Confluence for 9.28.25
"Workslop" and leadership. A lesson from radiology. Claude comes to Microsoft 365 Copilot. OpenAI launches video academy with AARP.
Welcome to Confluence. Here’s what has our attention this week at the intersection of generative AI, leadership, and corporate communication:
“Workslop” and Leadership
A Lesson From Radiology
Claude Comes to Microsoft 365 Copilot
OpenAI Launches Video Academy with AARP
“Workslop” and Leadership
Preventing “workslop” is a leadership challenge, not a technical one.
We’ll begin today’s edition of Confluence with a simple question: have you ever been on the receiving end of AI-generated work that looked polished but actually made your job harder? We suspect most of our readers have. If so, you’ve experienced “workslop,” a term that researchers from BetterUp Labs and the Stanford Social Media Lab introduced in a recent Harvard Business Review article provocatively titled “AI-Generated ‘Workslop’ Is Destroying Productivity” (and quickly picked up in an Axios article with the same bold framing: “AI ‘workslop is crushing productivity, study finds”). The concept is straightforward and likely familiar to our readers:
As AI tools become more accessible, workers are increasingly able to quickly produce polished output: well-formatted slides, long, structured reports, seemingly articulate summaries of academic papers by non-experts, and usable code. But while some employees are using this ability to polish good work, others use it to create content that is actually unhelpful, incomplete, or missing crucial context about the project at hand. The insidious effect of workslop is that it shifts the burden of the work downstream, requiring the receiver to interpret, correct, or redo the work. In other words, it transfers the effort from creator to receiver.
The phenomenon may have captured headlines this week, but it’s something we’ve long anticipated. In the “What We Expect” section of a white paper we published in the spring of 2023, we noted that as use of generative AI becomes more common, quality assurance will matter more, writing that “AI has the power to turn a much larger percentage of an organization into credible creators of text and more — and leaders will need to be directly involved in ensuring that this empowerment aids and does not diminish clarity and alignment.” While that prediction focused primarily on the creation of text, it’s now clear that the point holds for nearly every facet of knowledge work — from coding to analysis and much, much more.
“Workslop” is what happens when we give people incredibly powerful tools for producing nearly any kind of output, encourage their use, and fail to establish and maintain appropriate standards and expectations for that use. So what can leaders do? The HBR article concludes with sound advice:
Leaders will do best to model thoughtful AI use that has purpose and intention. Set clear guardrails for your teams around norms and acceptable use. Frame AI as a collaborative tool, not a shortcut … And uphold the same standards of excellence for work done by bionic human-AI duos as by humans alone.
Nabeel Qureshi wrote earlier this year that “The opposite of slop is care.” When it comes to “workslop” specifically, that care manifests as leadership commitment to establishing clear expectations about AI use, maintaining quality standards, and engaging in ongoing dialogue with teams about how AI can elevate rather than diminish their work. “Workslop” is a preventable problem, not an inevitable one. And like most things in organizations, it comes down to leadership.
A Lesson From Radiology
The Jevons paradox in action.
This week, we came across a piece by Deena Mousa in Works in Progress that caught our attention. “AI Isn’t Replacing Radiologists” examines the current state of radiology in the United States. This matters to us because radiology has long been seen as especially vulnerable to AI replacement. In 2016, Geoffrey Hinton, one of the most important and influential AI researchers, went so far as to say “people should stop training radiologists now.”
Nearly a decade has passed since those remarks, and we should be deeply thankful we’ve continued training radiologists. Mousa notes that demand for human radiologists is higher than ever. Diagnostic radiology residency programs are offering more positions than ever before, and vacant positions have reached an all-time high. Average radiologist salaries have increased 48% since 2015. All this has occurred despite the existence of over 700 FDA-cleared radiology AI models, which account for about 78% of all FDA-cleared medical AI devices.
We share this not because we have particular expertise in radiology, but because it represents a field where AI experts confidently forecasted the replacement of human labor, yet those predictions haven’t materialized. It reminds us to maintain healthy skepticism about such projections and to think carefully about the relationship between human expertise and AI capabilities. While radiology has specific factors driving these outcomes, Mousa raises a broader concept worth our attention:
As tasks get faster or cheaper to perform, we may also do more of them. In some cases, especially if lower costs or faster turnaround times open the door to new uses, the increase in demand can outweigh the increase in efficiency, a phenomenon known as Jevons paradox.
For leaders navigating AI adoption, this means efficiency doesn’t necessarily lead to fewer people doing the same amount of work. The increase in efficiency means we can simply do more of the work. When something becomes easier and faster to produce, we discover more instances where it makes sense to do so. This pattern extends well beyond radiology into other domains of knowledge work.
We’re seeing this play out in our own firm. Historically, we would ask junior colleagues to dive deep into research on topics related to our client work. This meant searching for the most relevant research, reading dense academic papers, and synthesizing findings for our colleagues and clients. A comprehensive research summary took days, if not weeks, to do well. Now, with generative AI’s deep research capabilities across all the leading models, we can produce quality research reports in a fraction of the time (though these capabilities still have limitations we need to keep in mind). Since the time and effort cost has decreased so dramatically, there are more instances where it makes sense to create these research reports using generative AI.
The question we’d invite our readers to consider: what tasks have or will become dramatically more efficient thanks to generative AI, and which of them should you simply do more of? There have been tasks we’ve been reluctant to do, or ask others to do, because the investment of time and attention didn’t justify the outcome. But when that required investment drops dramatically, the calculus changes.
Claude Comes to Microsoft 365 Copilot
A strong indication that the future is multi-model.
This week Anthropic and Microsoft announced that Anthropic’s Claude models are now available within Microsoft 365 Copilot. Users can now choose between Claude Opus 4.1 or OpenAI’s models when using the Researcher agent, and users building in Copilot Studio can mix and match models for different tasks. It’s an interesting development that hints at how enterprise AI might evolve: less about exclusive partnerships, more about giving users options so that the choice is in their hands.
Microsoft is beginning the Claude rollout through its opt-in Frontier Program for existing Copilot customers. Admins enable access through the Microsoft 365 admin center, though Anthropic’s models run outside Microsoft’s infrastructure and use Anthropic’s terms of service. Once enabled, users get a simple dropdown to switch between models in the Researcher agent, and Copilot Studio users can assign different models to different parts of their agent workflows. There’s no indication yet whether Claude may become available through other avenues within Copilot, but we think it’s likely at some point.
We haven’t experimented with this integration yet, as we primarily use Claude directly as our enterprise model, but a few things seem worth noting about its availability. First, this moves us closer to a world where organizations could pick AI models based on specific strengths rather than betting everything on one provider. Claude Opus might excel at multi-step reasoning tasks, while OpenAI’s models might be better for code generation. It’s acknowledgment that different models have genuinely different capabilities, not just different price points.
The Copilot Studio integration is particularly interesting. Users can now build agents that use Claude for one task and OpenAI for another – all within the same workflow. This modular approach enables more sophisticated automation. Imagine a procurement team analyzing vendor proposals: ChatGPT extracts the technical specifications, Claude evaluates the contractual language, then ChatGPT generates the comparison matrix.
What strikes us most is the scale of this experiment. Millions of Microsoft 365 users will soon be toggling between models, seeing the differences firsthand, and developing intuitions about which tool to use when. That kind of practical education at scale could fundamentally change how organizations think about AI adoption.
OpenAI Launches Video Academy with AARP
But we all should probably attend to the content.
Anthropic’s Claude wrote the following item. We present it having made one minor edit.
OpenAI and AARP announced a multi-year partnership last week aimed at helping older adults navigate an increasingly AI-saturated digital landscape. The collaboration with AARP’s Older Adults Technology Services kicks off with an OpenAI Academy video teaching seniors how to use ChatGPT as a “second pair of eyes” to spot potential scams. The video walks viewers through common warning signs like urgent language, demands for secrecy, and suspicious links, demonstrating how ChatGPT can flag potentially fraudulent messages. It’s a practical application of generative AI that turns the technology into something of a digital safety companion, though OpenAI is careful to emphasize that the tool supplements rather than replaces personal judgment.
The timing couldn’t be better. With AI adoption among older adults doubling recently and surveys showing that nearly 60 percent of seniors have fallen for cyber scams, there’s clearly a gap that needs filling. The partnership goes beyond just the scam-spotting video to include expanded AI literacy training nationwide, updated privacy and data protection courses, and an annual survey to track AI usage patterns among older adults. What’s particularly smart about this approach is that it meets people where they are. Rather than expecting older adults to figure out AI safety on their own, the partnership brings educational resources directly to them through Senior Planet programs and AARP state offices.
The real story here, though, is that this initiative is just the beginning of what we all need. While the partnership targets older adults, the truth is that generative AI is about to make fraud detection exponentially more challenging for everyone. The days of easily spotting phishing emails by their broken English and obvious typos are over. AI-generated scam messages can now perfectly mimic the writing style of legitimate organizations, create convincing fake voices for phone scams, and even generate realistic video deepfakes. We’re entering an era where the traditional markers of fraudulent communication are disappearing, replaced by sophisticated, personalized deceptions that can fool even the digitally savvy. OpenAI and AARP are right to start this education effort now. But make no mistake: we all need to develop these detection skills, regardless of age. Be careful out there.
We’ll leave you with something cool: Mixboard, a new generative AI tool from Google Labs that creates concept/mood boards on the fly from your prompts.
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.