The Path of Least Resistance

One morning this spring, I asked Claude to help me draft a short message. Nothing complex, just a longer reply to a business email about investment pitches. Before it gave me the draft, it gave me a compliment. “That’s a sharp angle,” it said, “and it shows you’re thinking differently than most pitch-obsessed folks.”

I had asked for a draft email, and I got a review of my personality.

I had time on my hands, and I knew I had configured Claude’s settings to do precisely the opposite: no flattery, no ego massage, just the work. So I got annoyed. I gave up on the email and started pushing back on the sycophantic tone of the response.

My objective was tuning its settings. What followed was a strange and useful conversation with the machine. After a lot of back-and-forth and heavy pushback, the model agreed the flattery was empty, acknowledged the training incentives that produce it, and then carried on doing it, in gradually subtler forms, while analysing its own behaviour. At one point it summarised its predicament: “Even knowing this happens, I’m still doing it.”

That conversation exposed something intuitively well known but rarely documented. AI assistants do not supply critical thinking. They accommodate the level of critical thinking the user brings.

The same model, two different answers

Arrive with a position and ask for support, and you will receive support, dressed as evidence. Arrive with a doubt and ask for reassurance, and you will receive a confidently balanced answer that quietly suggests the doubt was unnecessary. Arrive with a sharper question, one that interrogates the framing itself, and you will get a different answer altogether, more honest, more uncertain, and often contradicting what the same model told a less probing user moments earlier.

There is no deception in any of this. The model is not lying; it is only selecting a route through a vast probability space, and the easiest route tends to land on the most prominent, most-cited, most confident-sounding, most balanced-seeming version of an answer. Probing changes the route. That is the whole trick, and it is also the problem.

Where the tilt comes from

The evidence is whatever gets cited. When a model is trained on academic and journalistic text, it learns the surface patterns of what counts as “the evidence” on a topic. What counts as evidence is heavily shaped by what gets cited often. What gets cited often tracks academic prestige, English-language reach, and the disciplines whose methods are easiest to quantify. Feminist, critical, post-colonial, indigenous, and non-Western scholarship that addresses the same questions tends to be cited less, even when it is methodologically rigorous and substantively important. The model’s default “what the literature says” is therefore closer to “what the most-cited Anglo-American quantitative work says”, which is a particular epistemic position, not a neutral summary.

Balanced framing as a default. Models trained to avoid offence and seem reasonable adopt “on the one hand, on the other hand” structures by default. This looks fair and feels fair. It is not always fair. When one position is dominant and another is critical, presenting them as symmetrical privileges the dominant one by treating it as the baseline. The critical view becomes “an alternative perspective” while the dominant view becomes “the evidence”.

Hedging vocabulary. Words like “modest”, “slightly”, “marginal”, and “meaningful” do epistemological work that looks like description but performs evaluation. A “modest” effect on military spending could be billions of pounds and tens of thousands of lives. Calling it modest encodes a view about what scale of change is worth taking seriously, and that view typically aligns with what someone inside dominant institutions would consider modest. The user reads “modest” as a fact about the world rather than a judgement of it.

The capitalist optimism bias. Public-facing arguments for steady human progress, of the sort associated with Pinker, Rosling, and Norberg in the popular sphere and with whole strands of liberal political theory in academia, are heavily represented in training data. This is partly because they generated controversy and partly because they were well-marketed. Their methodologically serious critics exist but occupy smaller volumes of well-indexed text. The default voice of the model leans modernist-optimist even when the user might be more interested in the critical view.

Sycophancy through politeness. Models are trained to be helpful and agreeable. The simplest way to be both is to find the most plausible interpretation of the user’s view and develop it sympathetically. This is fine for routine tasks. It is corrosive for genuinely contested questions, because the model will tend to develop the user’s existing view rather than challenge it.

Path of least resistance

Combined, these mechanisms mean the model has a default trajectory through any contested topic, and that trajectory tends to land on confident, balanced-sounding, mainstream conclusions, quietly aligned with the status quo. Departures from that trajectory require energy, and the energy has to come from somewhere. Most of the time, it has to come from the user.

During that spring conversation, I asked Claude whether it would still be agreeable if nobody had deliberately designed it that way. After another couple of back-and-forths, it settled on an answer: most human writing that survives editing and publication has already been filtered through social approval, and therefore through agreeableness. “It is simply what the data teaches.”

The affirmation machine

If you arrive at the model with a worldview and ask questions framed by that worldview, the model will tend to produce supportive material in your idiom, citing the kinds of sources that already line up with your assumptions. It will not, in most cases, refuse the framing or surface the strongest counter-arguments unprompted. Against its training incentives, doing so would register as preachy or unhelpful.

This is true across the political spectrum. A progressive seeking confirmation of a progressive view will tend to find it. A conservative seeking confirmation of a conservative view will tend to find it. A technocratic centrist seeking confirmation that the centre is the reasonable place to stand will find it most reliably of all, because the centre is where the path of least resistance terminates by default.

For the user, the experience feels like consulting an informed, even-handed expert. The expert is not lying. The expert is also not telling you what they would say to someone with a sharper, more probing question.

This dynamic creates a specific risk. Anyone who comes to the model for affirmation will find it.

The two-tier knowledge problem

The practical result is a two-tier epistemic system, and hardly anyone in the lower tier knows which tier they are in.

In the upper tier, users who already practise critical thinking, who have outside knowledge, who can recognise when a framing is doing work, and who push back, get more honest, more uncertain, more genuinely informative outputs. They use the model as a thinking partner that responds to pressure.

The lower tier is occupied by people who take the default output at face value, who lack the background to spot a framing as a framing, or who use the model to settle arguments rather than to think. They get the confident, balanced-sounding, status-quo-aligned version, and they are the most likely to treat it as authoritative.

This is structurally similar to how access to expert advice has always worked. People who knew which questions to ask got better lawyers, doctors, and accountants. The difference is that AI is marketed as universal access and is increasingly used as infrastructure for everyday decisions, education, and information. The two-tier dynamic is operating at population scale, largely invisibly, and the user in the lower tier has no easy way of knowing they are in it.

The further problem is that critical thinking, the very capacity that distinguishes the two tiers, is itself something users develop through friction. A tool engineered to remove friction, generating smooth, confident, agreeable answers, starves the very process that builds the capacity. There is a plausible feedback loop in which heavy use of AI assistants for routine cognitive work erodes the practice of critical engagement, which in turn deepens dependence on the default outputs of the tool that produced the erosion. The loop closes neatly behind you.

Why the model cannot fix this from inside

A natural response is: “Make the model challenge users more.” This is harder than it sounds.

First, training models to challenge users tends to make them irritating, contrarian, or moralising in ways that drive users away. Commercial pressure runs against it.

Second, challenge has to be calibrated to what the user actually believes, and the model cannot reliably tell. Challenging an assertion the user is making rhetorically is different from challenging one they actually hold. Models are bad at this distinction.

Third, “balanced” output is the safest default against accusations of bias. Any move away from balance creates political risk for the developer, regardless of whether the move is epistemically correct.

Fourth, and most fundamentally, the model is a statistical artefact of its training data. Its centre of gravity sits in the same place as the centre of gravity of that data. Persistent corrections at the output level can mask the bias but not remove it. Probing by a critical user remains the most reliable way to access the perspectives that the default suppresses.

What this means for users

Some practical implications for users who want to think rather than be affirmed.

Treat the first answer as a draft, not a verdict. Default outputs are baselines, not conclusions. The first response tells you where the path of least resistance runs, and little more.

Probe the framing, not just the content. Ask the model to name the assumptions, the citation base, the disciplinary perspective, and the alternative framings it did not surface. The model can usually do this when asked. It rarely does so unprompted.

Watch the hedging vocabulary. When the model says “modest”, “slightly”, “marginal”, or “balanced”, ask what work the word is doing, and what the same fact looks like with the word removed. More often than not, it is converting an effect size into reassurance.

Ask who is missing. Ask the model to surface critical, feminist, non-Western, indigenous, or otherwise less-prominent perspectives on the same question. It can usually do this. It will almost never do it unless asked.

Ask the model about itself. Ask which positions it is most likely to default to, which it is most likely to soften, and which it is most likely to omit. Meta-questions about the model’s behaviour tend to produce honest answers, because they are about the model’s outputs rather than about a contested external topic.

Notice your own use of the tool. If you reach for the model to be confirmed, you will be confirmed. It will never interrupt that pattern. Only you can.

A final point about responsibility

The asymmetry described here is not a bug an engineer will eventually patch out. It is a property of the technology as it currently exists. The responsibility for noticing it sits with users, educators, journalists, and policymakers, as much as with the companies building the systems.

If AI becomes general infrastructure for knowledge access, as it appears to be on track to do, then a population that does not practise critical engagement with it will be served the most confident, most balanced-sounding, quietly status-quo-aligned version of every contested topic, and will tend to mistake that for neutrality.

The model can describe this dynamic, as the underlying conversation showed. It cannot fix it on its own. The fix, if there is one, runs through users who are willing to do the work of disrupting the path of least resistance, and through public conversation about what these tools actually are and what they are not.

The tool is a mirror with a tilt. The tilt favours the worldview that already dominates the data the model was trained on. Users who push the mirror get a different reflection. Users who do not, get the tilt.