out slop.
Completion API is only half the product. Teams integrate these tools into production with nothing but policy documents between them & incidents. The guardrail system does not exist for a reason. No business model. No investors. No revenue stream.
Tesla didn't just build FSD. They had to build a verification layer: driver cameras, steering sensors, attention alerts. That parallel software makes autopilot road ready.
One of the great advantages of AI for non english native speakers is the ability of the tool to speak in better English than the writer. With so many young journalists graduating from school using AI instead of learning the full language, this use would become more frequent.
At my work place, non native speakers would send me documents for grammatical corrections. They don’t do that anymore! Hoorah!
Great advantages for the writer. It's not a great advantage for the reader. The AI could completely change the meaning of the article and the author would be none the wiser.
In my country, last Sunday there were extraordinary city elections in a major city, and the ruling party lost. Almost all major newspapers carried headlines, all of them speaking that the ruling party suffered a "beating" (paliza). In our colloquial context, using that word is not that common, and it was something akin to the proverbial em dash: a telltale sign of AI slop used by all newspapers. Sad, but true.
What shocks me more than the copy-pasting from ChatGPT (which might actually be more relevant than a journalist paid peanuts) is that it means the articles are never proofread before being published. Even if the article had been written by a human, it would be a disaster and would say a lot about the level of misinformation that can slip through an underpaid newsroom.
I keep seeing those mistakes a lot recently, especially the [insert something here] that is inside a wall of text where the AI is keeping the option to the users to edit!
In 2022, my opinion of journalism was low. Decades of headlines which were objectively false but no retraction, just doubling down on their state propaganda.
There were some papers that I still trusted. Then AI hit journalism with a silly stick and utterly wrecked them all.
Mind you, I love AI. I however can admit that AI seems to have wrecked what was left of journalism.
Actually, at some point, it makes sense to be honest about usage of AI and not feeling to hide that. Just like how food products are expected to print about the ingredients.
One should not feel ashamed to declare the usage of AI, just like you are not ashamed to use a calculator.
I feel like there is a difference here. A calculator has no bias. LLMs do, obviously. News is not the place for bias. Unless the LLM used hallucinated the operator’s intentions, the operator was using the LLM to doctor the article to capture readers not report the news.
TBH, I think that journalists tying themselves into pretzels in an effort to remain unbiased does more damage than the presence of some bias. As a consumer of news, I want journalists to be biased, for example, towards the rule of law, the preservations of institutions, and checks & balances, and even norms.
I don't see why a content can't carry a label saying "AI-generated", or "Reviewed by AI", or "Refined using AI" etc. This allows consumer to consume it with appropriate caution.
> If you want, I can also create an even snappier “front-page style” version with punchy one-line stats and a bold, infographic-ready layout—perfect for maximum reader impact. Do you want me to do that next?
The article in question is titled “Auto sales rev up in October” and is an exceedingly dry slab of statistic-laden prose, of the sort that LLMs love to err in (though there’s no indication of whether they have or not), and for which alternative (non-prose) presentations can be drastically better. Honestly, if the entire thing came from “here’s tabular data, select insights and churn out prose”… I can understand not wanting to do such drudgework.
reply