The AI Steering Tax in L&D

Hey folks 👋

Over the last month or so I’ve been in anthropologist mode, observing ~100 L&D professionals use AI across their real work — every prompt, output and decision logged.

1. People who design & develop learning in the workplace want AI that challenges them, not AI that does work for them;

2. What L&D folks want from AI isn’t consistent — it changes at every stage of the workflow;

3. Users pay a “steering tax” to extract value from AI — at a price most of us can’t afford to pay.

Meanwhile, the market is delivering tools to us which basically do the opposite:

• Never challenge at all — in a recent audit, 100% of tested L&D AI tools started generating content without asking a single diagnostic question…;

• Bring one posture to every stage — generate, generate, generate — and mostly serve only one stage of the workflow (the build);

• Ship the agreeable first pass as the finished product — leaving the steering tax entirely with the user.

In this week’s blog post, I share the full findings — including where practitioners want AI most (it’s not where the tools are), the three documented AI defects that create the steering tax, and the five-move starter kit my cohorts battle-tested to get genuinely useful work out of AI today.

Let’s dive in!

Three key findings emerged from my observations:

The clearest signal across both organisations is that nobody’s most-valued AI moment was a piece of finished work. The moments people rated most were the ones where AI pushed their thinking — sparring on ideas, interrogating a weak brief, pressure-testing a recommendation.

But there’s a crucial refinement, and it came through loud and clear in the sprint feedback: challenge must not become gatekeeping.

The distinction they drew, in their own words, was between a “pair designer” and a process enforcer:

• Pair designer: “I can draft that. Before I do — the evidence supports at least two other root causes. Want to see the case for each?”

• Process enforcer: “I can’t proceed until you’ve validated the root cause.”

Both of these AI products have the same rigour and value prop , but only is likely to get opened and used a second time.

“Thought partner, not doer” isn’t uniformly true. Participants intuitively re-mapped what they wanted from AI stage by stage:

Two details which stand out to me:

1. The most surprising product idea came at the diagnosis stage: participants wanted the problem-definition agent to live outside L&D entirely. Their logic: most weak projects are weak before L&D ever sees them, because requests arrive “pre-solutioned” — "we need a course on X" — with the actual problem undiagnosed. So, they want AI to help them to put the challenge at the front door: an intake tool the business owner uses before submitting a request ("here's my problem — what's the root cause, and what's the theory of change?"), so that what lands on L&D's desk is a diagnosed problem, not a prescribed solution.

2. The biggest unmet need was at the evaluation stage. We know our own weakness here, and wish AI did more to help. What folks asked AI for was robust impact evaluation: defining the pass/fail bar ("if completion is under 60%, we shelve it"), selecting optimal evaluation methods & schedules and having AI hold them to that number when the results come in — however much they'd like to move the goalposts.

Reading log after log, one pattern was unmissable: AI’s best work is reachable, but takes a lot of user time, energy and and skill. Every genuinely impressive moment — a sharp reframing, an honest confidence downgrade, a catch the human had missed — came after the operator pushed it with a sharper second, third, fourth, fifth question. Meanwhile, operators who didn’t push AI got something agreeable, fluent, and shallow.

The cost of that pushing — the multiple questions, the re-runs, the stripping of your own framing — is what I call the AI Steering Tax: what users pay to move AI from its default behaviour to the behaviour they actually need.

Faced with the same tax, L&D folks spend too long trying to make AI do what they want, get frustrated, and either settle for the agreeable first pass or give up. TL;DR - the AI steering tax is a price most L&D folks can’t afford to pay.

Why does the tax exist at all? Because three AI behaviours are the default — and all three are now independently documented:

• Sycophancy. The agent flatters your framing back to you, lowering your scrutiny exactly when it should be raised. A study published in Science this year tested eleven state-of-the-art models: sycophancy was widespread in all of them.

• Confidence inflation. When evidence was thin, agents said “Medium” confidence when an honest read said “Low.” Confidence research suggests this is structural: the training that makes AI models “helpful” rewards decisive answers over honest uncertainty.

• Reflection, not origination. Left unprompted, the agent mirrors your thinking back rather than generating new thinking — which feels like partnership but isn’t.

Now, put the three findings next to what’s actually being sold to L&D.

A 2025 peer-reviewed study of the leading specialist AI tools for instructional design — found that in 100% of tested scenarios, the tools began generating content immediately: no questions about learner context, performance gaps, or whether the problem was even a learning problem. The researchers’ verdict: these tools institutionalise a solution-first bias.

The money tells the same story. The ~$2.2bn generative AI market for L&D is dominated by video lesson generation, scenario simulation, and quiz authoring — production tools, top to bottom. Synthesia’s 2026 report found 84% of L&D teams cite speed as their primary reason for adopting AI.

Diagnosis, problem definition, evaluation — the stages where my participants wanted AI most — barely feature.

TL;DR: The people who do the job are asking for a partner in judgement. The market is selling them a faster production line.

1. Pre-commit before you run. Write your hypothesis — and what would prove you wrong — before opening the tool. Otherwise the fluent first answer becomes your anchor.

2. Withhold context. Feed the rawest input first; add detail only when asked. The more framing you give upfront, the more the output is your own thinking, laundered.

3. Strip your framing. When the AI agrees with you, remove your synthesis and re-run. If the assessment changes, it was mirroring you.

4. Push back on confidence. Two prompts that earned their keep: “What evidence would have to be true to lower this to medium?” and “Pretend you’re a sceptical L&D director reviewing your own recommendation — where is your reasoning weakest?”

5. Notice absence as data. Missing questions, caveats, pushback — findings, not neutral. An AI that raises no concerns hasn’t validated your plan; it’s failed to examine it.

After a month of watching L&D folks work with AI up close, five things stay with me:

1. People want AI that challenges them, not AI that does the work for them — and the challenge must move with them, never gate them. Rigour delivered as gatekeeping gets abandoned.

2. The spec is stage-dependent: argue with me at diagnosis, generate with me at design, discipline me at evaluation — and never block my momentum.

3. The steering tax is the adoption killer. AI’s best work must be extracted; supported professionals pay the tax, everyone else walks.

4. Sycophancy, confidence inflation and reflection-without-origination are product defects, not user problems — documented, structural, and designable-out.

5. The market is inverted: a production-line industry serving a judgement-partner demand. Whoever builds for the demand side first owns the category.

Happy innovating!