Dionn Schaffner took her first AI class in 1991 at Stanford, when computing time was scarce enough that students had to schedule access in server-filled basement labs. A single modern smartphone now contains more computing power than that entire facility. She opened Running Remote 2026 with this not as nostalgia but as a calibration: the democratisation of computing power is not new. What is new is its pace, and what that pace demands of leaders trying to coordinate distributed human teams around it.
Her central argument is that the core skill of AI adoption is not tool fluency, not prompt engineering, not technical expertise. It is judgment — the human capacity to decide what matters, whose experience counts, and who owns the outcome. And unlike most skills, judgment can be built into your systems deliberately rather than relying on it to emerge from individuals.
The strider bike versus training wheels
Schaffner used the difference between training wheels and strider bikes to frame her approach to governance. Training wheels let children move before they develop balance — they prioritise speed over capability. Strider bikes (which have no pedals) teach balance first, using the child’s own feet as feedback, before adding speed and complexity. Children who learn on strider bikes typically skip training wheels entirely and go straight to cycling.
The analogy maps directly to AI governance. Training wheels governance is reactive, restrictive, and creates dependence — it tells people what not to do rather than building the capability to navigate on their own. Infrastructure governance provides safe streets to ride on: data guidelines that clarify what can be ingested, sandbox environments for practice, model guidance through pre-approved tools, clear escalation paths. It enables sustainable speed by building the capability, not constraining it.
Human judgment and the AI slop problem
The industry has evolved from the era when AI use was hidden — where people avoided certain phrases or writing patterns to avoid being detected — to the current moment where ‘drafted with AI’ or ‘sent by your AI agent’ is openly announced in email signatures. The question is no longer whether people use AI. It is whether the human judgment that makes AI output valuable is present in what they produce.
AI slop is Schaffner’s term for what happens when judgment is absent: generic, personality-free content that could have been written by anyone, reviewed by no one, and carries no accountability. It is a leadership failure, she argued, not a tool failure. The same way ‘sent from my iPhone’ stopped being an excuse for typos years ago, ‘AI generated’ will stop being an excuse for low-quality output. The critical element is not attribution — it is ownership. There must be a human who says: this piece of work belongs to me, and I stand behind it.
Outcome accountability design
Many organisations have C-suite mandates to use AI without any coherent theory of how to measure actual impact. Logons, sessions, token usage — these are activity metrics, not outcome metrics. Meanwhile, shadow AI proliferates as employees use personal accounts or unapproved tools outside governance frameworks.
Schaffner drew on her Stanford education to explain why tool-specific training misses the point. She learned multiple programming languages not to master each one but to understand fundamentals — data, logic, flow — that transfer across any language or tool. The same principle applies to AI. Teach fundamentals, not tools. What was here yesterday will be replaced by what is here tomorrow.
Her define-build-connect framework: define what good outcomes look like clearly enough that people can self-direct; build sandbox environments with representative data and reusable components where experimentation is safe; connect learnings across the organisation through channels and forums where what one team discovers becomes available to everyone else.
Architecting dissent and participation
Psychological safety is more difficult to create in distributed environments than in co-located ones, for a structural reason: the informal dissent pathways that exist in physical spaces — the pre-meeting coffee conversation, the hallway check-in, the look across the table — do not exist remotely. This is not a cultural failure. It is an architectural one. And architectural problems require architectural solutions.
Schaffner described a research project that created an AI Devil’s Advocate — a tool that allowed employees to share dissenting views privately, which the AI would anonymise and package for leadership. The result was significantly more dissenting viewpoints reaching decision-makers, because the social risk of speaking up had been removed.
She offered three structural mechanisms for any distributed team. A designated challenger: rotate the role of arguing the opposite position deliberately, as a structural feature of decision-making, not as a personality trait. An async dissent channel: a low-stakes Slack channel specifically for raising concerns, which is particularly valuable for introverts and for people whose cultural background makes direct disagreement uncomfortable. Separating generation from evaluation: AI generates options and analysis; humans evaluate with judgment.
The closing audit question she left the room with: does your team have a structural way to say no, or does it require courage every time? If the answer is courage every time, you have designed a system where dissent is heroic rather than normal — and heroic dissent is rare, expensive, and unreliable. Design the safe street. Balance before speed.
