Podcast: Can AI Save Journalism? (with Peter Stuart)

On this week’s episode of the podcast, I am joined by Peter Stuart, the co-founder of Velora, an AI operating system for specialist publishers. Peter returns to the podcast to discuss the rapidly evolving landscape of AI in journalism and to explore how editorial teams are integrating these tools into their daily operations. Among other things, we cover:

• How AI tools like LLMs can effectively augment journalistic research and fact-checking processes without compromising editorial integrity
• If the increasing efficiency of AI content pipelines will ultimately diminish the perceived value of human-led investigative reporting
• Why some media organizations remain hesitant to integrate AI while others aggressively leverage it for content curation and distribution
• When the line between AI-assisted writing and pure machine generation becomes indistinguishable to the average consumer of digital news
• How AI platforms can help journalists reclaim time for high-value reporting by automating mundane administrative and formatting tasks

Thanks to the sponsors of this week’s episode of the Mobile Dev Memo podcast:

• ⁠⁠INCRMNTAL⁠⁠⁠. True attribution measures incrementality, always on.
• Xsolla⁠. With the Xsolla Web Shop, you can create a direct storefront, cut fees down to as low as 5%, and keep players engaged with bundles, rewards, and analytics.
• ⁠Branch⁠. Branch is an AI-powered MMP, connecting every paid, owned, and organic touchpoint so growth teams can see exactly where to put their dollars to bring users in the door and keep them coming back

Interested in sponsoring the Mobile Dev Memo podcast? Contact Mobile Dev Memo advertising.

The Mobile Dev Memo podcast is available on:

• Spotify
• YouTube
• Apple Podcasts

Transcript

Eric Seufert: Welcome to the Mobile Dev Memo podcast. I am your host, Eric Seufert, and I am joined again by Peter Stuart. Peter, welcome back.

Peter Stuart: Thanks for having me, Eric.

ES: I had you on six months ago in December. We talked about a topic I find very fascinating, which is the application of AI to the news and media publishing space. You gave us a really good overview of what was happening in that space.

My sense is a lot has changed in those six months. When we first spoke, you had just launched your publishing platform, Velora. You had maybe encountered a strand of negative sentiment about the application of AI to publishing. Please introduce yourself to the audience or those who didn’t hear the first podcast and then walk me through what’s changed in the world of publishing in terms of the application of AI and the perceived acceptability of using AI in the publishing workflow since we last spoke.

PS: Happy to chat about that. Once again, thanks for having me on. I am Peter, the co-founder of Velora, along with my technical co-founder, Danny Bellion. I come from a background of having been a cycling journalist since I was 21 years old. Before that, I was a general journalist and a student journalist with Eric.

I went through the heyday of magazines. My last role was as the editor of Cyclingnews, which is the world’s biggest cycling website. About a year ago, I came up with the loose idea of developing something that was going to be more AI-native as a concept for publishing. In conversation with Danny, whom I met at a barbecue after not having spoken to him for six years, we got chatting and we decided to give it a go. We simultaneously quit our jobs and went for launching this platform.

As you mentioned, it was originally a cycling website and it still is a cycling website, but our focus was always on developing a platform behind it that would ultimately take the heavy lift out of modern journalism and modern publishing. A large part of that has just been the decomposition of the writing process and the publishing process. Danny, having not been in the publishing background, looked at what I did day-to-day and looked at where he thought an agent or an AI tool could do the job better or at least augment the job substantially.

Where we started with the cycling publication, we quickly pivoted to producing this as a public-facing platform that other publishers could use. Since we chatted, we have about a dozen different publishers using the platform on a license basis across totally different sectors. We are just trying to help more publishers be able to use what is a suite of different AI tools in a very tight orchestration to take the heavy lift out of their publishing workflow.

Sometimes that is just basic stuff like tracking news visibility. Sometimes that is more involved in terms of the CMS uploading, finding images, and sometimes it is extending through to AI drafting articles and doing some more automated content. Key to it has always been the visibility element, the research element, the verification element, and the CMS manual upload agentic process, which are a help for anyone in publishing regardless of their view on AI writing or AI drafting.

ES: What has happened in the news and media publishing space with respect to the adoption of AI in the six months since we spoke last?

PS: When we last spoke, it was probably when there had been a lot of stuff published about the quite slow uptake of AI in journalism across the UK certainly. I think the Reuters Institute had a stat that about 20% of journalists had never used an AI tool in any journalistic task at all. Only about a quarter were using it on a daily basis, and a large part of that was just AI transcription tools like Otter.

Since then, there has been some more public embracing of AI tools. You have a high-profile partnership between News UK and Symbolic, which is an AI editorial suite. Outside of the UK, we certainly had other countries that have had more fundamental AI infrastructure become part of their workflow with multiple agent-generated news with almost effectively totally automated local news processes or news production.

In the UK, I think, as with the US, a lot of the development has been slightly hamstrung by a lot of high-profile, embarrassing cases of AI being misused or causing editorial errors. That is where most of the news has hobbled. Most AI use is still quite discrete. Speaking to trainers in the AI space, while I think more companies are looking at ways to use AI, most companies are still quite early in their process of currently still getting to terms with a Claude subscription, Claude Pro work tools, and Gemini basic chatbots. We haven’t seen the full agentic process go into many publications and certainly not in a public way.

ES: For the most part, any major news publication now has an LLM-generated summary at the top. I don’t think readers or consumers object to that. In fact, people probably find it really helpful. I started including that in the posts I send by email because I think it is important for people to get a sense of what this thousand words is about before they commit to it. If you just read a two-sentence summary that is obviously generated by an LLM, it gives you a sense of whether this is worth your time or not, and I think people find that helpful.

Where is that boundary? Where does it go from being a helpful perk that you are providing to me versus having deceived me by making me read slop? Where do you draw the line? A summary is an acceptable use case, but writing the full article with an LLM probably wouldn’t be seen in the same light. Where is the line? If I add whole paragraphs that were LLM-generated, where do you think you would draw a line there on the output side of things?

PS: What you see with lots of publications is they have developed language that will state if an AI is used to create a substantial part of the article, we will make it transparent. That is the language you might see with the Guardian or the BBC. I think the BBC might specifically say that writing is not allowed or research is not allowed with AI tools basically to be even more clear-cut.

There is a furry line when you say AI is used for any part of the writing process versus AI is maybe generating articles quasi-autonomously. To my mind, it gets super cynical and a bit distasteful when something is positioned as commentary or opinion and it is clearly AI-generated. That is when you feel as a reader, have you really articulated this yourself or is it just the result of two prompts or even one prompt like ‘write me a hot take’.

It is really important to consider how that is different to something that might be a financial write-up of a quarterly report. At that point, really all you’re trying to do is you might have a huge document that is emailed to you in press release form or an SEC filing and you’re just trying to delineate the key information for your audience. At that point, I’d say if you have a robust pipeline in place and you have verification in place and the information is correct, it doesn’t really make a difference in my perspective if that is AI-written or human-written from scratch. The core thing is an informational journey for the reader, and if they have found that was fulfilled appropriately, then AI use or no AI use isn’t really a problem.

Generally, the amount of AI-written content on the internet is actually far more than most people appreciate. Even people at the Guardian and prestigious publications get called out with Pangram-type screenshots saying your sports writers have been using 100% AI-generated content for the last six columns they published. They haven’t often been able to rebuff that in a very efficient or robust way because most publications don’t have control over what every single writer is doing all the time. There is always the argument that the AI detector is not 100% reliable, so it’s not like a fair jury.

In terms of the layer that sits below the nationals and the prestigious titles, AI-generated copy is already effectively ubiquitous. You’re thinking about all the trade magazines, the B2B publications, and the individual site owners that exist there. A lot of these people are financially sensible and responsible for their output and their income. They are naturally going to look at what is the most efficient way to get content out to a level which their audience expects. There is also the amount of offshoring that is a combination of offshoring and AI use. It is also super common across lots of sectors. You might have financial, showbiz, or sporting titles that have teams in the Philippines or India that are using ChatGPT to mass-produce content.

A lot of this gets penalized by Google from time to time, but lots of it doesn’t if the domain is established and has lots of authority. The actual broader ecosystem for how AI is used has changed substantially. My expectation now with most publications, when I see content if it has regular output of quite commodity content, is that AI has played a part in the creation of lots of the copy. Whether that is obviously detectable is probably more of a technical question than it is ethical, because people that use AI a lot are probably getting better and better at disguising the obvious AI-isms.

That kind of becomes to me the most insulting bit when someone’s not even tried to cover up the fact that it’s ‘that’s not weakness, it’s fragility’ or something and it’s like, come on, this is freaking obviously ChatGPT talking, get rid of that. You kind of end up more resentful of the person’s technical gap and not thinking why didn’t you cover this up. All that’s to say, I don’t necessarily promote this as a good outcome. The machine-written ecosystem of publishing is kind of scary in a way, but we have to be honest that people aren’t going to make life difficult for themselves. If you have basically the world’s most efficient intern sitting right next to you willing to do all your writing for free, you’re not going to labor through it yourself if you’ve got a mortgage to pay and you’re conscious of hits and time.

In the broadest term, that has changed quite a lot across the whole ecosystem of publishing. Where that sits for big publishers, I’m not sure. Truth be told, while I was saying before that they’ve got AI tools in place and are trying to work on that, what I hear from people that are trainers in the space is that a lot of them, because they’ve had such hostility and such fear amongst the staff, they still are maybe a year or two years behind what other sectors might be at. They are now currently getting into using Claude Co-work or setting up an agent, but they’re not quite where other industries maybe who are saying, even in marketing, a lot of marketing teams have got quite efficient content automation pipelines in place already. It’s kind of maybe a bit of a given. Increasingly even from the SEO teams that used to be very anti-AI, they’ll accept the assumption that AI has played quite a big part in the writing because it’s just naturally efficient.

ES: I think it’s a great point. Where does the audience delineate acceptable use? That is probably what ultimately is going to define what acceptable use is and not just these legacy media organizations trying to maintain the status quo because it’s just not going to be economical. I was writing today about this company called Bending Spoons that filed for an IPO. I was writing up the F-1, which is the foreign equivalent of an S-1. Historically, I would have just read that whole thing and it would be a hundred pages and take two hours. Now I just feed it to ChatGPT and I have a list of questions. Tell me where I can find this information.

Also, if it’s giving me just a two-sentence paragraph that summarizes the growth, there is no discernible difference between what I would have written and what it wrote. It’s just a bunch of numbers concatenated with ‘ands’ and commas. You wouldn’t be able to tell. For that purpose, if I copied and pasted that, I don’t think it is a violation of trust on the part of the reader. That’s not analysis, that’s just reporting metrics.

What have you heard? What is the range of attitudes across big media organizations? Does it tend to map by size? The BBC says absolutely not and then some small blog says we just don’t care. What is the range of attitudes and how does that map to size, prestige, or age of the organization?

PS: Really big organizations haven’t changed particularly quickly, and that’s probably not to do with ethics as much as to do with just the inertia of a very large business and the way it operates. The BBC has thousands and thousands of journalists. The Times has hundreds. Those people have their own workflows and their own desks, and telling them to switch to something else is going to be really hard at the best of times.

Quite commonly, these big organizations might roll out a Gemini license and then uptake is absolutely tiny. I’m actually headed to a conference next week for publishing across Europe and it’ll be interesting to see more direct on-the-ground information there. I think you have a scene where in the Nordic countries or Scandinavia or even Belgium, you have much more embracing of AI tools and AI agentic processes. Mediahuis, which is a big publisher that owns a hundred different publications including nationals, has a whole agentic news workflow that they’re very open about. I don’t know what the bylines look like on that. I don’t know whether it is subsequently attached to someone’s name or it’s transparently machine-written.

Generally, places in the UK that do have start to go down the AI drafting route are probably not super vocal about it. The problem with that is that I think trust falls through the floor when someone knows something is AI-written. I think the assumption is that if someone is using AI, it’s a total laziness play. It’s not to do with actually trying to improve the content or be more efficient. There are lots of studies on this that trust of AI media is super low. Trust of media generally is also actually super low, so people are naturally not going to trust AI in media if they don’t trust media anyway.

There’s a big range from full manual-written policies that will be very carefully adhered to and then people that are experimenting a bit more boldly. A lot of it is going to be on a very basic level a kind of economic cost-per-article thesis. If you have The Times, Sunday Times, The Economist, they don’t publish a huge amount of articles for the staff they have. They’ve got a very expensive subscription base. Their readers probably expect a large investment. A writer might have always spent one or two days writing an article.

By comparison, you might have digital writers at lots of publications that might have to do 10-plus stories a day. At that point, if the story was commodity content anyway, it doesn’t really change anything if it’s machine-written. It’s not surprising when a lot of those titles will flip to a machine-written model. I’d expect that the more prestigious titles, the FT, the Times, the BBC, will continue to have a very rigorous copy-is-human-written policy. Whether that actually changes anything or actually makes the experience better for the reader, I kind of question.

Some of these people, like the FT, have actually offered totally AI-generated content as a service, like Newscast that was done a few years ago using a third-party AI news wire service called NowaWire. I don’t think that’s ever made it to consumer readers, but it was envisaged as a B2B service. Again, a lot of that stuff probably was experimented with quite early when the models were still not really quite up to it. I think there’s definitely been a sea change when you have GPT-4.5, probably the moment I’d say that stuff jumped forward quite dramatically in terms of capability of writing. The actual models now are super capable, and now it’s not a question of quality, it’s really just about ethics and accountability on the writing front.

ES: How do you view that accountability piece? You could view the brand or just the association of the publication with rigorous journalism, exhaustive fact-checking, unbiasedness, as a license to embrace this because you give them the benefit of the doubt that they’ll embrace it responsibly. It’s actually interesting that you said the Northern European countries are embracing the use of AI in news because they tend to be a little bit more restrictive with anything that has to do with labor and workers’ rights. If you have this reputation, if you’re the Walter Cronkite halo effect, then you’d think that readers might give them more leeway to introduce this because I trust that they’re using this, but they’re also validating it, verifying it. They’re using it in ways that are aligned with their brand image, and so I’m going to accept it from them, but from another publication, I might not just because they don’t have that same level of trust. Is that roughly a good way to look at it?

PS: To be honest, I think my view on this stuff is that I think the level of trust is very high with those brands, also the level of skepticism. If there’s a commonly shared LinkedIn post about some headline in a newspaper being all wrong, it’s always like ‘look AI messed up again’. Probably not, actually. If someone miswrote a headline, I don’t think AI would do that. It’s quite good at doing that sort of stuff. It’s probably just some sub-editor that wasn’t paying attention. You get the odd end-of-a-newspaper entry that had ‘for the next question, let me know if I can help out by putting this in an email’. That stuff happens quite a lot.

Then you have more high-profile mess-ups, like the New York Times when they had an AI-generated summary that was attributed as a quote, or you had a book review plagiarized from the Guardian, and you had Ars Technica had that huge issue with AI-generated quotes as well. There was a whole list of issues. When titles don’t have a very clear or sensible AI policy, some individuals in that business use AI tools in a very untechnical, unwise way. They have a big scandal and then they totally freeze up and say we don’t use AI at all, it’s all humans. It’s like that recoiling back just because that fear of trust is so strong.

My view of this stuff has always been that if you have super robust AI policies that involve actual AI use and really clear traceability, that is what I don’t think there is a lot of process in place for. Here is our research agent, here is where that’s checked by a human, here is where a draft might be checked by a human, here is where an edit is checked by a human, here is where an edit is completed by a human, here is an n-gram comparison of the edit versus the draft, and you can totally see how much has been changed by the person. That would be a state where you’d think, okay, this seems like it’s a robust, auditable process.

The EU has the Transparency Act for AI-generated content. For that, they actually offer an exemption of news publishers that have very clear auditing in place. If they say we clearly see who set this off to be written, which research tool was used, you don’t have to disclose and say this is AI-generated content. How that law is going to work in effect, I have no idea because how are you going to enforce that across the whole of news media.

I think there’s probably a bit of a technical gap where titles haven’t begun aggressively working on an AI infrastructure that makes sense for them. They are faced with so much audience hostility, people just hunting down some sort of a scandal to slit them up and be like, look, the media got it wrong and they’re using AI for lazy purposes. Alongside that, you have some super ill-advised projects where one publisher in the US had said they have the right to publish articles that are AI-generated with writers’ bylines with or without their consent. That’s just a crazy management decision where it is totally obviously going to have a big backlash.

Trust of AI being used well will be super low for brands where there’s already a massive public skepticism anyway, which typically is the bigger brands that are more public.

ES: I think about the use of Grok on Twitter, for instance. It feels like a very effective way of just combating misinformation because you use the LLM, which also has tool access and can do searches in real-time to validate what people are saying. There’s a lot more AI slop on Twitter now, but you can also use the integrated LLM, which is Grok, to validate what people are saying. If they’re just spouting nonsense, you can call them out in an utterly neutralized way because you can see the Grok response and then someone attaches a Community Note. I think that actually is a really ingenious technique for slowing the spread of false information. Right there, it’s visible on Twitter of the dual use or the dual impact.

PS: I actually think about that all the time. A really interesting specific example of that in action in real-time was about two weeks ago. There was an article published across the international media which was about these Thai police officers that had apparently gone in drag to bust a drug ring. It was based on this picture of them all in drag at the police station having busted and seized loads of drugs. It was probably three or four days it was running until eventually it came out the image was AI-generated and everyone had to issue corrections.

The thing that stunned me about that story is that I looked at it and it wasn’t actually picked up in the UK press at all that this was the case. I took the Telegraph’s picture that was on their actual website, ran it through SynthID, Google Gemini’s inbuilt AI detection tool. SynthID flagged that it was an AI-generated image based on the SynthID pattern. It’s totally bewildering that nobody throughout this whole news cycle thought to just put it into Gemini to be like, is this AI? That tool is free, anyone can do that.

Lots of AI-generated content won’t have the SynthID because it might be done from other AI image generation tools, but the most common one is Gemini or ChatGPT, and ChatGPT also now has the SynthID in it. You could easily use AI to actually combat misinformation, but there’s not necessarily the processes in place to do that. That’s one element I think is really interesting, that people think AI is misinformation, but actually, AI also is a solution to that even if it’s AI-disseminated misinformation.

The point about AI tools being used to combat misinformation is super interesting because while we on one hand talk about the lack of trust in AI-generated content, I constantly see Grok is true on Twitter. I think people go to ChatGPT to check everything. I check medical stuff on ChatGPT. I feel confident to be like, well, I’m an intelligent human on top of this. If it says go to hospital, I probably need to use that information in reference to what I know about the world to be like, okay, that’s fair advice.

On Google, typically medical stuff is kind of terrifying because you’re like, here is an info page about having some awful terminal disease that might have the symptom. AI is super effective actually providing information and being grounded and also weirdly trusted by seemingly all sides. I don’t think someone who is a Republican would be like, ‘@Guardian, is this story true?’ They would innately distrust them. Everyone seems to just align themselves to believing what an AI tool will say.

My best guess is that there’s no implicit agenda within an AI tool. It’s just existing to react to the questions you provide it. You don’t feel like it’s trying to imprint on you, and it probably might be. I’m sure there’s lots of studies into the right or left-wing traces within answers and stuff, but ultimately people just seem to just be happy that it’s responding to them in the way they address it, which probably will mold to whatever political ideology you have. Typically, I think it does tend to mediate against any extreme. If you’re a very left-wing person, ChatGPT may push you more center, and the same with right-wing. Generally, I think it engages you on your level, which I think is why people seem to have this trust that they can just engage in dialogue in a way that doesn’t work with the wider news media where I think you constantly really have to align yourself to one political way of thinking to agree with one publication’s output.

Ultimately, the news is ambient and huge. It’s all happening all the time. Now more than ever, newspapers report on what Donald Trump said on Truth Social and that will be the main component of a story. They don’t need to spend as much time being at a press conference to get a quote. Stuff is just happening all around us. Newspapers increasingly have to choose their segment of reality along a certain polemic of the way they want to communicate that.

In the world of AI, increasingly platforms are already distributing that information naturally in a way that is suited to its audience. TikTok or Instagram may serve reels that provide news to people in a way that’s very different to historic news production. AI tools offer a layer on top of that where they can totally interpret everything that’s going on and produce a really useful answer and an explainer to somebody that’s quite reassuring in real-time. Those things just make an AI tool quite trustworthy for the general population. I think that’s going to be more and more the case. If people are using an AI tool to validate a journalist’s perspective, the game is totally inverted because the person that’s trusted to be the source of information is now being corrected by the AI tool, which is kind of madness.

ES: My point was just that this allows us to have a better sense of whether we can trust the underlying factual scaffolding of things that are reported. You’ve got to trust that the tool itself is unbiased. If you’re seeing news, opinion pieces aside, I’m assuming now that if I can assume they used an LLM-based fact-checker to validate what was written, it’s probably trustworthy, or at least it hews to the preponderance of information that exists on that topic. That should make me give more trust in the news that I read if I’m assuming that these tools are applied to the output, not generating the output, but just applied to it as some sort of filter. There’s just more accessibility of fact-checking, and then maybe I have a better sense of trust for everything that I read going from the biggest publication down to someone’s Substack.

PS: I was thinking about that today. There’s a sense that okay, you’ve got AI not trustworthy, human trustworthy, but obviously, people are not trustworthy. As an editor, the amount of copy that came across my desk was huge. I didn’t have time to check, okay, let me go and Google that this city actually is in Italy or whatever. I’d just be like, cool, well, that looks fine, the journalist presumably knew what they were talking about, no spelling mistakes, let’s go. Don’t quote me on that, obviously check the stuff that’s high stakes, but you don’t have a constant rolling fact-checker.

As technology changes in most sectors, you have an expectation that if a tool is available and it’s capable, there begins to become a dynamic where not using it becomes the negligent thing. If a doctor refused to use a CT scanner or didn’t do X-rays, no, I just don’t believe in that, I’m an old-school orthopedic surgeon, it’d be like, no, use the freaking technology, otherwise, if something goes wrong, then that’s suddenly a culpability issue for you. I feel like it’s very feasible that could become the case with journalistic publishing work because you’ve got a grounded search on Google Gemini Pro or something. It is incredibly capable at returning information and not just information that is based on training data. It will find search-based validated information in real-time and be like, okay, this 100% is true based on this source or this actual primary evidence based on what someone said on X. There are gaps there, and you sometimes have hallucinations cascade through news media. Even then, there are points where you could develop an anti-hallucination validator to go find the original source. You can actually introduce tiers of validation and verification that would be totally impossible for any news team unless they had a crazy budget of hundreds and hundreds of heads.

In that sense, I feel that AI can and will ultimately make stuff a higher bar of verification and truth when it’s used in an intelligent and well-deployed way.

ES: To your point, you might even want to disclose the fact that you didn’t use it. You might want to disclose that no LLM was used to fact-check this article. Then I would probably be more skeptical of what was written there, that it wasn’t at least validated against common terminology or common reference metrics if it was using some source that was seen to be biased in some way and that wasn’t flagged. I want to know that because I just assume now that anything I’m reading on the internet has at least gone through a pass. I do it. Give me spellcheck and grammar check at a minimum. The widespread availability of these tools should fortify a lot of the stuff that we read. I’ve got one of my prompts that I take an article through before I publish it, especially if it’s a more controversial topic: read this adversarially. Tell me the top five criticisms that someone would have against this if they were very hostile to the ideas that I’m articulating here. That’s really helpful because it helps to soften some of the edges or deflate some of the more tendentious things and make it less likely that someone’s going to explode when they read it. You just see that effect. If you focus on just these tools as a layer that the human-written output passes through to improve them, I don’t think anybody could argue that that is a benefit.

PS: On the Velora platform, we have a whole pipeline from article idea to verified final CMS-ready content. We have actually developed a whole separate pathway which is manual writing. In that pathway, you just write into our platform, but as you go, you can select text and you can say ‘challenge this’ or ‘find a new angle’ or ‘fact-check it’. It does that against research documentation that exists against the topic. That has been super interesting because I don’t think at any point you can say you have an objection to that in terms of AI interfering with the process because you can take or leave the advice.

Typically it’s super effective at being like, you only have your own perspective in writing, it is effectively someone else, like a live real-time editor saying, you know what, that doesn’t really make sense, does it? Or don’t you think that people are going to question whether that’s a valid argument in light of this? I find that’s super helpful as I’m writing stuff like analysis-type stuff. Highlight this, challenge this, angle this. I feel like that could become the standard way to ultimately see writing reach a new level of quality.

We think about AI tools helping coding or coding agents that really accelerate in a way that most people no longer have an objection to. No one’s saying ChatGPT creates slop, it’s how most of the internet now is being built. Writing could be the same. As the technology develops, I think we’ll see tools that actually augment writing in a really positive way and actually lead to writing being better because you’re giving people that might not have had the experience of having an editor sudden having an LLM tool to say this could be a bit better. Most seasoned writers have had that in human form. They’ve had someone go through and say you need to tighten up your writing, we need to knock the edges off this. That’s a really key part of maturing as a writer. I don’t know many people that wouldn’t have had that at some point in their careers that have suddenly gone on to be very accomplished or technically strong writers.

That’s a real positive. The flip side is that when one editor is very edit-heavy at a title, everyone’s content tends to conform to that person’s writing. The same thing could happen with ChatGPT. If the whole world is using ChatGPT as their de facto editor, then everyone’s writing will naturally trend towards what ChatGPT inherently thinks is good. But within that, there are flaws there, there are risks, but I think there’s still a lot of scope for originality and for just treating an LLM as a really healthy adversarial or supporting writing assistant or co-writer.

ES: Yeah, and I’m not talking about ‘give me advice on the prose’. I’m talking about ‘fact-check this’ and ‘give me a sense of what the holes are in it’.

I want to talk about the Velora platform because I think what you’ve built is, if you just had a one-liner elevator pitch for the platform, a layman would probably say you’re going to contribute to more AI output on the internet. That is absolutely not what the Velora platform is. What I’m impressed by with Velora is the capabilities with respect to finding, processing, classifying, sorting, ranking news and ideas as a tool. That to me is what an LLM is truly good at. An LLM does next-token prediction and it generates text, but I think a lot of the misunderstood value is just doing all those things: classifying, ranking, sorting, summarizing. Talk to me about how the Velora platform is used as a publishing tool because that’s the piece that I think people don’t appreciate about the power of AI in publishing. We talked about one piece just now, which is fact-checking, but talk to me about the whole suite, the whole end-to-end use case of this stuff that has nothing to do with generating text.

PS: Ultimately, whenever you think about AI tools, you can take objections to them and say I don’t like the idea of this, I don’t think it’s trustworthy for writing copy, which is all totally fine. But there are some things you can’t debate that an AI tool is better at. One of those things is certainly taking a huge amount of information and quickly processing it and understanding it in a very top-line way.

News discovery in terms of signals and understanding of what is interesting to your audience, you can ingest a thousand sources every hour and an AI, typically a person might scroll their competitors and their beat, might look at social posts, but they might have 30 signals or sources. An AI can look at thousands and it will sort in an intelligent way. This is what is relevant to you at this particular moment, these are the most high-interest topics for you to pursue as a starting point. In my mind, there’s just no competition there. There’s a value in someone looking through the whole news of the picture, but you can’t possibly do it as well as an AI tool.

Another element you can’t do as well, or perhaps you need to do in symphony, is research. I think definitely there’s value in doing your own research, but as you said, you can’t look through dozens and dozens of sources, hundreds of pages of PDFs in five minutes. It’s not something that a human being is able to do. If you had enough time and resource, maybe, but ultimately, there’s a limitation that you simply cannot scope this much information, you can’t draw so much research together, you can’t understand the picture widely like that. So that is something that I think an AI tool again is fundamentally better at in symphony with a person.

Then at the other end, the pure manual processes. Finding an image that’s relevant, that’s like a long personal drain on time. Populating the fields in a CMS, that is a very laborious, menial task that people have historically done. CMSs have been built because people have to do them, have to populate these fields. They have to say here’s a title field, here’s a subhead field. When you really think about it, that doesn’t really need to have any human interface. If an agent or an AI tool can understand that that’s where that copy needs to go, we only created a complex infrastructure to enable someone to put it there because there’s no other way of taking copy from here and putting it there. When an AI can do it, you can just abstract that step entirely away from the person looking at it or doing it. Similarly SEO or alt text, it’s all just menial work that people end up having to do.

That final level, just creating a validation layer that’s really robust on top of human validation, and then actually creating distribution, so say publishing a piece but then creating social assets that accompany it, creating suggestions for social copy, stuff like that. This is all stuff that takes a long time, but very few of those steps really add anything at all to the creative process that people actually value in journalism or a piece of writing. It’s just a lot of unnecessary surrounding labor and noise.

Within that, in the central part of the platform, we do have drafting, but that’s totally down to an editorial team’s discretion whether they decide to do something manually by using an AI-augmented writing tool like where it will challenge or make suggested angles, or they just ask for a first-run draft. Within that first-run draft, we also have a ‘grill me’ phase, so it will be the AI tool will come back and be like, what are you thinking for this story? What’s your actual angle? What are you trying to focus on? There’s this bit that was in the interview, there’s this bit that’s in the PDF document. What’s actually important to you? It will go through seven or eight steps of that and then it will then go away to form a more cohesive brief that you can look through and then it will go through to a draft.

At that point, some of that content that I’ve gone through has been kind of heartbreaking because I’ve done that for a feature. I thought, okay, I’ll try this grill me process and I’ll use all this information. I looked at the final version and I thought, wow, I don’t really know what I can do here because it’s just exactly what I’d expect but in a way that was super neatly stitched together. Then I can change a bit of wording here or there, but I’m not really adding anything. It’s just customizing. But when you’ve added that level of information, the draft you’ll get back is super high quality. That’s probably what’s lost in the one-shot slop AI writing world where you’ve got someone just going ‘boom, here’s a LinkedIn post’ that says nothing and means nothing versus here’s a lot of information, here’s a lot of context, and then an iterative writing process.

We kind of do all that and we do that for multiple types of content, like news or features. The idea is that you give time back to the journalist to do something interesting. That’s the key core principle. What is interesting is not whether you spent 25 minutes writing a quarterly financial report or two hours in the morning looking through a thousand sources to find out what was the most interesting bit of information in the morning in the Canadian fishing beat. That stuff doesn’t really add. What’s interesting is going out into the world, speaking to people, actually looking at a whole topic and saying I think the most interesting angle here is this from my lived experience and actually adding something human to it. We often think the human bits are in the writing, but if it’s totally formulaic programmatic writing anyway, someone isn’t really adding their perspective, it’s just a drain on time. We have to be realistic about is that something someone should be doing. Ultimately that stuff people will be doing less in the future, whether it’s because of AI or whether because that content just becomes less valuable because it didn’t really take any effort to create, it’s not scarce in any way. That’s the interesting question that’s going to surface in the next few years. Journalism will have to get to a point where people do more interesting things with their time. They actually find interesting stories, they have interesting conversations, they form interesting perspectives. Without that, there’s no real moat against someone just churning out absolute tons of content at scale using an AI tool without any sort of high benchmark of verification, interest, or journalistic ethics in it.

ES: I think that’s something that people that have never written as a professional activity just miss. How much reading is involved in publishing something? There’s a lot of times when just now I’m writing about Apple’s new Foundation Models framework and I had to read two technical papers this morning. The blog post will be 800 words, but it took me 45 minutes to read. To not read a technical paper takes a very long time to read, but just to scan it and get the big-picture idea is very time-consuming. With a technical paper, I don’t really trust just dropping it into ChatGPT, but there’s a lot of stuff like an earnings report. I could spend 45 minutes reading this whole thing word for word. I could drop it in ChatGPT and say I’m looking for these pieces of information. Tell me where to find them. Give me three anchor words that I could search for to go exactly to that point and then skip all the questions that I don’t care about. It’s just so much more efficient. Maybe I get to write two earnings analyses in a night instead of just one because I didn’t have to read both earnings call transcripts word for word. That’s not AI-generated output. It has nothing to do with actually generating the words. It’s more about how you inform me and give me the latest information as fast as possible so that then I can ingest that and then write my analysis. Find me new information that could add more context to my understanding of this topic. That is really underappreciated. How do I find it? It is unknown unknowns. Where do I even go to look? If there’s a tool that knows it’s got a huge bank of resources and it can scan basically every 10 minutes to find stuff that’s relevant to me, that’s just helpful. It doesn’t generate the output; it makes the writing process more efficient, and that really is the true advantage of using these tools.

PS: That’s it. To visualize that a bit more clearly, for us that’s like we have in the platform we have an RSS and a web scraping section. We have a social media section. We have a YouTube section. That stuff’s all pulled super frequently. Some stuff might originate from international press or a different sector, might originate from a competitor, but a lot of it’s just from social media that’s happening. YouTube transcripts, finding a very specific part of a long podcast that’s actually newsworthy or an interesting starting point. That stuff that is just genuinely impossible to do on a personal level. What’s cool about the platform is that the journalist decides what the sources are. It’s not just saying ‘hey ChatGPT find me an interesting story’. You’re in charge of saying, I think these international outlets are really important and I’m going to over-index on them. I think these social accounts are the ones I really want to follow. These are the YouTube channels I want you to look at every single day and tell me what’s happening up there. That’s where you get a bit more originality and curation put back into the process compared to just a general sweep of saying ‘Claude tell me what’s interesting’.

ES: All right. I know we’re at time. Take this moment to promote the Velora platform. How can people test it out? How can they add it to their writing workflow?

PS: We are at velora.build. I’m Peter Stuart on LinkedIn. You can just come in and add me on LinkedIn, ask to use the platform. We let people use it. We’re still in the early phase so we want people to use it and feedback and tell us what they like or hate or think’s really good. Just reach out and I’m happy to do demos. We’re also happy to work with people on their specific use case because we’re just super fascinated in how this scales out to someone working in fishing or birdwatching or super specific insurance financial stuff. So far we’ve found that it’s actually able to help people in almost all cases because some of these tools just are totally title or sector agnostic, which is really fun. Reach out to me and please don’t hesitate. I’m friendly and I’m always happy to have a chat and even if you don’t want to use the platform, just happy to have a chat about AI. My favorite topic is AI and journalism. Don’t be a stranger.

ES: Thanks, Peter.