Language Is the Lens: How LLMs Quietly Shape What And How We Think

Language is the interface we use to experience the world. We notice what we have words for, we connect ideas that our terminology and grammar allow, and we evaluate reality through the metaphors we habitually reach for, arguments are “won” or “lost,” time is “spent”, bringing us agency over our own reality. This doesn’t mean language determines thought in a hard, inescapable way, but it does mean our words tilt the playing field of attention, memory, and judgment. An insight that was crystallized further during a conversation Srinidhi had with Laura Pirillo from Milan.

To preserve that agency in AI-assisted legal drafting, this piece also includes deliberate friction as a feature, and not a bug. This is included to show the opportunities that arise when the lawyers retain ownership of framing, alternatives, and consequences rather than sliding into model defaults. LLMs and the chatbot interfaces are the most prolific language partners many of us now have. They autocomplete not just sentences, but also categories, frames, and possibilities. In doing so, they subtly modulate the foundation for and the way we reason. The effects are rarely dramatic. They’re cumulative, ambient, and easy to miss precisely because they arrive as helpful suggestions.

The soft power of suggestion

Think about what an LLM does: it predicts plausible continuations based on the input. That means it supplies defaults – and this matters much more than you think in the continuation of the user’s processing of language. Behavioral science shows that in ambiguous situations people gravitate toward the option presented as the baseline. A suggestion from a chatbot often feels neutral, but it quietly sets an anchor. The first phrasing proposed becomes the frame within which you evaluate alternatives and the model’s tone, empathy level, and vocabulary become the template you’ll edit rather than invent. You’re still in charge, but your starting point is no longer empty, it’s already colored. This “starting-point effect” scales from sentences to strategies.

Legal work is the art of selecting, trimming, and situating language in context to persuade a court, close a deal, or de-risk a dispute. That’s why large language models (LLMs) and chatbots deserve special scrutiny in the field. They don’t just speed up drafting, they propose the frames within which interpretation happens. And for junior professionals, there’s a specific hazard: over-reliance that erodes the hard-won skill of judgment that comes from wrestling with words.

Counter-anchoring as productive friction.

To counter the consequences of the power implied by the default suggestions we can add a productine intellectual friction. Before adopting the model’s first frame, generate one human-first counter-frame and note two trade-offs between them. Then decide. This micro-friction keeps authorship with the lawyer and can reduce the quiet pull of the starting point created by the LLM.

The soft drift toward the median, multiplied by vendor concentration

LLMs produce plausible continuations; in practice, that often means a competent, median register of cautious verbs, hedged claims, broadly acceptable phrasing. Useful, yes, but high value legal work might often about deferring from the median. The genious can bein choosing one precise contour of meaning among near-synonyms that diverge sharply in disputes. The risk compounds when most firms rely on the same one or two legal-AI vendors: the models’ defaults become the industry’s defaults. For example, “Commercially reasonable efforts” starts to crowd out “best efforts,” “diligent efforts,” – not because counsel decided that’s right for the matter, but because the common tool in use offered it first.

Again, an alternative approach to have AI assist in diverging from the median and increase the probability of better lawyering is to involve an ownership check as friction against median drift. This would mean that you for any clause the model proposes, record who (you or the model) originated the risk allocation, who integrated it with related definitions, and how alternatives were evaluated. One margin line suffices. This makes acceptance a reasoned choice of alternatives, and not merely a passive default.

When a widely used system smooths a complex indemnity into something generically elegant, the same carve-outs and survival quirks may get sanded away across firms using the same technology in the same way. A probable scenario is the that over time, a vendor-driven style guide can masquerade as market practice, juniors learn “model-speak” instead of firm lawyer-skilled judgment, and negotiations between opposing counsel using identical engines collapse toward a narrow middle. The result is a quiet homogenization of language and with it, a normalization of compromises that no partner consciously chose.

A lot of early-career growth comes from the friction of interpretation. If a chatbot supplies an instant answer, juniors might glide over those micro-decisions. The result isn’t incompetence; it’s a subtle flattening of their internal error detectors, the very instincts that let a lawyer interrogate language rather than merely consume it.

Whose language is it, anyway?

From an international perspective, there’s also a distributional question: which dialects, registers, and worldviews dominate the corpus? If most text comes from particular institutions, regions, or professions, then the model’s “neutral” voice is not really neutral; it’s an echo of those linguistic centers of gravity. This can subtly re-import their assumptions into your work, especially if you operate in a different context.

A practical method that will increase the probability of the context correctly shaping the LLM’s output, can be done by adding so called Context Clarity Audit (micro friction before auditing). A few concrete examples are:

• Venue and decision-maker: name the forum and archetype (e.g., regulator, VC-backed GC, Distric Court).
• Register and dialect: specify the target voice (local practice norms, brevity preference, idiom constraints).
• Add a context risk posture to diverge from the median: aggressive vs. conservative, and why.
• Instruction to AI: either mirror the chosen context or intentionally contrast it. State which and why. This can lead to locking of the context before LLM-language gets to homogenizing it.

What gets lost when AI drafts first

1. Hypothesis generation. Good lawyers don’t just answer questions; they pose better ones. When a model supplies a polished answer, humans tend to evaluate within that frame instead of asking whether the frame is wrong. That narrows issue spotting. Fricition adding possibility: After reading any AI answer, write one question the answer presupposes and one it excludes; then you can test a reframed prompt.
2. Edge awareness. Legal value often accumulates at edges – where doctrines meet, where facts don’t quite fit the template. Templates (and template-like model outputs) are least sensitive right there.Fricition adding possibility: From the first output, you can run a short “controlled detachment” sprint where you outline edge arguments without AI and then then compare to the AI-assisted outline and mark where divergence of nuances.
3. Idiolect and persuasion. The advocate’s voice – the metaphors they choose, the cadence they build – is a tool for moving decision-makers. Over-standardized prose can read competent but forgettable, reducing persuasive lift.Fricition adding possibility: By grounding in alternatives, you can preserve idiolect and better control persby keeping a short “signature moves” list ( 3–5 metaphors or turns of phrases) and then deliberately re-introduce them after you perceive the LLM- smoothing.
4. Precedent literacy. Wrestling with authorities to reconcile tensions is how juniors internalize hierarchy and analogical reasoning. If a model paraphrases cases too smoothly, the messy, educative parts are hidden.Fricition adding possibility: By requiring  a for of trace-back note for each key case with mimicking the relevant authorities you can create the friction to induce learning: Reverse engineer with including the controlling facts, a more doctrinal hinge, and the analogy that links it to your facts in the context?

Practical guardrails that keep thinking sharp through frictional approach

1) Human-first, model-assisted drafting. Make it policy that juniors write the first analytical pass before consulting AI. That could be an issue list, a strengths/risks matrix, or a bullet-point argument flow that is created with legal reasoning. Only then ask the model for alternatives. This preserves an original frame to compare against the model’s suggestions which will create intellectual friction that stimulate thinking.

2) Prompt for plurality, not singularity. Ban “Give me the clause.” Prefer “List three clause formulations that differently allocate risk, and explain trade-offs.” Force the output to surface choices, not just a median and preferably create a barrier of explanation from the lawyer to release the choice By adding that legal reasoning is induced.

3) Redline, don’t replace. Require associates to mark up the AI draft visibly, with margin notes: why this verb, which case supports this reading, what we’re conceding by using this definition. Partners can scan those annotations to assess actual reasoning, not just surface polish.

4) Deliberate friction blocks. In training matters, impose short “AI-off” sprints – e.g., 45 minutes to outline arguments from record cites only. Then compare against an AI-aided outline. The delta is instructive: where did the machine compress nuance? Where did the human miss a path?

5) Model pluralism for high-stakes questions. Route the same prompt through two models (or two prompting frames: efficiency vs. enforceability) and require juniors to explain divergences. Disagreement is a diagnostic that wakes up human judgment.

6) Ownership justification mapping. For any accepted AI suggestion, record in one line what was accepted, what was rejected, and why. By doing this the lawyer is forced to think through before acceptance and make implicit adoption decisions visible to others in the workflow.

Friction as a feature in AI tools: using AI to surface ownership

Deliberate micro-friction through the likes of  counter-anchoring, controlled detachment, justification mapping are imposed to be slowdowns for their own sake and hurdles in the quest of the use of technology to reduce frictions and make professional life easier. They are created to secure authorship and agency in lawyering. They show and secure that the lawyer framed the problem, surfaced alternatives, and chose words with intent, and not  merely curated, dismissed or accepted model output. Used sparingly and at in the right tasks and workflows, friction increases the intellectual processing in lawyering without materially hurting throughput.

A culture that rewards thinking, not throughput

The efficiency gains of AI are real and in client service, speed matters. But speed is table stakes and judgment is the differentiator. Promotion criteria and feedback rituals signal what juniors practice. If the only celebrated metric is volume, over-reliance on AI is a rational response.

Bottom line

Law’s bread and butter is language and interpretation. LLMs are astonishing accelerants, but acceleration without attention can flatten the very capacities that make a lawyer valuable – framing, distinguishing, and choosing words that move outcomes. The goal isn’t to keep juniors away from AI but it’s to ensure they meet it with a formed point of view, a repertoire of alternatives, and the habit of asking what a different word would do in the hands of a judge, a regulator, or a tough counterparty. Used that way, AI becomes a sparring partner that sharpens judgment rather than a crutch that replaces it.

As a quick diagnostic tip: Ask who did the “thinking” – your team or the tool? If it wasn’t you, add enough friction until it is.