Working With AI
Reading Packets

Reading Packets

The research behind the series, in a bit more depth than the courses have room for, with a few questions worth sitting with.

Each course in this series rests on a small number of studies and a couple of long-standing ideas from cognitive psychology and decision science. None of this is required before starting a course. All of it rewards fifteen minutes if you want to know where the frameworks came from, and each section ends with a short reference list so you can go to the source directly.

For Build Your AI-Powered Second Brain

Where the second brain idea comes from

The idea of writing down thinking so it can be found and reused later is old. What made it a discipline rather than a habit was Niklas Luhmann, a German sociologist who, working largely alone, produced roughly seventy books and several hundred articles over his career. Luhmann credited an unusual tool: a card index, built up over decades, that by some widely cited estimates held close to ninety thousand cards. He described it as a conversation partner rather than a filing system, something that talked back by surfacing connections he had not gone looking for. The practitioner literature on this method, most notably Sönke Ahrens's How to Take Smart Notes, distilled the working habits into rules a modern reader can apply directly: one idea per note, and notes linked to each other rather than buried in folders. Those two rules are where this course's "one file per subject" principle and its emphasis on writing for retrieval both come from.

The AI-era half of the argument comes from research on cognitive offloading, the use of an external tool or store to reduce the mental effort a task would otherwise take. Risko and Gilbert's 2016 review in Trends in Cognitive Sciences lays out the mechanism: offloading is a trade, not a free lunch, and how much a person offloads depends on how they judge their own memory, a judgment that can itself be wrong. A more specific and more famous finding sharpens the stakes. Sparrow, Liu, and Wegner's 2011 study in Science, widely known as the "Google effects" study, found that once people expect information to be available externally, they become worse at recalling the information itself and better at recalling where to find it. That is the empirical basis for this course's warning that a stale file is more dangerous than no file: once you trust a store exists, you stop keeping the content in your head, and the store's accuracy becomes load-bearing in a way it was not before.

Questions worth sitting with

  1. Luhmann treated his card index as something that talked back, surfacing connections he was not looking for. Where in your own system could an unexpected connection actually surface, as opposed to a folder that only returns what you already knew you filed there?
  2. The Google-effects study found that people stop remembering content once they trust it is stored externally. What would it cost you if a file in your second brain quietly went stale and you did not notice, because you trusted it was current?
  3. Offloading is a trade: less effort now, more dependence on the store being right later. Which parts of your work are safe to hand over completely, and which ones still need to live in your own head regardless of what you write down?

References

For Communication & Influence

Where the workslop and signal research comes from

In September 2025, BetterUp Labs and the Stanford Social Media Lab surveyed 1,150 full-time U.S. desk workers and gave a name to a problem everyone had noticed but no one had measured: workslop, AI-generated work that looks polished but lacks the substance to move a task forward. Forty percent of respondents had received it in the past month, and those who had estimated that roughly fifteen percent of everything crossing their desk qualified. The researchers' most striking finding was not the time cost, close to two hours per incident, but the social one: over half of recipients rated the senders of workslop as less capable, and 42 percent said they trusted those colleagues less afterward. The paper's framing of this as an "invisible tax" is deliberate. Nobody bills for the two hours; it simply disappears from everyone's week, and the trust does not come back at the same rate it left.

The reader's side of the same problem shows up in Korbyt and Reworked's 2026 State of Workplace Communication survey of 1,175 U.S. employees. Its most counterintuitive result is that roughly half of workers say the volume of communication they receive is "about right," while 44 percent still report tuning out, a gap the researchers read as passive disengagement: satisfaction with volume can mask the fact that people have simply stopped absorbing what arrives. The same survey found that 81 percent of workers believe they can usually tell when a message was written by AI, and that 92 percent want AI used specifically to reduce the volume of information they have to process, not add to it. Read together, the two studies describe a closed loop: AI makes it cheap to produce more communication, readers are getting better at discounting what that produces, and the loop only breaks for a message that is worth the specific person's specific attention.

Questions worth sitting with

  1. Half of the people in the BetterUp/Stanford study said workslop made them see the sender as less capable. Has a piece of your own writing ever cost you that kind of credibility, and did you know it at the time?
  2. Korbyt and Reworked found that people who call their communication volume "fine" can still be quietly tuning out. What would that look like from someone on your own team, and would you notice it before it showed up somewhere else?
  3. Ninety-two percent of workers want AI used to reduce information, not add to it. Where in your own use of AI this month were you adding volume instead of cutting it?

References

For Pressure-Test My Strategy

Where the stress-test method comes from

PwC's surveys of corporate directors and C-suite executives capture a gap that is easy to underestimate. In one survey, 99 percent of executives said their boards should be using AI in oversight; in a companion survey of directors, only 35 percent said their boards currently do. A gap that size, between what leadership expects and what boards have actually adopted, tends to close fast once the tools are this accessible, which is the practical argument for stress-testing a recommendation before that gap closes on you.

Two older ideas from decision science do the rest of the work. The planning fallacy, named by Kahneman and Tversky in a 1979 paper on intuitive prediction, describes a specific asymmetry: people are systematically more optimistic about their own project's timeline and cost than an outside observer looking at similar past projects would be. The correction, anchoring a forecast to a reference class of comparable efforts before adjusting for what is genuinely different this time, is the direct source of this course's insistence on asking what a projection's reference class actually is before believing its bend.

The pre-mortem comes from a 1989 study by Mitchell, Russo, and Pennington on what they called prospective hindsight: generating an explanation for an event by imagining it has already happened, rather than merely asking whether it might. Their finding, later popularized by Gary Klein's 2007 Harvard Business Review article on the technique, is that treating an outcome as already real measurably improves people's ability to identify its causes, by an often-cited margin of about thirty percent over simply asking what could go wrong. The mechanism is specific: "what could go wrong" is answered from a posture of optimism, while "explain why this already failed" is answered from a posture of explanation, and explanation surfaces more and sharper causes than speculation does.

Questions worth sitting with

  1. If the gap between what boards expect and what they currently do keeps closing, what in your next recommendation would not survive a first pass from a director's own AI, run before the meeting rather than during it?
  2. Mitchell, Russo, and Pennington's subjects got roughly thirty percent better at spotting causes once they treated the failure as already real. Have you ever run an actual pre-mortem, out loud, on a decision that mattered, rather than just asking what might go wrong?
  3. The planning fallacy says your inside view of your own project is systematically more optimistic than the outside view of similar ones. What is the nearest reference class for the recommendation you are working on right now, and have you actually looked at how it performed?

References

BackAll courses NextAI Reference Guide
Working With AI. A self-paced series built for use with any AI assistant.