Critical Thinking
The full learning plan. Work through it sequentially or use the navigation to jump to what you need.
Skill Snapshot
Without Critical Thinking, well-presented conclusions get treated as verified ones. A confident report, a polished AI analysis, or an argument that sounds airtight all get acted on before anyone checks what they're standing on. The cost surfaces later — a recommendation built on a stale assumption, a number nobody traced back to its source, weeks spent fixing the wrong thing.
- A confident, well-structured recommendation arrives and the instinct is to act on it
- A conclusion conveniently confirms what you already expected to find
- A number gets repeated often enough that nobody traces it back to its source
- A colleague's argument sounds airtight, but something about it doesn't sit right
- AI produces a polished, confident analysis — and the question becomes whether it has earned that confidence, not whether it sounds right
The ability to take any conclusion — your own, a colleague's, or an AI's — and produce a clear judgment of what's sound, what's questionable, and what's missing, before it gets acted on.
Overview
Critical Thinking is the discipline of testing whether a conclusion holds — checking its evidence, its reasoning, and its bias — before deciding what to trust, what to question, and what's still missing. Unlike a skill that produces something new, this one takes something already produced — an analysis, an argument, a recommendation, a report an AI generated in seconds — and tests it before anyone builds on it.
The analogy that isolates it best: think of a fact-checker reviewing a finished article. They don't write the piece, and they don't rewrite it. They take it as given and test whether each claim holds up — tracing assertions back to sources, checking whether the evidence actually supports what's being claimed, and flagging what's solid from what isn't. The article might be entirely accurate, or it might have one unsupported claim buried in an otherwise solid piece. Either way, the fact-checker's job is the same: test, don't assume.
Where the boundaries are
Analytical Thinking produces the picture this skill evaluates. It decomposes a problem into its component parts and traces symptoms to causes — the output is a structured analysis, whether produced by a person or an AI tool. Critical Thinking doesn't redo that work. It tests whether the picture that's been produced actually holds up: are the parts correctly identified, is the cause-tracing supported by evidence, does the structure hide anything important. The two skills are sequential — Analytical Thinking first, then Critical Thinking — not interchangeable.
Creative Thinking generates what comes next. Once a judgment exists — this part is sound, this part is questionable, this is missing — deciding what to do about it, including generating alternatives or proposing a fix, is Creative Thinking's job. The boundary matters because the pull toward "here's what we should do instead" is strong the moment a flaw is found, and converging on a fix before the evaluation is finished is one of the most common ways an evaluation gets cut short.
Decision-making sits a step further still. A completed judgment — even paired with a generated set of alternatives — still requires choosing a course of action, weighing trade-offs, and committing. That choice is a separate act from either evaluating what's in front of you or generating what could replace it.
Critical Thinking is the discipline of testing whether a conclusion holds — checking its evidence, its reasoning, and its bias — before deciding what to trust, what to question, and what's still missing.
Out of scope
- Generating alternatives, solutions, or recommendations from a judgment — that's Creative Thinking
- Producing the initial analysis, decomposition, or report being evaluated — that's Analytical Thinking
- Formal logic, epistemology, or academic argumentation theory — the focus stays on applied workplace judgment
- Choosing a course of action based on the judgment — that follows this skill but is a separate act of decision-making
Learning Objectives
By the end of this learning plan, you will be able to:
- 01Separate a claim from the evidence that actually supports it, regardless of how confidently the claim is stated.
- 02Apply a source/method/currency/motive filter to judge whether a piece of evidence is strong enough to build on.
- 03Identify where bias may be operating — in a source, in the data, or in your own reaction to a conclusion.
- 04Trace a conclusion back through the Ladder of Inference to find the rung where an assumption took over from observation.
- 05Construct the strongest reasonable version of a claim — including an AI-generated one — before evaluating it.
- 06Produce a judgment of sound, questionable, or missing without converging prematurely on a fix.
Principles of Evaluation
Five axes for testing whether a conclusion holds. Unlike a build sequence, these aren't strictly ordered — depending on what's in front of you, you might start anywhere. Together they cover the ground a judgment needs to stand on: what's claimed versus what's shown, whether the evidence underneath is strong enough, where bias might be operating, where an unexamined assumption took over, and what an evaluation does — and doesn't — produce.
Claims versus evidence
Every conclusion arrives wrapped in language that makes it sound true: confident framing, specific numbers, an authoritative tone. None of that is evidence. The first move in evaluating any conclusion is separating what's being claimed from what's actually been demonstrated — which sentences are assertions, and which are backed by something you could independently check.
This is easy to skip because a well-formed claim and a well-evidenced claim often look identical on the page. "Support tickets increased 40% due to a UI change" reads the same whether that's a measured fact or a plausible-sounding inference dressed in measurement language. Until you've located what a claim is standing on, treat it as a hypothesis — something to test, not something to act on.
The quality of the evidence
Not all evidence carries the same weight, and treating it as if it does is one of the most common ways a sound-seeming conclusion turns out to be wrong. Evidence varies along at least four dimensions worth checking deliberately, every time:
Where did this come from — a primary source, a report about a report, or an unsourced assertion?
How do they know — measured, observed, inferred, or just stated?
Does it still hold — or has the context changed since this was true?
Who benefits if this is true — and does that shape how it's being framed?
Running this filter turns "is there evidence" into "is this evidence good enough to build on." A claim can pass on three of four and still be unsafe to act on — currency in particular is easy to skip, because evidence that was true rarely announces when it stops being true.
AI-generated analysis often presents figures without showing where they came from — source and method are frequently invisible by default. The question is the same one you'd ask of any report: does this figure trace back to something checkable, or did it just appear, stated with the same confidence either way?
Bias has three addresses
Bias is in the source — a report written by someone with a stake in a particular outcome. It's in the data — a dataset filtered, categorized, or collected in a way that shapes what it can show before any analysis begins. And it's in you — bringing prior beliefs that make some conclusions feel more obviously true than others, regardless of what the evidence actually supports.
The third is hardest, because it's invisible from the inside. A conclusion that confirms what you already suspected gets less scrutiny than one that surprises you — not from carelessness, but because confirming evidence takes less effort to accept than disconfirming evidence takes to process.
One check against all three: before critiquing a claim, restate it in its strongest possible form — a practice known as steelmanning, after Anatol Rapoport's rules for constructive criticism. If your objection only works against a weaker version of the argument than the one actually made, you haven't found a problem with the argument. You've found a problem with how you first heard it.
AI tools are unusually good at producing conclusions that sound reasonable whether or not they confirm what you expected — which leaves your own reaction as the main filter still standing. A surprising AI conclusion tends to get checked; a convenient one often doesn't, even though both came from the same process.
The leap from data to conclusion
Between "here's what happened" and "here's what it means" sits a leap, and that leap is where assumptions get smuggled in without anyone deciding to put them there. A 40% increase in tickets is data. "A UI change caused confusion" is an interpretation of that data, resting on an assumption — that UI changes typically cause this kind of spike — which may or may not apply here.
The Ladder of Inference, covered as this plan's mental model, makes this leap visible: each rung from raw observation to action involves a choice that feels automatic but isn't. Most flawed conclusions aren't wrong because the underlying data was wrong. They're wrong because a leap partway up the ladder went unexamined — and the higher up the ladder a conclusion sits, the more confidently it tends to be stated, which is exactly backwards from how much scrutiny it's received.
What this produces — and what it doesn't
The output of this entire process is a judgment: this part is sound, this part is questionable, this is what's missing. That's it. It is not a fix, a redesign, or a list of alternatives — reaching for those is Creative Thinking, and reaching for them too early is one of the most common ways an evaluation gets cut short.
The pull toward "here's what we should do instead" is strong, especially once you've found a real problem with a conclusion. Finding a flaw creates momentum toward resolving it, and stopping at "this doesn't hold" can feel incomplete in a way that's uncomfortable to sit with. Resisting that pull means treating an honest, incomplete picture — knowing something doesn't hold up without yet knowing what replaces it — as a legitimate, useful place to stop.
AI tools rarely stop at a judgment — ask one to evaluate a plan and it will typically follow with suggested fixes, whether or not you asked for them. That's not wrong, but it collapses two separate steps into one, and the second step can arrive before the first has actually been tested. Take the evaluation; set the suggested fix aside until the judgment itself is finished.
Bias Field Guide
Eight biases that shape what gets noticed, accepted, or waved through — in a source, in data, and in your own reaction to a conclusion. Click a card for the fuller picture: what it looks like in knowledge work, how to spot it, and how to counter it.
Select a bias above for a fuller breakdown — definition, how it shows up in knowledge work, and how to spot and counter it.
Mental Model
The Ladder of Inference
The Ladder of Inference describes seven rungs most people climb in milliseconds, without noticing: starting from raw data and facts, we select certain data to pay attention to (shaped by what we already expect to see), add meaning to it based on past experience, form assumptions, draw conclusions, develop beliefs, and take action — each rung built on the one below it, each one a choice that felt automatic.
Argyris's caution is that the lower a rung sits, the more self-evident it feels — and the less likely anyone questions it. A fact, however obvious it seems, isn't really substantiated until it's verified independently. The Ladder's value isn't memorising seven labels; it's the habit of climbing back down a conclusion before acting on it, rung by rung, until you find the one that was actually an assumption wearing a fact's clothes.
The leap: a pattern from general experience ("UI changes cause spikes") applied to this specific, unverified situation.
Embedded into a team's working practice — not just an individual's — the Ladder becomes something to show, not just to use privately. Walking a colleague through which rung a conclusion actually sits on, and inviting them to check it, turns a private check into a shared one. It's also a non-confrontational way to question someone's reasoning: "which rung is this on?" lands very differently from "I don't think you're right."
Common Mistakes
Mistaking confidence for credibility
Fluent, assertive language reduces the effort needed to accept a claim — something well-organised and stated with certainty feels more accurate, independent of what's underneath. AI-generated text does this especially well: it's trained to sound certain regardless of how thin the underlying basis is.
Apply the same source/method/currency/motive check no matter how confident the framing sounds. Strip the tone away and ask what's actually left.
Confirmation creep
A conclusion that matches what you already believed gets less scrutiny by default — not from carelessness, but because confirming evidence is cognitively cheaper to accept than disconfirming evidence is to process.
Treat "this immediately feels right" as a flag rather than a green light. Apply more scrutiny to claims that confirm your priors, not less.
Strawmanning instead of steelmanning
Picking apart a careless or weak version of a claim feels productive — you get to "win" without doing the harder work of engaging with its strongest form.
Restate the claim as strongly and fairly as you can before raising an objection. If the objection doesn't survive that version, you haven't found a problem yet.
Treating "no flaw found" as "verified true"
Evaluation feels complete once you've checked what you know to check — but unknown unknowns don't announce themselves.
Keep "I checked X, Y, Z and they hold" and "I haven't found a problem yet" as two different statements. The second is an incomplete judgment, and should be labeled that way rather than rounded up to confirmation.
Behavioral Indicators
Observable behaviors across a single spectrum: the target zone in the middle, flanked by the signals that show when the skill is under- or overused.
- Separates what a claim asserts from what it demonstrates before reacting to it
- Applies the source/method/currency/motive check regardless of how confidently something is presented
- Notices when a conclusion confirms a prior belief and gives it more scrutiny, not less
- Restates a claim in its strongest form before finding fault with it
- Traces a surprising or high-stakes conclusion back through its reasoning to find where an assumption took over
- States a judgment — sound, questionable, missing — without immediately reaching for a fix
- Treats AI-generated analysis as a hypothesis needing the same checks as any other source
- Keeps "no problem found" and "verified" as separate statements
- Asks "what would change my mind about this?" before accepting a conclusion
- Can say specifically which parts of an analysis they trust, which they don't, and why
- Accepts conclusions because they're confidently stated or come from an authoritative-seeming source
- Acts on AI-generated analysis without checking its basis
- Stops at "this feels right"
- Picks apart minor phrasing rather than substance
- Treats "nobody objected" as evidence of correctness
- Applies scrutiny unevenly — harder on claims they disagree with, lighter on ones they expected
- Repeats a conclusion that's gone stale without checking whether the situation changed
- Accepts a remaining cause "by elimination" once one alternative is ruled out, without separately checking it
- Treats every claim as equally suspect regardless of stakes, slowing decisions that don't warrant it
- Evaluation becomes its own end — endless questioning with no judgment ever stated
- Steelmanning becomes a way to avoid ever disagreeing
- Uses "not fully verified" to stall indefinitely, including on low-cost-of-being-wrong decisions
Practical Examples
An AI ticket-volume analysis
An AI tool reports ticket volume up 40%, points to a recent UI change, and recommends reverting it plus adding headcount for slower resolution times. Leadership approves both immediately.
Three months later, volume hasn't moved. The 40% figure was partly a tagging artefact, a pricing change that shipped the same week was never considered, and the headcount addition missed the real bottleneck — triage.
Each claim gets traced to its source before action. The volume figure is checked against the tagging change. The UI-change theory is steelmanned against "what else changed that week?" — surfacing the pricing change.
"Resolution time" gets checked against what it actually measures, and the real bottleneck — triage — becomes visible. The report was a hypothesis worth testing, not a finding worth acting on.
Removing a "redundant" QA step
A colleague says "we've always had problems with this step — it just slows everything down." The framing goes unchallenged, and the step gets cut from the release process.
Someone checks when "always" was. The delays were real eighteen months ago, before a tooling fix resolved the underlying issue.
The colleague's experience is genuine but stale — the step hasn't been the bottleneck since. Cutting it would remove a now-harmless check while leaving the actual current bottleneck untouched.
Self-Reflection Activities
Three prompts to audit your current use of Critical Thinking. Each takes under five minutes to complete honestly.
Knowledge Check
Five questions — two conceptual, two applied, one synthesis. Select your answers, then reveal results.
Confidence is a feature of presentation; support is a feature of evidence. They vary independently — a poorly-supported claim can be stated with total confidence (this is common in AI-generated outputs, which are trained to produce fluent, assertive text regardless of underlying certainty), and a well-supported claim can be stated tentatively. Treating confidence as a proxy for support is one of the most common ways flawed conclusions get accepted.
Conclusions that confirm what you already believe get less scrutiny by default, not from carelessness, but because confirming evidence is cognitively cheaper to accept than disconfirming evidence is to process. Checking your own reaction is how you catch the bias that's hardest to see — the one operating in you.
Before evaluating a proposed cause, check that the figure itself is measuring something consistent — definitional drift in a metric like "churn" can produce a number that looks like a real change but isn't. Then check what else changed at the same time, since a single proposed cause competing with an unconsidered confound is exactly the gap that produces a confident but wrong conclusion.
Shared experience in the same meetings isn't independent verification — it's the same observation seen twice. The bias check here is on yourself: your agreement feels like confirmation because it is one, drawn from the same limited vantage point as the original claim. Checking outside that vantage point, and checking whether the claim is current rather than a stale-but-true story, is what the evaluation actually requires.
Disproving one cause doesn't validate the others — "not yet examined" and "confirmed by elimination" are different states being treated as the same thing. The honest judgment names exactly what's been established: one cause unsupported, two still untested. That's Principle 05 in practice — a judgment, stated precisely, without rounding up to either "all good" or "all bad."
5-Day Habit Builder
Five daily practices, each under 15 minutes. Days build on each other — each step uses output from the previous day. Open each day to see the full practice.
"This tool reduces manual processing time by 30% in three months, for teams similar to ours."
List only what's there: one case study, one customer, no methodology shown, no date.
Evidence provided: one named customer, one headline percentage, a short quote. Nothing about how "processing time" was measured, what the baseline was, or when this happened.
Source: where does this come from? Method: how was it measured? Currency: when did this happen? Motive: who benefits from this number persuading you?
Method: unclear what "manual processing time" includes or how it was baselined.
Currency: undated — could be from an earlier product version.
Motive: the vendor benefits directly from this number persuading you.
A recommendation, a previous good experience, fatigue from evaluating alternatives — name anything that might be giving this vendor a head start.
If a vendor you didn't favor showed you this exact case study, would you accept it as readily?
Steelman test: a vendor with no recommendation behind them, showing the exact same case study, would prompt more questions — which means those questions are legitimate now too.
Write the action or belief you're closest to — even provisionally.
Belief → conclusion → assumption → meaning added → selected data → data. Identify which rung holds the unexamined leap.
Belief: it'll meaningfully reduce our team's workload.
Conclusion: it'll deliver similar results for us.
Assumption: "similar to yours" means comparable enough to transfer.
Added meaning: a 30% result from a "similar" team is a strong signal.
Selected data: the 30% headline, out of whatever else the case study contained.
Data: the case study text itself.
The leap sits between "selected data" and "added meaning" — treating one headline number from one unverified-similar customer as a strong signal.
What's sound, what's questionable, what's missing — in plain terms.
Read it back. Does it stop at the judgment, or does "missing" quietly turn into "so we should..."? If so, separate the two.
Questionable: source-biased (vendor-selected and -published), method unclear (what's measured, what's the baseline), "similar to yours" unverified.
Missing: an independent reference customer, the underlying methodology, and the date of the case study.
This isn't evidence the tool won't work — it's evidence we don't yet know whether it will. That's a separate question, requiring separate information-gathering.
Five days complete. The Foundation stage begins here — the habit builder has given you the starting practice. Return to it any time the skill feels rusty.
Progression Path
Three stages of developing mastery in Critical Thinking. Each stage has a new capability and an observable signal of progress.
The Careful Skeptic
You apply the checks deliberately and consciously — source, method, currency, motive; steelmanning before objecting; tracing a leap down the Ladder. The process takes effort and reminders. You're still occasionally swept up by confident presentation, but you catch it more often than before.
You can take a confidently-stated conclusion — AI-generated or otherwise — and produce a written judgment of what holds and what doesn't, rather than accepting or rejecting it wholesale.
You catch at least one flawed claim — a number traced to a bad source, a recommendation based on stale data — before it gets acted on.
The Bias Spotter
The checks become habitual — you run them before deciding to. Your focus shifts to noticing patterns in your own reactions, which kinds of claims get a free pass, and you start noticing when other people's evaluations stop short: a "no flaw found" treated as "verified," a cause accepted by elimination, a conclusion that went unchecked because it confirmed something convenient.
You can evaluate someone else's evaluation — noticing when a colleague's "I checked it, it's fine" only checked what was easy to check.
Colleagues run things past you before acting on them — not because of seniority, but because you reliably catch what others miss.
The Calibration Point
The checks run fast enough not to slow decisions down, and you can tell which situations need heavy scrutiny and which don't — avoiding the overuse end of the spectrum as readily as the underuse end. You help others build the same instinct by making the checks visible and shared, not just applying them privately.
You can facilitate a group evaluation — bringing a team to a shared judgment about whether to trust a finding, without either rubber-stamping it or getting stuck in endless doubt.
The team's default response to a confident AI output or a compelling pitch becomes "what's this based on?" — the question itself has become the norm, not something you have to keep introducing.
Quick-Recall Summary
Critical Thinking is the discipline of testing whether a conclusion holds before deciding what to trust. It means separating what's claimed from what's evidenced, checking that evidence against its source, method, currency, and motive, watching for bias in the source, the data, and yourself, and tracing surprising conclusions back to the point where an assumption took over from observation.
The output is a judgment — sound, questionable, missing — not a fix. What you do with that judgment is a different skill; this one gives you the confidence to know which parts of the picture you can actually stand on.
What to work on next
Critical Thinking tells you whether a conclusion holds. Analytical Thinking gives you the structure to build reliable ones in the first place. If you haven't worked through Analytical Thinking yet, it's the natural companion to this skill.
There's also a curated set of reads, videos, and podcasts below if you want to go deeper into the concepts behind this one first.
Explore Further
Curated for what each resource adds beyond this learning plan — not a description of what it covers, but a reason it belongs here.
| Type | Title | Author / Source | Est. Time | Why This Specifically | Links |
|---|---|---|---|---|---|
| Book | The Fifth Discipline | Peter Senge · Doubleday, 1990 (rev. 2006) | 6–8 hrs | The original source for the Ladder of Inference, with the full organizational context Argyris and Senge built it for — beyond what any single-page summary covers. | Goodreads |
| Article | How to Use the Ladder of Inference for Better Decisions | Asana, 2026 | 15 min | The clearest practical walkthrough of all seven rungs with workplace examples, including how to use the Ladder in team conversations rather than just privately. | Read |
| Paper | AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking | Michael Gerlich · Societies, 2025 (open access) | 30–45 min | The primary research behind this plan's framing of AI and Critical Thinking: a 666-participant study finding a measurable negative correlation between AI tool use and critical thinking, mediated by cognitive offloading. | Read |
| Article | The Paradox of AI Assistance: Better Results, Worse Thinking | EDUCAUSE Review, Dec 2025 | 15 min | An accessible synthesis of Gerlich and related studies, written for a workplace and education audience rather than an academic one. | Read |
| Factsheet | Evidence-Based Practice for Effective Decision-Making | CIPD | 15 min | A practitioner framework for weighing evidence quality from a professional body — useful grounding for the source/method/currency/motive filter in Principle 02. | Read |
| Book | Thinking, Fast and Slow | Daniel Kahneman, 2011 | ~20 hrs | The foundational text on System 1 / System 2 thinking and confirmation bias — goes deeper on "bias has three addresses" than any shorter source can. | Goodreads |
| Book | Intuition Pumps and Other Tools for Thinking | Daniel Dennett, 2013 | ~10 hrs | Dennett's formulation of Rapoport's Rules — the structured version of steelmanning referenced in Principle 03. | Goodreads |