Legal work is unusually exposed to generative AI. Tools like CoCounsel and Harvey have moved from pilot to daily use in many firms; document review, contract redlining, and first-pass legal research now run through some kind of model. The cautionary cases are well known — Mata v Avianca in 2023 famously involved a lawyer who filed a brief citing AI-fabricated cases — but the steady-state question for working lawyers and paralegals is narrower: what cognitive acts do you keep doing yourself, every working day, so that the AI integration doesn't hollow out the judgment that defines the job?
This post lays out a defensible answer, with citations and a concrete routine.
Why this matters — the published evidence
The 2024 MDPI Societies study on AI and cognitive offloading (MDPI / Societies) found an inverse correlation between AI usage frequency and self-reported critical-thinking engagement on knowledge tasks. The effect was strongest among younger users with heavier daily exposure. The study does not claim AI use causes critical-thinking decline — survey correlations cannot — but the direction is consistent with the broader cognitive-offloading literature (Risko & Gilbert, 2016, Trends in Cognitive Sciences).
The 2025 EDUCAUSE Review piece on the "productivity paradox of AI" names the pattern directly for knowledge work: AI tools produce better artefacts and weaker thinking habits in parallel if you don't design around it. The pattern shows up in legal work because legal work is, at its core, a thinking job whose artefacts are language.
The 2026 Anthropic engineering-skills piece (Anthropic blog; InfoQ coverage) is about software engineering rather than law, but the structural argument transfers. The piece argues that engineers who outsource the originating cognitive acts to AI develop "skill atrophy" on those acts. Reading-heavy, judgment-heavy disciplines like law have the same exposure.
The close-reading muscle
The cognitive act most exposed to legal-tech AI is close reading. CoCounsel will summarise a case. Harvey will produce a research memo. Both are useful. Neither performs the act that defines the lawyer — sitting with the actual text, in primary sources, and noticing what is and is not there.
Close reading is the load-bearing variable in legal judgment. The signal that a case is distinguishable, that a clause is ambiguous, that a brief is overreaching its own holding — all of it lives in the close reading. AI summaries flatten the texture that close reading depends on.
The Mata v Avianca 2023 case is the public-facing reminder. The lawyer's mistake was not using AI; it was failing to do the close reading that would have caught the fabricated citations. The cautionary tale, properly read, is not about AI hallucination. It is about what happens when the close-reading muscle stops being exercised.
A daily routine for working legal professionals
The routine below is built around a working day. None of it requires you to abandon legal-tech AI. All of it requires fifteen to twenty minutes a day.
1. Read one primary source unmediated, every day. A case, a statute, a clause from a contract you are working on. Read it without an AI summary alongside. The volume does not need to be large; the act has to be unmediated.
2. Make the first synthesis yourself. Before asking CoCounsel or Harvey for a summary, write your own one-paragraph version. Then compare. The cognitive act of producing the synthesis is what keeps the muscle warm; the AI version is the check, not the originator.
3. Verify every citation from the primary source. Not a sample. Every one. Treat AI citation lists as a first-pass index and the primary source as the truth. Mata v Avianca is what happens when this step gets skipped at scale.
4. Reason out loud before redrafting. When you spot a problem in a draft — yours or AI's — say what is wrong before fixing it. The articulation is the rep. Silent fixing trains the editor, not the reasoner.
5. Keep one purely unmediated brief or memo a week. Even a short one. Even an internal one. The point is to preserve the daily existence of the originating cognitive act, which is the thing the AI integration is most likely to quietly take over.
What this is not
The category is loud and the honest position is narrower than the headlines.
This routine does not predict that AI will deskill the legal profession. It does not claim that any lawyer using CoCounsel or Harvey is at risk. It does not say AI use causes any clinical condition. None of those claims is supported by the published evidence and Senwitt does not make them.
What the evidence does support is narrow: cognitive acts that get fewer reps tend to weaken, and AI integration in legal work shifts which acts get reps. The routine above keeps the load-bearing acts on the calendar. That is all it claims to do.
It is also not a substitute for firm-level practice management. Whether your firm permits AI on which kinds of matters, how cite-checking is audited, and how junior associates are trained when first-pass drafting is AI-mediated are all real questions that an individual practice routine cannot solve. The routine is what an individual lawyer or paralegal can do regardless of how those questions get answered.
How Senwitt fits
Senwitt does not replace the work above. What it offers is a daily, on-purpose, seven-minute Set across six Senwitt Skills — including Reading, Reasoning, and Writing — that keeps those muscles warm even on days the rest of your work runs through AI. The for lawyers page lays out the case in more detail. The deliberate-practice frame Senwitt is built on comes from Ericsson, Krampe & Tesch-Römer, 1993 — daily, effortful, on-purpose engagement with the specific skill.
The longer argument on what the cognitive-offloading research actually supports lives on the research/cognitive-offloading page. The research/ai-overreliance page covers the workplace pattern.
A specific note on the paralegal and junior-associate case, because the cognitive-offloading risk lands differently early in a legal career. In the pre-AI model, the close-reading muscle was effectively trained on the job — the first three years of legal work involved unreasonable volumes of case review, statute parsing, and cite-checking that produced the close-reading reflex by repetition. In an AI-mediated model, much of that volume is performed by tools, and the practitioner who used to develop the muscle by doing the work can now finish the work without developing the muscle. The risk is not that the artefacts are worse; it is that the reflex never gets installed. The daily routine above matters disproportionately for early-career practitioners precisely because the structural reps that used to come from the job itself are no longer reliably there. The honest framing is that the daily practice is not a complement to the work for someone five years in; for someone in the first three years it is the part of training that AI integration has quietly displaced, and it has to be put back in by hand if the close-reading reflex is going to exist when the harder cases arrive.
