Confluence for 5.17.26
It's not just a trend, it's a real problem. Slop is in the eye of the beholder. AI use and self confidence. The model as data hub.
Welcome to Confluence. Here’s what has our attention this week at the intersection of generative AI, leadership, and corporate communication:
It’s Not Just a Trend, It’s a Real Problem
Slop is in the Eye of the Beholder
AI Use and Self Confidence
The Model as Data Hub
It’s Not Just a Trend, It’s a Real Problem
As AI literacy rises, unedited AI output is a growing credibility issue.
We’re hearing a version of the same story from multiple clients lately. Someone on the team submits a first draft for review, and the reviewer can tell immediately that the draft came straight out of an LLM, with minimal human shaping. Although the writing might be good and the formatting might be neat, the whole thing reads like it was written by someone who has never met the audience, the project, or the organization. Something’s just off. The reviewer then has two jobs: do their own review, and redo the drafting work that should have been done before the document reached them. It’s a problem we don’t expect to disappear any time soon.
Last week we wrote about Claude’s visual tells: the thin colored bars, the outsized numbered lists, the declarative slide titles that signal AI-generated documents. The linguistic tells are the older, more catalogued sibling. Wikipedia’s editors, who have been cleaning up AI-generated submissions at scale, maintain an extensive guide called “Signs of AI Writing” that catalogues the patterns: overuse of em dashes, formulaic parallel constructions, transition-heavy prose, vocabulary tics like “delve” and “tapestry” and “underscore,” and a tendency to inflate the significance of every topic. Barron’s recently reported that a single one of these constructions, the “it’s not just a ___, it’s a ___” pattern, quadrupled in corporate filings and press releases between 2023 and 2025, according to data from AlphaSense. Colleagues across our own firm have started keeping running lists of AI writing tells they see across engagements: “move” (as in “the most-cited move is”), “genuinely,” “and honestly?” to name a few. The list is ever-changing—recently, structural imagery like “load-bearing” or “X is doing heavy lifting here” has been the darling of LLMs everywhere—and once you see them, you can’t unsee them.
And that’s the rub. The people most likely to spot these tells are the ones who use AI most themselves (and the research proves it). The tells become obvious once you’ve seen the patterns in your own output and learned to edit past them. As AI fluency spreads through an organization, the tolerance for unedited AI output drops, too. Submitting a raw chatbot draft to a colleague who uses AI daily is the professional equivalent of forwarding an email you clearly didn’t read. The transparency dilemma and competence penalty we’ve covered before both apply here: disclosure carries a cost, but getting caught carries a bigger one.
We’ve long been advocates of transparency and discussion in how teams are using generative AI in their work, whether it’s comparing approaches and outputs, building shared editing standards, or simply talking openly about what’s working. As organizations contend with the question of how to manage AI-produced output, those that are most successful are the ones who identify, examine, and discuss what they’re seeing and how they’ll maintain their expertise moving forward. The expert in the loop is what separates AI-assisted work from AI-generated work, and the distinction matters more every day.
Slop is in the Eye of the Beholder
AI tells are real, but spotting AI is still harder than we think.
We wrote last week, and above, about the AI “tells,” new and old, that we’re seeing more of every day. Those tells are real, and as we argue above, they’re becoming easier to spot as more people use AI and learn to recognize each model’s default style. All true, and important to know about. But the broader conversation around these giveaways can leave some with the impression that AI-generated or AI-assisted work will always be easy to recognize—that it will never be good enough to fool us, or that we’ll always know the real thing when we see it.
Not so fast.
The counter-argument for this played out on X this week, when a user posted an AI-generated image in the style of a Monet and asked others to describe what made it inferior to the real thing. Responses came flooding in, most describing in great detail why the image was obviously AI-generated: it was emotionless, soulless, lacking in composition, the brush strokes were wrong, things looked upside down, it had no appreciation for depth or contrast, and so on (you can see some of the most entertaining responses here). But the twist, it turned out, was that the image was in fact a real Monet, one of his Water Lilies (Nymphéas) series. The image was presented as AI, and so the commenters saw it as AI, and they identified false AI-indicators with such confidence that the work of one of the most famous painters in history was repeatedly labeled as clearly, obviously, undeniably bad and inhuman, as AI slop at its finest.
So there are real caveats to those “obvious” tells we’ve been writing about. Yes, the more we use AI, the better we tend to get at spotting it elsewhere. But the most proficient users are also usually much better at editing known tells out, or prompting the model to avoid them in the first place. Most of us at the firm, for example, keep a section in our custom preferences that explicitly directs Claude (the not-Monet one) not to use the most common linguistic AI-isms. It’s an imperfect screen, but we generally find that skillful prompting, custom instructions, and well-built Claude Skills or the like can produce work that aligns far more with an individual or organization’s voice than with the model’s default. That alignment is a big part of what makes AI such a powerful tool in the first place.
The jury is also still out on things like AI detection software. The evidence so far shows it is unreliable and tends to be biased against non-native English speakers, something academia in particular is struggling with right now. Research also shows, as we’ve written before, that AI-associated words and phrases are making their way into how humans write and speak, even without AI. Humans and AI are starting to sound alike.
All of which is to say: none of us, no matter how proficient we are or how good our tools, should be too confident in our ability to tell AI-generated work from the real thing. If anything, that mindset is itself becoming a sign of a certain kind of AI naïveté: if you think AI can never be good enough to fool you, you’re probably not using it very much. The deeper lesson, though, comes back to credibility. Once people think they see AI, or the tells they associate with lazy or unedited AI work, they tend to perceive it as slop. For communicators and leaders, that’s what matters. These little tells are worth watching for, even if they don’t always indicate AI use, because of what they may signal to an audience about the quality, rigor, and authenticity of the work and its author. If Monet isn’t immune, neither are any of us.
AI Use and Self Confidence
How we work with AI says something about our confidence in our own skills.
We’ve been watching how generative AI use affects our own abilities for a while. As we offload work previously done by humans to AI models, we’ve seen and expected effects on our skills, sense of meaning and purpose, and other consequences yet unknown. Confidence is a dimension we hadn’t thought as much about, but it belongs on the radar.
The American Psychological Association recently published research on it. A total of 1,923 adults completed 10 tasks designed to approximate the cognitive demands of most knowledge work. They used commercially available models between June 2024 and December 2024, and were expected to have reasonable experience with generative AI. When participants said AI did “most of the thinking” on a task, they reported lower confidence in their own ability to do that work. That finding by itself makes sense. But the next one is more useful.
AI use itself wasn’t associated with reduced confidence. Confidence varied with how participants engaged with what the model produced. The people who actively revised or rejected AI suggestions reported greater confidence in their reasoning. The people who accepted outputs with minimal modification reported less. Authorship and autonomy, the researchers suggest, are shaped by patterns of oversight more than by the presence of AI itself.
Using generative AI is increasingly frictionless. The interface is designed to give you an answer. The whole experience optimizes for getting that answer in front of you fast. If we want to keep our thinking sharp, we have to put the friction back in ourselves.
Write your own draft before you ask for one, then compare. Ask the model to steelman the opposite of what it just told you. Tell it to argue with your conclusion. Ask what you’re missing. Reject the first output and demand a second take from a different angle. The habit underneath all of these is refusing to accept what comes back as the final word.
This is harder than it sounds. The path of least resistance, accepting and copying and moving on, is genuinely faster in the short run. But there’s a real trade-off involved. Most AI models do just want to give you the answer, and introducing this cognitive friction takes deliberate effort on our parts.
The Model as Data Hub
A powerful intelligence that can work (and could make mistakes with) your data.
This week OpenAI announced a preview for its Pro users of a new personal finance function. The function uses Plaid, which is a leading financial data integration platform, to connect ChatGPT to your personal financial institutions. In the same way that tools like Plaid allow the easy connection of financial information to services like TurboTax in the United States for annual tax prep, the OpenAI/Plaid integration allows ChatGPT to have access to credit card accounts, bank accounts, investment accounts, and more. The promise is that it allows you to use ChatGPT’s intelligence to analyze, understand, and do more with your personal finance data.
The ability to aggregate personal finance data online is nothing new. What’s new is a layer of generative intelligence between you and the data that can not only present it to you in common-sense language (and tables, and graphs, and images) but can also answer your questions and work with the data in an interactive way.
Because we have an OpenAI Pro account and because we like to try new things, we spent time this week giving it a shot. Our first ask was for a spending breakdown across bank accounts for the past month. ChatGPT provided that information in a few seconds, after which we asked for a payee-level analysis across budget categories. ChatGPT quickly provided that breakdown as well, and we found it comprehensive.
We know models can hallucinate, so we then asked ChatGPT to verify its math in a spreadsheet table. It did so, and its math was accurate. Then we asked for an analysis of our investment portfolio, but asked ChatGPT to do so as Nassim Taleb, author of “The Black Swan,” as opposed to a typical bulge-bracket bank financial advisor. And it provided a very thoughtful “Talebesque” critique.
After some back-and-forth like this, we have to say we had a “Wow, this is impressive” moment.
We expect Google and Anthropic to have equivalencies soon. But ChatGPT’s personal finance service signals something on the frontier that we think will become much more central to how these models work: a large language model serving as an interactive, intelligent hub for personal and professional data.
We’ve written many times about Model Context Protocol, or MCP, the open-source approach created by Anthropic that allows large language models to access databases of all sorts. While people have been able to connect Claude, ChatGPT, or Gemini to data sources like their Spotify account or their Oura Ring for some time now, we think two recent developments are changing how this feature will present itself for everyday users.
First, the number of connectors is exploding. There are connectors for Uber, for ZoomInfo, for Spotify, for Apple Health, for many other services where people or organizations have accounts and data, and one can now use a model to interact with that account and data, both to push and pull information. For example, you can not only use Claude to find restaurants with Resy, but you can have Claude make reservations with Resy as well via its MCP.
The other, and it’s what we find most interesting and impressive about the ChatGPT finance service, is how it bundles multiple connectors as a suite of services that allows a more powerful approach to working with and understanding personal finance data. We presume there are a set of custom prompts and programming scripts working under the hood to ensure the math is correct and to make ChatGPT’s financial commentary helpful and accurate. It’s not too difficult to see how similar bundles could apply to other domains of life, healthcare being the first and most obvious example.
The past several years we’ve treated large language models as a source of content generation, and, increasingly, learning. Professionally, we use them to produce real work, and now agents, mostly in software development but increasingly in other areas, are starting to work autonomously to run processes and create output.
There’s a new capability coming, though, which is the model as the center of your information: to aggregate and present it to you, and to serve as a layer of intelligence that helps you understand and make decisions about your data. Of course, hallucinations are a real risk in this world, because you can’t make good decisions if the intelligence layer is making up bad facts about your data. But our work with OpenAI’s finance suite suggests they’ve largely solved that problem, at least in this service.
If we’re trying to figure the shape of things to come, this is one place where we can start to see the contours: the ability to talk to the information in your life, and have an assistive intelligence that understands it, applies domain expertise to it, and acts as an advisor as you make decisions about it, not just within a domain, but across them. Such an ability is something nobody saw coming three years ago, and it’s something we think will have significant applications in both your personal and professional worlds in the years to come.
We’ll leave you with something cool: Mayo Clinic researchers have built REDMOD, a model that identifies indications of pancreatic cancer in normal-seeming scans, years before symptoms.
AI Disclosure: We used generative AI in creating imagery for this post. We also used it selectively as a creator and summarizer of content and as an editor and proofreader.