AI Manifesto: Stage 4, Upskill
Train Scepticism, Not Button-Clicks
Train Scepticism, Not Button-Clicks
In most companies, “AI training” has become shorthand for prompt tips and a few examples of what the tools can do. It’s understandable. It’s also incomplete.
Stage 4 in the framework is Upskill, and it asks a question that cuts to the chase (and the clear risk): “Can our people use AI confidently, critically, and safely?” The framework doesn’t centre “how to get output”. It centres how to evaluate output. Train people to evaluate AI outputs critically, not just accept them. Cover prompt craft, verification, and knowing when not to use AI. Build AI literacy at every level, from leadership to frontline. Develop AI champions inside teams who can support peers. The guardrail is the quiet line that prevents a lot of pain: no tool rollout until people know how to question what it produces.
This is where the manifesto’s “Humans Decide. AI Assists.” becomes operational. If humans remain accountable, they need the practical competence to challenge outputs, detect uncertainty, and recognise when a task is unsuitable for AI in the first place.
What Aviation Gets Right
Aviation is the cleanest comparison because it shows what “human in the loop” looks like when the consequences are real. Modern cockpits are full of automation. Pilots are trained to use it and trained to override it, because over-reliance is a known hazard. Research and safety literature describes automation-induced complacency as a risk where people become less vigilant because the system seems to be behaving. You can swap autopilot for generative AI and the pattern holds: the output looks plausible, so the human stops thinking.
Why Verification Can’t Be Left to Memory
Healthcare offers a different lens. Atul Gawande popularised the idea that in complex environments, failure often comes from inconsistency and missed steps, not lack of expertise, which is why checklists can change outcomes. The point for AI is not that every workflow needs a checklist. It’s that verification can’t be left to mood and memory. Upskill is where you teach structured habits: how to check sources, when to escalate, what needs a second pair of eyes, what you never paste into a model.
Request the full AI Manifesto
The Leadership Gap Nobody Talks About
Good Upskill programmes also fix a common leadership gap. Executives don’t need to become prompt specialists. They do need fluency in commissioning AI change. They need to ask the right questions: What is the definition of “good” for this use case. Where does accountability sit. What is the human review step. What data is in scope. What would cause us to stop. Those questions connect back to manifesto themes like engineering trust, treating data as a privilege, and treating bias as a defect.
Why AI Champions Are Not Optional
The “AI champions” element matters because adoption spreads sideways. People copy peers more readily than policies. Champions also reduce the shadow-use problem: the quiet spread of tools without shared standards. If you want human oversight to be real, you need local expertise that can translate principles into day-to-day practice, and you need a safe route for raising concerns without being labelled anti-innovation.
The Skill Atrophy Risk
The best Upskill efforts also address a less obvious risk: skill atrophy. If AI drafts everything, juniors lose the necessary repetition that builds exposure, experience and judgement. If AI summarises everything, teams stop reading primary sources. If AI explains everything, people stop asking the hard question: “What do we actually know?” This is why Stage 4 isn’t “train people to use AI”, it’s “train people to use AI critically”. The demise of critical thinking is one example of this skill atrophy.
AI Assists. Humans Decide.
Done well, Upskill produces a culture where AI output is treated as a draft, not a verdict. People learn to ask for evidence and context. They understand the difference between low-stakes assistance and high-stakes decisions. They know when to stop.
In other words, they become capable of keeping the manifesto’s promise: AI assists, humans decide.
Request the full AI Manifesto