Confluence for 4.13.25
Shopify memo on generative AI. Why you still need to write. ChatGPT gets better memory. Getting more out of Deep Research.
Welcome to Confluence. Here’s what has our attention this week at the intersection of generative AI, leadership, and communication:
Shopify Memo on Generative AI
Why You Still Need to Write
ChatGPT Gets Better Memory
Getting More Out of Deep Research
Shopify Memo on Generative AI
Setting “reflexive AI usage” as a baseline expectation.
An internal memo from Shopify CEO Tobi Lütke created several stirs last week: first when it started to leak, and again when he posted it on X. The memo is worth reading in full, as it represents one of the most assertive pieces of internal communication we’ve seen regarding employee expectations for using generative AI. To sum it up, the memo states that “reflexive AI usage is now a baseline expectation at Shopify.” This is a position much more progressive than we see in most organizations, and there are several points in Lütke’s thinking of which other leaders might take note.
First, Lütke doesn’t position AI usage as a suggestion — it’s an expectation. He believes generative AI will fundamentally change how work gets done, and he’s clear that opting out simply isn’t viable at Shopify: “Frankly, I don’t think it’s feasible to opt out of learning the skill of applying AI in your craft; you are welcome to try, but I want to be honest I cannot see this working out today, and definitely not tomorrow.” This stance acknowledges the reality that AI proficiency is rapidly becoming as fundamental as email or spreadsheet skills are today.
Second, Lütke’s goes beyond mere encouragement by setting expectations for how AI is going to show up in actual processes. The memo requires AI as part of the prototyping process, integrates AI usage questions into performance and peer reviews, and directs teams to explain why AI doesn’t meet their needs when requesting additional resources or team members. These decisions won’t be right for every organization (and they certainly leave open questions even for Shopify), but they demonstrate serious commitment from leadership. We believe that simply saying “use generative AI more” is insufficient to create proper symbolism around generative AI use if you want it to expand in your organization.
Perhaps most important, Lütke says that “everyone means everyone” — including himself and the executive team. When senior leaders exempt themselves from technological change, it undermines the message. By making it clear that this applies at all levels, Lütke bolsters the credibility of his position and signals that this isn’t just another initiative that will fade in time.
The memo still leaves plenty unsaid and as, Ethan Mollick notes, it raises additional questions:
1) What is management’s vision of what the future of work looks like?
2) What is the plan for turning self-directed learning into organizational innovation?
3) How are organizational incentives being aligned?
4) How do I get better?
Still, the memo is a useful example for leaders who believe (as we do) that generative AI will have far-reaching consequences for how organizations work. The key for other leaders will be finding ways to create an equivalent urgency that make sense for their teams and cultures. Not every company will take Shopify’s approach, but every company will need to develop a clear stance on generative AI adoption — preferably before it happens by default, without strategy or direction.
Why You Still Need to Write
The ability to put words into text, yourself, will matter for a long, long, time.
Thousands of years ago, Plato argued against writing in his great work Phaedrus:
And so it is that you by reason of your tender regard for the writing that is your offspring have declared the very opposite of its true effect. If men learn this, it will implant forgetfulness in their souls. They will cease to exercise memory because they rely on that which is written, calling things to remembrance no longer from within themselves, but by means of external marks.
What you have discovered is a recipe not for memory, but for reminder. And it is no true wisdom that you offer your disciples, but only the semblance of wisdom, for by telling them of many things without teaching them you will make them seem to know much while for the most part they know nothing. And as men filled not with wisdom but with the conceit of wisdom they will be a burden to their fellows.
ChatGPT 4.5 translates this into “plain English” as:
Because you’re so fond of your own invention (writing), you’ve misunderstood its real impact. If people rely on writing, they’ll stop using their memory. Instead of recalling things from their own minds, they’ll depend on external notes or marks.
You’ve invented a method to remind people, not to strengthen their memory. What you offer isn’t genuine wisdom but just the appearance of it. You’re filling your students’ minds with facts without teaching true understanding, so they’ll seem knowledgeable while actually knowing very little. Filled with false confidence rather than real wisdom, they’ll end up causing trouble for others.
It’s a caution about second-order consequences: the indirect, unintended effects (good or bad) that follow a decision or innovation, the consequences of the consequence. Second-order consequences tend to be numerous, hard to predict, and sometimes significant.
We spend a lot of time writing about second-order consequences in Confluence, as we think generative AI is going to create many of them. One nest of those consequences wraps around content creation, and in particular, writing. Generative AI has become an amazing writer, and with sound prompting and stylistic guidance, in speed-for-quality comparisons we feel generative AI already exceeds the ability of most people on the planet. It’s natural to want to use generative AI to write, and a lot of people do, and in many cases, they should. But they should do so knowing that this choice has second-order consequences.
One is skill erosion. Nobody is going to lose the ability to write by becoming overly reliant on generative AI, just as nobody has lost the ability to remember things by becoming overly reliant on writing. But people certainly don’t use memory now as they did prior to Plato. One could argue that we are less proficient at certain types of memory, and we expect that outsourcing all your writing to generative AI will make you less proficient at certain types of writing (although we are not quite sure yet what those types will be).
A second is “slop,” which is the low-quality, sometimes cluttered content that large language models can produce when people don’t prompt them well. The internet is filling with slop, and while the research shows that humans (and computer programs) cannot differentiate well-prompted AI-generated writing from human writing, anyone with real time working with large language models gets pretty good at detecting slop. Those who don’t recognize slop as slop simply recognize (and struggle to understand) poor writing. Slop is not great, and we don’t relish a world where 9 out of 10 news sites you visit online are filled with low-quality, AI-generated text. We also fear a world where people quickly crank out an ever-increasing volume of AI-generated PowerPoint or Microsoft Word slop inside organizations. We need clarity in organizations, not ambiguity.
But there’s another second-order consequence to generative AI writing that has our attention, and our colleagues who write our firm’s sister-newsletter, Admired Leadership Field Notes, wrote about it yesterday: damaged authenticity and sincerity. We won’t repeat too much of that Field Note here and instead encourage you to read it (like all Field Notes, it’s quite brief). But here’s the gist:
Using AI to compose or assist in writing emails, policy statements, instructions, and presentation decks (among many other everyday messages) will save time and improve quality. Everyone will be expected to do so without any criticism.
But more personal messages, such as blog posts, personal letters, heartfelt expressions, advice, and requests for assistance (among many other symbolic messages) will be scrutinized and treated harshly if and when AI is used to create them. Not by everyone, of course, but by those who think that authenticity matters. Which will be a sizeable audience.
People will soon increasingly seek out music, art, novels, and other forms of creative expression that originate directly from artists and that do not use AI … sincerity and authenticity cannot be delegated to any intelligence other than your own.
We’ve been talking about authenticity in this space almost since the start, as we believe it really matters for leadership and communication. As content generated by AI becomes more common, people will start to question the authenticity of the content they see, and this is especially meaningful for leaders and communication professionals who believe their credibility and character matter.
It’s a question of expectations, and it informs a debate we’ve been having within our firm for almost two years as we created and continue to update our internal principles for generative AI use: what will people expect, and what is true to our values as a firm? Today nobody would be critical of someone using computer-automated spell check and grammar check on their documents, just as nobody would be critical of a senior leader delivering comments at a town hall written by a communication professional. But would you be critical if you were to learn that generative AI wrote the piece you are reading now, and we didn’t reveal that to you? (It did not … this one is all human, although as our disclosures at the bottom of each issue make clear, we do use generative AI in the process of creating Confluence, and sometimes that does involve creating content.) And how would you react if you were to learn that ChatGPT was the author of the very kind congratulatory note sent to you by your Senior Vice President thanking you for your service on a recent project? Or that Claude AI wrote the toast a friend delivered at your son’s wedding?
This is a big watch out for leaders and communication professionals. If you take your credibility seriously, you must stay in-tune with the expectations people have for when it’s your hands on the pen or keyboard. And those expectations will shift. The simple rule we’ve written about here many times is “the icky standard”: If it would feel awkward for your audience to learn how you involved generative AI in producing your content, you need to do that work yourself.
We welcome a world in which generative AI is deeply involved in content creation. Used well, we think it raises all boats on its tide of speed and quality. We use it every day, to the great benefit of our people and our clients. But in the short term (and if we are honest, also in the long), writing is still going to matter, be that for reasons of skill or slop or sincerity, just as memory still matters. No matter where all this goes, you still need to write. We think leaders and communication professionals should write every day, even if in a journal. The ability to convey words in text with sincerity and feeling is, for us, an essential part of being human, no matter how much artificial co-intelligence is part of our day-to-day existence.
ChatGPT Gets Better Memory
The model will soon remember everything you’ve talked about.
This week, ChatGPT began rolling out a considerable upgrade to its memory features. The update, which is initially available to Pro and Plus users (except in the UK, EU, and some European countries due to regulatory considerations), essentially allows ChatGPT to remember everything you’ve told it across the entire history of your chats. OpenAI CEO Sam Altman described this change as bringing the technology closer to his ultimate goal of “AI systems that get to know you over your life.”
The idea of “infinite memory” is not new, and Google announced it was coming to its Gemini model back in February. But ChatGPT has a larger market presence and is much closer to being a household name than Google’s Gemini. ChatGPT being able to remember things is also not new. OpenAI first announced its intent for the model to have (selective) memory over a year ago, and the model does regularly store things away and reference them as needed. What is new, though, is a memory that stores everything you talk about. Imagine talking with a close friend or colleague who has an uncanny ability to recall major and minor things you mentioned weeks, months, or even years ago.
But there are interesting questions implicit in that comparison: Would you want a colleague to remember everything you’ve mentioned, professional or personal? Would it be useful — or even desirable — for a friend to understand every step you took to draft that important document at work? While you can toggle ChatGPT’s memory on and off (and even use a “Temporary Chat” that has no memory) we’re interested to see how people who fluidly use ChatGPT manage this now-blurred boundary line. And for leaders, long-term AI memory raises important questions about how organizational knowledge and personal information might intermingle in these increasingly personalized AI systems.
Despite these unanswered questions and others that will surely arise once we get more time with this feature, we’re looking forward to using this capability. It should be a significant step toward more natural AI conversations that build on shared history, but the real test will be how well we can leverage these memory capabilities while maintaining boundaries between professional and personal use.
Getting More Out of Deep Research
We tested a new approach shared by Alex Lawsen and were very impressed with the results.
We wrote in February about ChatGPT Deep Research’s stunning ability to produce in minutes high-quality reports and analyses that would have taken us days. We continue to find utility in Deep Research and use it several times a week (in some cases, several times a day) to get smarter on a wide range of topics, both professional and personal. We were intrigued, then, to see Alex Lawsen introduce a new approach that claimed to get even more out of Deep Research tools (including not just ChatGPT’s but Google Gemini’s as well). The approach, which Lawsen lays out in a post on his Substack Speculative Decoding, involves creating a Claude Project that produces optimal Deep Research prompts. As Lawsen describes it, the Project “handles all the repeatable high-effort prompt engineering (like source selection) for you. It asks for your preferences, clarifies what you need, and produces a well-structured prompt that you can then feed into any DR model.” We decided to put it to the test.
Creating the Claude Project is relatively simple, and Lawsen lays out the detailed instructions in his post. It took us about 10 minutes. With the Project up and running, we started with a simple query:
I want to create a report on the implications of emerging media technologies like AR, VR, and metaverse-adjacent tech on media richness theory — and, specifically, what that might mean for corporate communication.
The Project then asked us seven short follow-up questions: two about the purpose of the research, three about its scope and focus, and two about the output format. We provided simple answers to those questions and, within 30 seconds, received a detailed, nuanced prompt to use in Deep Research. The exchange took about two minutes. Here’s the prompt we received:
Research Prompt: Media Richness Theory in the Era of Spatial Computing
Goal: Conduct a comprehensive analysis of how emerging spatial technologies (AR, VR, MR, and metaverse platforms) are reshaping media richness theory, and identify the strategic implications for corporate communication practices. Examine both theoretical extensions to the original framework and practical applications for communication professionals. Cover near-term applications (0-3 years) as well as longer-term potential (3-10 years). The analysis should identify specific opportunities, challenges, and recommended approaches for organizations seeking to leverage these technologies for internal and external communication.
Output format: Deliver a comprehensive research report (6,000-8,000 words) following a progressive disclosure structure that allows readers to engage at different levels of depth. Begin with an executive summary that captures all key findings and implications in no more than 10% of the total document length. Include clearly linked sections that lead readers from high-level insights to supporting evidence and detailed analysis. Use informative headings and subheadings that preview content rather than generic labels.
The report should include:
An executive summary with linked bullet points to corresponding detailed sections
A thorough introduction to both media richness theory and the current state of spatial computing technologies
A detailed analysis section examining how each major technology extends or challenges the core principles of media richness theory
A framework for evaluating these technologies against communication objectives
Practical applications and case studies highlighting current implementations
Strategic implications for corporate communicators, including specific recommendations
Future directions and areas for continued research
A balanced consideration of limitations, risks, and ethical considerations
Use concrete examples, visual representations (described in text), and comparative analyses where appropriate. Balance theoretical depth with practical applications, ensuring each informs the other.
Warnings:
Evaluate sources based on domain-specific credibility hierarchies:
For theoretical extensions to media richness theory, prioritize peer-reviewed academic articles in communication and information systems journals, followed by books from established academic publishers, and conference proceedings from recognized academic associations.
For technology assessments, prioritize research from independent technology research firms (e.g., Gartner, Forrester), technical documentation from platform developers, and analyses from specialized technology journalists with established expertise.
For corporate applications, prioritize documented case studies, white papers from consulting firms with expertise in communication technology, and industry publications with rigorous editorial standards.
Be skeptical of promotional materials, vendor claims lacking independent verification, and speculative predictions without substantive evidence.
Avoid common misconceptions and pitfalls:
Don't assume media richness is always beneficial; consider contexts where leaner media may be preferable.
Don't conflate technological capability with effective implementation; address organizational and human factors.
Avoid technological determinism; consider how organizational culture, individual preferences, and power dynamics influence technology adoption and use.
Be cautious of overgeneralizing findings from one context (e.g., consumer applications) to another (corporate communications).
Distinguish between empirically supported claims and theoretical predictions yet to be validated.
Additional Context: Media Richness Theory (MRT), originally developed by Daft and Lengel in the 1980s, proposes that communication channels vary in their capacity to process rich information. The theory suggests that effective communication results from matching media richness to task equivocality, with richer media better suited for ambiguous tasks and leaner media for straightforward ones. The theory traditionally ranks communication channels along a continuum from rich to lean: face-to-face, video, audio, and written communication.
Emerging spatial technologies potentially disrupt this continuum by:
Creating new forms of presence and embodiment that may rival or exceed face-to-face interaction
Enabling novel forms of synchronous and asynchronous collaboration
Introducing new modalities for information representation and interaction
Blurring boundaries between physical and digital communication environments
Potentially altering cognitive and emotional engagement with communication content
Corporate communication encompasses multiple domains, including:
Leadership communication (executive messaging, vision/strategy dissemination)
Employee communication (change management, engagement, training)
Team collaboration (project coordination, knowledge sharing)
External stakeholder communication (customer interaction, investor relations, public affairs)
The analysis should consider how spatial technologies might differentially impact these domains and examine implications for:
Message fidelity and interpretation
Audience engagement and retention
Credibility and trust
Bandwidth and efficiency
Accessibility and inclusion
Organizational structure and culture
Additional considerations should include:
Technology implementation challenges (cost, infrastructure, adoption barriers)
Measurement approaches for evaluating effectiveness
Privacy and security implications
Comparative advantages relative to existing channels
Integration with broader communication ecosystems
This research should provide both theoretical foundation and actionable insights for communication professionals seeking to understand, evaluate, and potentially implement spatial technologies within corporate communication contexts.
So, yes: detailed. We then fed this prompt, verbatim, into ChatGPT’s Deep Research and six minutes later received an extremely thorough report that, when pasted into Microsoft Word, stretched to 72 pages. Here’s the table of contents:
So, yes: thorough. But is it good? We read the report with an eye toward novel insights and practical application. We found that the report included plenty of both. It was by no means perfect or publication-ready, but it certainly made us smarter and deepened our thinking on a topic which we think will be increasingly important in the years to come. And it took all of 15 minutes from the initial prompt in Claude to the final report in ChatGPT.
To conduct a rough test of this approach against the approach we had previously been using for Deep Research (simply creating our prompts “by hand”), we used Deep Research to create another report on the exact same topic, but without using our new Claude Project. The resulting report was 23 pages and again took only a few minutes to generate. It was strong, but clearly not as comprehensive or compelling as the first report we created with our more thorough prompt. We then gave both reports to both Claude 3.7 Sonnet and to GPT-4.5 and asked each model to evaluate them against each other. Both concluded that the first report — which, again, utilized the more thorough prompt we created using our new, Lawsen-inspired Claude Project — was clearly superior:
Claude 3.7 Sonnet: Report #1 is clearly the superior report overall. It provides a more comprehensive, structured, and practically useful analysis of the topic. The inclusion of evaluation frameworks, detailed recommendations, ethical considerations, and future directions makes it more valuable for both academic understanding and practical implementation.
Report #2 is still a strong research report with valuable insights, but Report #1 goes beyond it in breadth, depth, structure, and applicability, making it the better resource for corporate communicators and decision-makers navigating these emerging technologies.
…
ChatGPT-4.5: Both reports offer valuable insights, but Report #1 stands out as the superior document overall. It excels in providing a thorough theoretical grounding in Media Richness Theory, delivers more detailed analyses and nuanced distinctions between AR, VR, MR, and metaverse platforms, and includes richer practical examples and case studies that clearly illustrate real-world applications. Additionally, it provides a structured evaluation framework that offers practical guidance for communication professionals. The primary strength of Report #2 lies in its concise readability and stronger emphasis on potential intellectual property implications, making it useful in contexts specifically focused on IP strategy. Ultimately, for a comprehensive and strategic exploration of emerging immersive media in corporate communication, Report #1 is recommended as the primary resource, while Report #2 is better suited for quick orientation or when intellectual property considerations are paramount.
Two months ago, Report #2 would have blown us away. Thanks to Alex Lawsen’s curiosity and generosity in sharing his approach, though, we know we can do even more. We’re looking forward to continuing to experiment with this approach, and we encourage our readers to give it a try as well.
We’ll leave you with something cool: A team from U.C. Berkeley, Stanford, and Nvidia have made a breakthrough on extending the length of AI generated videos, using old Tom and Jerry cartoons as a proof-of-concept.
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.
I really enjoy your newsletter, thank you. The Deep Research + Claude was something I'd never considered