A bi-weekly roundup of sycophancy in 9 minutes
August 22 - September 4

a scale weighing time and done, with done weighing down the balance

Remembering How To Human

Quick note: September 16-18th I'll be at HubSpot's UNBOUND conference, doing tarot card readings about AI fluency/anxiety. I've set up a Braindate listing if you're going to be there, or just look for the guy in the Learn to Scale hat and ask for your fortune.

I learned how to design trainings way back in my first job after college, in an AmeriCorps program at a charter school. Through a blend of rigor and care, the school tried to get kids from a 4th grade reading/math level up to a 7th grade level in 180 days.

I'm often surprised by people's reactions when they see the level of detail that I put into my lesson plans for running trainings, but when they hear that goal of 4th grade to 7th grade in 180 days, they start to understand why my outlines are voluminous. Every minute is accounted for. Transition sentences between each five minute block are pre-written with just the right amount of levity. Discussion questions get a forecast of anticipated right and wrong answers.

(side note for my future TED talk: it looks very similar to a script for a play. Imagine what training-as-theater feels like as a participant.)

With this week's release of the two most cutting-edge AI models, Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra, I thought it poetic to snip a section from my AI Fluency for a 100% Human Workforce training program to help remind all of you how to be a human in the face of eerily human machines. And since you all are TL;DR subscribers, I want to show a bit of my learning design thinking BEHIND the training on how to be human.

TL;DR- Please enjoy my musings on how to help humans take human-like AI outputs and check that they're human enough.

Welcome to module 2 of the AI Fluency course, Practical AI Quality Control & Human-in-the-Loop. By now, you would have already gotten a primer on generative AI, discussed what AI safety guardrails should look like in the language of day-to-day work, and crafted a pragmatic set of "I will/I won't" commitments that make those guardrails easy to follow.

Dan's rationale: AI, to people who haven't touched it, can be terrifying. By making the green light and red light around using it safely really practical, learners drop their guard and the fear moves to the backseat. This needs to happen first, otherwise my most scared learners will be second-guessing everything they and I do. By the time we're in this second session, people know how my training will feel and are ready to tackle more challenging topics.

You've joined the Zoom call and Dan reviews some of the safety guardrails we covered in the previous section. Double checking AI-generated work, things we won't use AI for (hiring, promotion, compensation, etc.), what tools are safe for sensitive information, etc. It's a quick friendly refresher-

-and then Dan drops a scenario that totally evades these guardrails, catching you off-guard.

Rationale: This is the Hook for the training. That emotional "uh oh" feeling connects a real-world situation to the training content. It's a more effective "What's In It For Me" than telling someone that the training is important.

Here's the scenario: AI helps you draft an email to one of your direct reports in seconds — warm, balanced, respectful, identifying "delegation and trust" as a growth area. It's on-brand, constructive, and doesn't throw anybody under the bus. It's good, so you approve it and send it.

What the polished text walked you past: this direct report had one of their team members quit over micromanagement on Tuesday. Everyone in the office knows and your direct report is raw about exactly this. The kind note lands like a public verdict. She checks out.

The snooker: "The machine didn't know about Tuesday. Fine. The draft was finished enough to make YOU forget her too. Hand-writing it would've put her face in front of you; approving a polished draft let you skip her. You will serve what AI serves you unless you stop and add back the one thing only you know: human empathy."

Rationale: This is a key learning outcome for this training: recognizing how AI will seduce you and what gets lost when it succeeds. By couching it in a narrative that ends with "You, the boss, allowed AI to put your humanity on copilot and it hurt someone," learners get to experience the failure without actually failing.

At this point, Dan starts a group discussion, "Good news, everyone- nobody did this...yet. Situations like this are inevitable, though. Let's dive back into that scenario and diagnose three things:

What did AI do?
Where exactly was the checkpoint?
What one question would've caught it?

Rationale: By de-escalating the heat but not excusing it, I'm reiterating that this activity will matter back on the job. The activity itself is a Pre-Mortem where we take a fictional future event and walk backwards to identify what could go wrong. It's a good start-of-session activity because it welcomes all levels of learners and is low-stakes enough that the most junior person's contribution is just as valid as everyone else's.

The group starts to analyze the scenario. You chime in with a "the AI sounded confident in its response" and Dan presses deeper, "What does 'confident in a response' really look like?"

Rationale: Discussions engage people because they're putting their own thoughts into the room, but often those thoughts need to be fleshed out more in order for the learning goal to be truly met. When facilitating a discussion, I have a specific destination I need to reach and by asking for clarification, it helps a learner think one step further, then one step further, until we've reached the full concept.

A little on the spot, you say, "Well, confident looks like...hmm, certainly a logical sentence. Grammatically correct, for sure. Confident responses have a concrete or specific request that matches the situation. And if the AI response has a little human joke or way of wording the email, it doesn't feel mechanical. It sounds...I dunno, confident."

Dan responds, "Awesome! What I'm hearing: grammatically sound, internally consistent, specific request, maybe a 'load-bearing' joke, this is AI confidence in a nutshell, right?"

You nod, "Yeah, that's right!"

Rationale: When someone is sounding off their thoughts, live, they rarely remember what they said. By restating back their thoughts in a more concise way, they are assessing their own work and embedding the conceptual recipe, thus making the concept stickier. They also adopt the confidence I had in my summary, giving them a sense of self-efficacy that yes, I can learn this.

Dan then turns to the room, "Ok, if this is what a confident AI-generated answer looks like, something that's kind and grammatically correct and aligned with what you wanted it to do...what was the checkpoint? Where could this feedback email have been caught before it ruined someone's day?"

Another person in the room raises her hand, "When a confident AI-generated response looks like it would complete my task, all I have to do is hit the Send button and move on. One click, and I can go do something else. AI gave me exactly what I wanted, but what I really needed was to change my goal and be a little more sensitive with feedback."

Dan screams. "OMG, that's such a great response! Yes, exactly that: the second the draft looked finished your job shrank from "Give feedback that might take 20 minutes to write" to "Approve feedback email in one click," you took the bait. This is one very real way that AI can hijack your critical thinking. Now, let's talk about the questions to ask yourself to defend against that hijacking..."

and the training continues...

That whole exchange would take about 9 minutes, which is ten percent of this 90-minute training module. But look at what happened in those 9 minutes:

Refreshed the course content from another training
Obtained buy-in to today's training on an emotional level
Defined AI sycophancy in a participant's own words
Showed how AI sycophancy shows up in day-to-day work

The real learning outcome here is that those 9 minutes spotlighted one reason why we choose sycophancy. When AI converts 20 minutes into one click, it's really hard to choose the uncomfortable 20 minutes. Even though you may "know" AI will tell you what you want, you won't notice the cost of that one click until your direct report quits on you.

Fable 5.1 and Astra came out this week. They're not only really good at turning 20-minute tasks into one click, but turning hours of work into zero clicks. And what does zero clicks actually look like on a Tuesday? Astra drafts the feedback, schedules the 1:1, and sends the recap before you've read it.

And even though we know this seduction is coming, we won't notice the cost of zero clicks until our critical thinking skills quit on us.

The other 81 minutes of that module exist. It's all about critical thinking, complete with load-bearing jokes.

The whole AI Fluency for a 100% Human Workforce training program can be zero cost to your organization, thanks to the Massachusetts Express Grant. Grab some time with me and I'll show you how.

Show Me The Other 81 Minutes

A TL;DR from the CRO

My critical thinking skills get their daily exercise by going outside and watching the turkeys: they seem like they can't fly, but they can? More analysis is needed.

-Roman Noodles, Chief Ruff Officer

This weekend I'm headed to a wedding that's summer camp. Not summer camp-themed, but summer camp. Color wars, music hour, boats and campfires, and some wedding stuff too but also CRAFTS.

I'm about 65% summer camp generated. Growing up, I attended orchestra camp, band camp, piano camp, rec park camp, archeology and art camp...is anyone surprised that the kid who was dropped into groups of strangers for 12+ years every summer freaking loves icebreakers as an adult?

One of my most vivid formative memories was at Summer Sonatina. Every morning you got to choose what non-piano activity you got to do for the day and on some Wednesdays they had horseback riding. It was always a hot commodity because there were only five spots, it wasn't every week, and you only got to sign up after your bedroom was "cleared" by a camp counselor.

If we knew that horseback riding was on the menu, my bunkmates and I would get up early, clean the room, and then try to sweet talk the counselor to check our room first so we could be first to sign up for horseback riding.

The ONE time we pulled off this heist, I got my name on the list. I was so excited. It was my first horseback riding experience. I think I was 12.

What I most vividly remember when I finally got plopped on the saddle was TERROR. I was atop a huge animal that was easily twice as tall as me. I was terrified of falling off, of the horse kicking me, of being trampled into Dan-goo. They had us wear helmets and to my 12-year-old brain, this was because horses regularly stomped on children.

On this Wednesday, death was neigh.

When the trainer tried to get us riding at a canter, I absolutely refused. The rest of my horseback riding experience was just a haze of watching everyone else while I white-knuckled my anxiety on the probably-most-chill pony at the stable.

The pony was the one-click option. Everyone else got the twenty minutes.

Don't worry, I've ridden horses since that fateful Wednesday and have been stomped zero times, but boy do those memories linger. I'm a little apprehensive that the hora will happen on horseback this weekend.

Keep riding through the fear,

Dan from Learn to Scale

Opt-out from the newsletter | Unsubscribe from all emails | Update your Preferences | www.learntoscale.us, Boston, MA 02119 |

PS. This is a personal one! I dug up a video I made while I was a camp counselor at that piano camp. We were teaching the context behind Berlioz's Symphony Fantastique where the "artist" (played by me) takes a non-lethal dose of opium and hallucinates that he's being marched to the scaffold and executed. You won't believe how it ends.

Dan Newman

The TL;DR, You Know AI Will Seduce You But You'll Click Anyways

Remembering How To Human

A TL;DR from the CRO

The TL;DR, My Entire Year of AI Runs on Two Gallons of Gas

The TL;DR, Fighting With AI Makes Me Less Stupid

The TL;DR, Predicting Dire Financial Rhymes