Full stack AI legal scholarship

From inchoate theory to law review draft in a collaborative morning

Academic legal scholarship challenges both law faculty and law students. This post explores the possibility of delegating significant responsibility for the process to AI. That prospect may appall many. But before we despair about the declining purpose of humans, let's see how well AI can really do. If, after all, the work of the AI is below that of a substantial fraction of humans trained in law, perhaps we need not worry too much. If, on the other hand, AI surpasses the abilities of all but the top experts – and at least works faster than they do – complex issues arise about the ability of AI to fill the space of legal research and the remaining purpose of academic legal study.

A note before we begin. Writing about interactions with AI often results in very long posts or posts with a lot of links that all but the most devoted readers may find it bothersome to click on. So, I am going to offer three things here. First, an Executive summary of the post for you to read. Second, a "how to" summarizing how to replicate the workflow described here. And third a brief podcast generated by NotebookLM that discusses the content of the post. But, of course, for the full benefit, read on!

Executive Summary

This article explores the evolving role of artificial intelligence in academic legal scholarship by documenting an experiment: can AI transform a vague notion into a polished, law review-ready article in a single morning? The author begins with a simple, undeveloped idea: that a series of individually constitutional state laws favoring religion can, in the aggregate, amount to a de facto "cumulative establishment" of religion, a phenomenon current legal doctrine is ill-equipped to handle.

To develop this thesis, the author engages in an iterative process with a custom AI tool, "Volokh my claim," designed to refine legal arguments. The AI's initial output provides a structured analysis, identifying potential weaknesses such as doctrinal mushiness, standing issues, and conflicts with recent Supreme Court precedent. The AI even proposes a quantitative "Aggregate Endorsement Score," a concept the author deems absurd but illustrative of the inherent difficulty in addressing the "death-by-a-thousand-cuts" strategy employed by some legislatures.

After several rounds of refinement, where the human author guides the AI away from unworkable quantification and toward more plausible doctrinal solutions, the process yields a sophisticated framework. The AI suggests borrowing the "pattern-or-practice" standard from Title VII, proposing remedies like structural injunctions, and developing a "plural-plaintiff aggregate injury" theory to overcome standing hurdles. The result of this collaborative effort is a comprehensive, 14,000-word draft article, complete with a literature review, empirical examples from 15 states, and a pre-emptive rebuttal of conservative counterarguments.

Ultimately, the author concludes that while AI can now perform 60-70% of the work involved in producing legal scholarship—adeptly handling research, structure, and initial drafting—it still requires significant human guidance to navigate nuance and avoid doctrinal dead ends. The final AI-generated article is impressive, insightful, and better than what most humans could produce in such a short time. However, it still falls short of a submittable final product, exhibiting weaknesses in footnote conventions and a need for deeper engagement with case law. The experiment demonstrates that while the job of a law professor is not obsolete, it is fundamentally changing. The future of legal scholarship may be one where humans act as expert directors, guiding powerful AI tools to explore and articulate complex legal theories, forcing a necessary conversation about the purpose and production of academic legal study in the age of AI.

Workflow

The process detailed in the article can be broken down into the following steps:

1. Conceptualization: Start with an "undeveloped idea," even one as brief as a single sentence. The author challenges the AI with the "equivalent of a grunt" to show that a fully formed thesis is not a prerequisite.
2. Initial Development (Iteration 1): Use a custom AI tool designed to develop legal theories. The author uses a CustomGPT called "Develop and improve legal theories" to transform the initial idea into a proposed claim, complete with a "Ruthless Volokh Evaluation," a "Critical Test Suite" of scenarios, and a "Fatal-Flaw Watchlist".
3. Human-Guided Refinement (Iteration 2 & 3): Engage in an iterative feedback loop with the AI. The author provides specific prompts to the AI, asking it to create a tighter version of its claim that addresses weaknesses and incorporates new ideas like remedies and legislative incentives. When the AI produces an "absurd" response—like a quantitative scoring system for the Establishment Clause—the author provides critical feedback to steer it toward more plausible, non-quantitative alternatives.
4. Thesis Consolidation: Once the theory is sufficiently refined, ask the AI to generate a concise, ~500-word abstract that crystallizes the argument. This creates a working summary of the proposed article.
5. Critical Stress-Testing: Switch to a different, more critical AI tool. The author uses the "Volokh my claim" CustomGPT to subject the abstract to a "brutal (but fair) criticism". This AI is prompted to identify fatal flaws, test the theory with difficult scenarios, and suggest specific deliverables needed to improve the scholarship.
6. Generating Core Components: Based on the critique from the previous step, prompt the AI to produce the specific deliverables it recommended. In the author's case, this included a revised thesis statement and abstract, an empirical table of state laws, a "standing work-around," pre-empted conservative objections, and a literature map.
7. Long-Form Drafting: Shift to an AI optimized for long-form writing, such as Gemini's Deep Research feature. Provide this AI with all the refined materials from the previous steps and prompt it to generate a full-length (e.g., 15,000-18,000 word) law review article, complete with its own research plan and references.
8. Human Evaluation and Final Polish: The final step is for the human expert to evaluate the AI-generated draft. The author notes the draft needs significant work before submission, such as fixing footnote conventions and deepening the case law analysis, reinforcing that the process yields a strong first draft, not a finished product.

By documenting the specific types of AI used, the prompts given, the iterative feedback process, and the final output, the blog post serves as a blueprint for other legal academics and students looking to leverage AI for "full stack development" of their own scholarly work.

The Podcast

6-minute version: https://notebooklm.google.com/notebook/963d7db8-0461-4cf3-a061-fa26902d9317/audio

15-minute version: https://notebooklm.google.com/notebook/b80c5f27-d658-4f5c-abf5-3891c63fdee9/audio

From Undeveloped Idea to Law Review Article

I'm going to focus on a concern of mine. It will be set forth in the "Undeveloped Idea" set forth below.

Undeveloped Idea

Over the past few legislative sessions, Texas and other states have passed a variety of laws that tend to favor religion or may favor a particular religion. Taken by themselves, those laws might not violate either the old Lemon test with a focus on entanglement and endorsement or the newer tests based on history and tradition favored by cases such as Kennedy v. Bremerton. But, if one puts the laws together, the state gets closer to establishing a theocratic system of the sort (I think) that would make the Founders uncomfortable. But how do you challenge a vibe? Which sets of laws get struck down? And how does this relate to facial versus as applied unconstitutionality?

A CustomGPT to Develop the Theory

To help develop this idea, which comes from a human mind, I enlist AI. In particular, I want to use a new CustomGPT called "Develop and improve legal theories." It has a superset of capabilities I developed earlier in another CustomGPT called "Volokh my claim" named in honor of Professor Eugene Volokh whose work on academic legal writing has helped both me and many of my students over the years. (Professor Volokh consents to the use of his name). The earlier CustomGPT, which is still perfectly serviceable, evaluates a legal theory that you propose and helps improve it. But, brutally, it requires that you have at least the draft of a thesis to begin with. The newer CustomGPT requires less human thought. Instead of a thesis, it just requires an undeveloped thought such as the one in the previous paragraph.

But I'm going to start with something even less precise than the "Undeveloped Idea" set forth above. I'm going to challenge AI by giving it the equivalent of a grunt. "Legal doctrine has difficulty coping with cumulative laws in separate acts which, when combined, establish religion." Let's see what my CustomGPT can do to move from this quarter-baked idea into a real legal thesis.

Here's the result.

Iteration 1

Proposed Novel Claim

“Under the Establishment Clause, courts should examine the aggregate of a state’s religion-related statutes and regulations—rather than assessing each enactment in isolation—because a series of individually permissible measures can, in combination, create a de facto establishment of religion. I call this the cumulative-entanglement doctrine: whenever plaintiffs plausibly allege a pattern of legislative actions that collectively privilege religion (or a particular faith), courts must conduct a holistic review that weighs the overall statutory landscape, much as environmental law demands a cumulative-impact analysis.”

1. Ruthless Volokh Evaluation

2. Critical Test Suite (7 tough scenarios)

1. Education Package – State A (i) authorizes public funding of religious charter schools, (ii) mandates posting the Ten Commandments in every classroom, and (iii) schedules a minute of “student-led prayer.” Does the package, viewed holistically, cross the line even if each component has passed recent scrutiny?
2. Religious Tax Exemptions Mosaic – Over 15 years, Legislature B enacts small exemptions benefiting only churches: parsonage allowances, utility-bill rebates, and land-use waivers. Does the cumulative dollar value trigger Establishment concerns, or is each waiver independently too de minimis to matter?
3. Multi-Agency Preference – Governor C issues separate executive orders: highway signs quoting scripture, prison chaplaincy expansion, and a tourism campaign around a mega-church district. Can plaintiffs aggregate cross-agency actions to allege state endorsement?
4. Single-Faith Saturation – City D funds ten cultural heritage grants—nine to Christian groups, one to a secular museum. Does numerical dominance alone prove cumulative establishment?
5. Patchwork “Equal Access” – District E opens school facilities to “all after-school clubs” (including religious ones) but simultaneously lets teachers promote a Bible club in class. Is the neutral-access statute laundered into endorsement when paired with on-campus promotion?
6. Time-Sequence Problem – Statute F (religious monument) enacted in 1990; Statute G (faith-based housing vouchers) in 2024. How long can gaps be before the earlier law drops out of the cumulative calculus?
7. Competing Faiths – State G funds Christmas and Hanukkah displays, then adds a Hindu festival grant. Does diversity cure or compound cumulative endorsement?

3. Fatal-Flaw Watchlist

1. “Mushiness” – What is the threshold for “too much religion” across enactments? Without numeric or qualitative benchmarks, judges may default to intuition.
2. Doctrine-shopping – Recent cases (Kennedy v. Bremerton, 2022) reject Lemon in favor of “history and tradition.” A cumulative test may revive Lemon-like balancing that the Court just abandoned.
3. Justiciability & Standing – Plaintiffs may lack standing to challenge all statutes if they are not harmed by each. Need a theory of “programmatic” injury.
4. Federalism Overreach – Aggregating state and local acts could intrude on diverse political subdivisions’ autonomy.
5. Empirical Vagueness – Requires factual proof of social perception across statutes—expensive surveys, expert testimony—making litigation unwieldy.

4. Knowledge-Limitations Statement

• I have not exhaustively searched 2024-25 law-review databases; a deep Westlaw/Lexis review is essential to confirm the doctrine’s originality.
• Empirical claims about public perception would need social-science expertise.
• Comparative-law angles (e.g., Canada’s “cumulative infringement” in Charter cases) may illuminate but require further research.

Ready for Revision?

Would you like me to draft an improved, tighter version of the claim that addresses these weaknesses and incorporates clearer standards for courts to apply?

The methods used here draw on Eugene Volokh, Academic Legal Writing (5th ed. 2016). Any errors are mine alone and should not be attributed to Professor Volokh.

Iteration 2

The prompt

You can judge for yourself, but I think ChatGPT did a fine job in starting to think through the problem. It raises all sorts of interesting questions such as time sequencing and justiciability that (I hate to admit it) I had not thought about. I'm now going to tell it, yes, I want you to revise it, but throw in a little more human insight. Here's the new prompt.

Yes, please draft an improved, tighter version of the claim that addresses these weaknesses and incorporates clearer standards for courts to apply? I'd also like for you to think hard about the problem of remedy. What subset of laws am I suggesting the court strike down? The minimal subset needed to make the remaining laws acceptable. But how does one judge minimality? The number of laws? The effect of the laws? Also think about the incentives for legislatures who do kind of want to establish religion to chop their legal regime up into little pieces in order to evade establishment clause claims.

The result

Refined Claim: Cumulative-Entanglement Doctrine 2.0

“The Establishment Clause is violated when the aggregate of a jurisdiction’s religion-related enactments exceeds the level of government religious involvement historically tolerated in the United States. Courts should (1) identify a presumptive statutory cluster, (2) calculate its Aggregate Endorsement Score (AES) using objective weights, and (3) if the AES is unconstitutional, annul only the minimal subset of enactments whose repeal brings the AES below the constitutional threshold.”

Step-by-Step Framework

Why This Version Fixes Earlier Weaknesses

1. Mushiness → Metrics: Clear four-factor scoring system with capped ranges replaces amorphous “too much religion” language.
2. Post-Kennedy Compatibility: History-anchored benchmark respects the Court’s rejection of Lemon while still allowing holistic review.
3. Standing & Justiciability: Plaintiffs need injury from one act; pendent-jurisdiction principles let courts evaluate the cluster once triggered.
4. Remedy Precision: Minimal-subset algorithm borrows from severability analysis and redistricting fix-only-what’s-unconstitutional doctrine.
5. Legislative Incentives: Ten-year window + penalty multiplier reduces payoffs from salami-slicing or time-staggering religious privileges.

Open Research / Empirical To-Dos

• Historic Baseline Data: Requires archival research on 1789–1868 state religious appropriations and symbols.
• Weight Calibration: Social-science input (surveys, experiments) needed to translate symbolism and pervasiveness into reliable 0–3 scores.
• Disruption-Cost Metric: Economics literature on reliance interests can supply proxies (e.g., years of participant investment).

Next Step

If you’d like, I can now stress-test this refined claim with a fresh Volokh‐style critique and updated test suite—just let me know.

The methods used here draw on Eugene Volokh, Academic Legal Writing (5th ed. 2016). Errors and omissions are mine alone; Professor Volokh was not involved in creating this response.

Iteration 3

Probably 99% of readers of this blog or members of the legal profession would regard this response as absurd. Maybe AI isn't so smart after all! The idea of quantifying establishment is completely alien to the law. Plus the notion of pendent jurisdiction seems misplaced. And, yet, maybe the very absurdity of the response points to the real challenges of cumulative establishment regimes. Let's see if I can point out the flaw to ChatGPT and have it come back with something better.

The prompt

No, don't stress test it because your response is absurd. No court is going to come up with a quantification of establishment clause. Plus your use of pendent jurisdiction seems misplaced. And, yet, maybe we would need something like your ridiculous scoring system in order to address cumulative establishment. So I would like you to think about whether there are any alternatives and, if there are not, what it says about a kind of loophole in establishment doctrine that legislatures can exploit. Also think about whether this cumulative law problem is unique to establishment. (I suspect it is, but it probably does exist elsewhere in constitutional law). Is there a way to escape the dilemma: either use some sort of accumulation algorithm or permit establishment to be created by a thousand cuts.

The AI's Thought Process

Here's how o3 described its own thought process.

I need to consider alternative approaches to the legal issues raised. The user doesn't find the previous framework convincing, so I’ll look for others and address concerns about loopholes, cumulative harms, and if similar problems exist in constitutional law. Additionally, the user's mention of a dilemma suggests I should explore the issue further—whether it's a "death by a thousand cuts" situation. I'll avoid creating new algorithms and instead explore more practical standards and comparable doctrines, like in voting rights or the Eighth Amendment.

Okay, the user's idea seems to focus on proposing alternative standards, not using purely quantitative methods. I might suggest a “pattern or practice” approach akin to civil rights law for analyzing religious establishment cases. I could propose "course of practice" standards, structural injunctions, or a "sunburst" remedy. The idea is to reflect on how the courts approach aggregated burdens and anti-circumvention principles, possibly integrating existing legal doctrines or relying on the predominant-purpose test. I'll need to look into how this could evolve in legal practice.

The response

Below is a sober—non-quantified—re-think of the “death-by-a-thousand-cuts” problem, plus a survey of where the same structural gap crops up elsewhere in constitutional doctrine.

1. Why the Court’s Usual Toolkit Breaks Down

• Piece-by-Piece Review. Establishment cases (Lemon, Lynch, Van Orden, Town of Greece, Kennedy) treat each display, voucher, or chaplaincy individually. That works when the challenged action is a single monument on a courthouse lawn. It breaks when a legislature intentionally slices a privileging regime into many small, facially innocuous parts.
• Intent Evidence Is Too Thin. One could try Arlington Heights–style discriminatory-purpose analysis, aggregating circumstantial proof across enactments. But legislators can leave spotless records—no floor statements, no “Christian nation” tweets—so purpose is hard to prove.
• Standing & Severability. A plaintiff has standing for the law that personally affects her, but not necessarily for the nine sister statutes that create the broader religious ecosystem. If the court strikes only “her” statute, the ecosystem survives.

Result: a doctrinal loophole that savvy legislatures can exploit—privilege religion incrementally and dare courts to swat one gnat at a time.

2. Alternatives to a Full-Blown Numeric Algorithm

Bottom Line

None of these escapes the core trade-off: either you articulate a reasonably objective trigger (which risks looking legislative), or you tolerate strategic fragmentation. The choice is between administrability and under-enforcement.

3. Is the Problem Unique to Establishment?

So the phenomenon is not unique, but Establishment lacks any meaningful aggregation doctrine, while the others have at least embryonic tools.

4. Escaping the Dilemma? Two Modest Proposals

1. “Context-Plus” Test (Judicially Manageable)Why it helps: Uses a simple numeric trigger (three acts) but no scoring. Leverages the Court’s own history-and-tradition vocabulary to keep the rule recognizable.
   • Step 1 – Context: Apply existing Kennedy “history & tradition” test to each law.
   • Step 2 – Plus: If plaintiffs demonstrate three or more contemporaneous religion-specific acts bearing the same legislative fingerprint (common sponsors, same session, similar beneficiaries), the court presumes an Establishment violation unless the state shows a tradition-based analogue of comparable breadth.
   • Remedy: Court invalidates whichever subset the state fails to historicise; legislature can reenact with broader neutrality.
2. Legislative Self-Auditing Incentive (Non-Judicial)
   Congress or state constitutions could require a “Religious Neutrality Impact Statement” (RNIS) for any bill referencing religion. The RNIS must list all operative religion-related laws and certify that the new bill will not push the combined package beyond historical norms.
   Courts would give Chevron-like deference only if an RNIS exists. No RNIS → hard look; possible aggregate strike-down.Why it helps: Shifts the aggregation burden to the policymaker, reducing judicial administrability worries.

5. What If We Accept the Hole in the Net?

If no administrable aggregation tool is crafted, legislatures retain a “Swiss-cheese” route to informal establishment:

• Small-scale statutes evade scrutiny because each looks de minimis.
• Political incentives favor many modest perks over one large, litigable one.
• The Court’s turn toward accommodation (Espinoza, Carson, Kennedy) emboldens this strategy.

That result is not logically compelled by precedent—courts could adopt a pattern-or-practice or context-plus approach—but it is practically predictable unless doctrine evolves.

Takeaways for Your Paper

• Frame the dilemma starkly: Either judicially manage some holistic test, or concede a doctrinal gap ripe for exploitation. There is no neutral third path.
• Offer a menu (Pattern-or-Practice, Context-Plus, RNIS) and candidly weigh their costs.
• Situate the problem comparatively so readers see that other constitutional fields have confronted—and partially solved—analogous aggregation challenges.

The methods used here draw on Eugene Volokh, Academic Legal Writing (5th ed. 2016). Errors and omissions are mine alone; Professor Volokh was not involved in creating this response.

Iteration 4

And that, dear readers, is pretty darned good (except for Part 4 of the response, which I did not care for). And now we have the makings of a thesis. Let's just ask ChatGPT to write an abstract of 500 words. Here's the abstract.

Abstract — Cumulative Establishment and Constitutional Blind Spots
(≈ 500 words)

Modern Establishment Clause doctrine is exquisitely sensitive to a single crèche on the courthouse lawn yet strangely indifferent to a legislature’s incremental construction of a religion-favoring regime. This article identifies and critiques that blind spot, which arises because courts examine each challenged enactment in isolation while ignoring how a succession of individually innocuous measures can, in concert, tilt the public sphere toward sectarian preference. The result is a doctrinal escape hatch through which determined lawmakers may steer a “mosaic of minor statutes” that collectively produce the very establishment the Constitution forbids.

Part I traces Supreme Court doctrine from Everson to Kennedy v. Bremerton, emphasizing the steady retreat from bright-line rules toward context-specific balancing. Paradoxically, the same contextual turn encourages atomized review: judges ask whether this voucher or that monument violates the Clause, seldom whether the state’s overall program has crossed a constitutional tipping point. Drawing on recent legislative campaigns that pair faith-based school funding with devotional classroom mandates and religious tax exemptions, the article shows how a deliberate “death-by-a-thousand-cuts” strategy can flourish under current precedent.

Part II explains why familiar doctrinal tools cannot plug the gap. Purpose inquiries modeled on Arlington Heights falter when legislators keep the record clean; standing doctrine bars plaintiffs from challenging provisions that do not personally affect them; and severability norms ensure that striking a single subsection leaves the broader architecture intact. The structure of litigation, not judicial goodwill, produces systematic under-enforcement.

Part III surveys non-quantitative aggregation models that could be imported from adjacent areas of public law without forcing judges into the awkward business of numerically scoring “religious endorsement.” A pattern-or-practice approach (borrowed from Title VII) would allow plaintiffs to plead a constellation of related statutes and shift the burden to the state to justify the ensemble. Structural injunctions, familiar from prison-conditions cases, could place remedial design in the government’s own hands while subjecting the full statutory landscape to unified oversight. Alternatively, a presumptive-cap rule limiting the share of discretionary benefits that may flow to religious entities each biennium would provide a bright line without granular arithmetic. Each option trades precision for administrability, but all address the loophole more effectively than current doctrine.

Part IV contends that the aggregation problem is not unique to Establishment. Free-speech zones, voter-suppression tactics, Eighth-Amendment prison harms, and fragmented land-use regulations exhibit the same pattern of individually permissible acts that cumulatively erode constitutional guarantees. Yet in most of those fields courts have at least embryonic tools—such as the Brnovich “totality” test or the Eighth Amendment’s structural decrees—to prevent strategic fragmentation. Establishment jurisprudence stands out for lacking any comparable safety valve.

The article concludes that without some form of holistic scrutiny, constitutional protection against governmental religious favoritism will remain porous by design. Whether courts adopt pattern-or-practice reasoning, structural injunctions, or another administrable proxy, the choice is no longer between perfect quantification and doing nothing. It is between acknowledging cumulative harm or acquiescing in its steady advance. A brief footnote sketches—and critiques—a more radical numeric algorithm, offered less as a prescription than as a thought experiment underscoring the urgency of doctrinal reform.

Switching over to "Volokh my claim"

We now have something I would not be embarrassed to show synthetic Professor Volokh. So I switch over to the "Volokh my claim" CustomGPT which subjects a thesis to brutal (but fair) criticism. I copy the abstract and paste it into the Volokh GPT.

Volokh Round 1

Here is synthetic Professor Volokh's response.

1. Merciless Volokh-Criteria Check

2. Brutal Test-Suite (7 Scenarios)

1. Voucher-Plus-Prayer Package: State funds 100% of private-school tuition vouchers (many recipients are religious), later mandates daily classroom “moment of silence” with teacher-led prayer option. Does your framework let plaintiffs plead both statutes in one shot despite different beneficiaries and timeframes?
2. Historic-Preservation Carousel: Legislature passes five small tax-credit bills over ten years—each benefiting a single church for “historic restoration.” Standing? Aggregation?
3. Religious-Exemption Snowball: State carves out successively broader conscience exemptions (vaccines, anti-discrimination, licensing). Individually neutral; cumulatively skew civil obligations. Can your cap or pattern test catch non-monetary privileges?
4. County-By-County Creep: Fifty counties each erect one Ten Commandments monument. No county crosses Van Orden limits alone. Statewide effect unmistakable. Can plaintiffs aggregate across local entities?
5. Sunset Flip: Legislature enacts twenty small faith-based grant programs, each expiring after two years but re-authorized on a rolling basis. Does the temporal gap defeat “pattern” pleading?
6. Reverse-Fragmentation: State brazenly enacts a big pro-religion package; court strikes it. Lawmakers respond by splitting it into ten bills. Show that your approach treats the split the same as the omnibus.
7. Mixed-Motives State: Half the statutes at issue carry clear secular justifications (historic preservation, social-service delivery). How does your burden-shifting model separate benign from biased motives without lapsing into Lemon purpose analysis?

3. Potentially Fatal Flaws (Ordered by Severity)

1. Administrability Black Hole – Judges need a workable trigger. “Constellation” language risks becoming another indeterminate balancing test. Without a crisp pleading rule, courts will default to business-as-usual.
2. Standing Roadblock – Aggregation means plaintiffs must sue over laws that do not injure them personally. Current standing doctrine forbids that. You acknowledge but don’t solve it. Unless you articulate a fresh standing theory (citizen suits? organizational harm?), the project stalls.
3. Separation-of-Powers Anxiety – Structural injunctions in Establishment cases look like federal courts dictating state tax and spending policy—catnip for the Roberts Court to nix. Show why your remedial design respects political-question boundaries.
4. Mushy “Presumptive Cap” – Any cap needs justification (why 30 % and not 35 %?). Else it smacks of ad-hoc line-drawing—the same defect that doomed Lemon.
5. Literature Gap – Unless you prove prior scholars haven’t already framed cumulative harms (especially in Blaine-Amendment and education-funding debates) reviewers will brand the article retread.
6. Empirical Thinness – Claim that lawmakers strategically fragment. Provide documented legislative memos, floor statements, or interview data; otherwise skeptics will call it speculative.

4. Knowledge & Research Limits

• Legislative-Intent Evidence – Empirical political-science or FOIA work needed; beyond AI scope.
• Comparative Data – State-by-state compilation of religious appropriation totals is essential to prove aggregation exists.
• Historical Precedent – You may need archival work on post-Everson funding patterns; few digital sources are comprehensive.
• Remedial Success Rates – Structural injunction metrics (cost, compliance) in prisons/Title VII come from docket studies we can’t replicate here.

5. Next Steps

1. Articulate a Bright Trigger Rule (e.g., “When three or more statutes within a decade confer >X% of their benefits to faith-based entities, burden shifts to the state”).
2. Draft a Standing Work-Around (aggregated taxpayer standing? organizational membership harm?).
3. Marshaling Empirical Proof – Build a table of at least 15 state-level packages showing cumulative tilt.
4. Pre-Empt Conservative Objections – Explain why your approach is neutral and would also stop anti-religious cumulative burdens (wins credibility).
5. Short Literature Map – One-page memo distinguishing your thesis from Schwartzman, Gedicks, Garnett, and Esbeck.

Would you like me to write an improved version of your core claim incorporating this feedback?

The methods used here are derived from those of Professor Eugene Volokh, Academic Legal Writing: Law Rev Articles, Student Notes, Seminar Papers, and Getting on Law Rev (5th ed. 2016). Errors and omissions are my own; Professor Volokh was not involved in creating this AI.

Volokh Round 2

I then submitted the following query to ChatGPT.

Thanks. So I would like (A) an improved version of my core claim in the form of (1) a short thesis statement; (2) a revised 500 word abstract; and (3) the various products you outlined in "5. Next Steps"

The output from this prompt is really too long to reproduce fully here. I'll show you the revised abstract and provide a link to a Google Doc in which I stuffed the full response.

Revised Abstract (~500 words)

Establishment-Clause doctrine obsessively dissects discrete disputes—a single scholarship, monument, or prayer—while ignoring how today’s piecemeal victories combine into tomorrow’s de facto state church. Because litigation is fragmented and standing rules are narrow, judges seldom confront the whole policy mosaic; instead they green-light each tile. This article exposes that “cumulative establishment” blind spot and argues that it now dominates church–state conflict.

Part I recounts the Court’s arc from Everson to Kennedy v. Bremerton, showing how the demise of rigid tests (†Lemon, endorsement, coercion) unintentionally entrenched atomized review. Legislatures responded with death-by-a-thousand-cuts tactics: universal school-voucher bills (Texas, Florida, Tennessee), mandatory Ten-Commandments displays (Louisiana, Texas), chaplain statutes, released-time rules, and special tax breaks for church property—all individually defensible, yet collectively transforming the public sphere. A comparative survey of fifteen states demonstrates the phenomenon empirically.

Part II explains why classic doctrinal levers—purpose inquiries, taxpayer standing, severability—cannot reach aggregate harm. Plausible evil motives are scrubbed from the record; affected citizens lack personal injury for most provisions; and surgical invalidation of one subsection leaves the structure standing. The constitutional under-enforcement is structural, not accidental.

Part III proposes a workable fix that avoids Lemon-style arithmetic. Borrowing from Title VII, a pattern-or-practice presumption would arise once (a) three or more faith-targeted acts are adopted within ten years and (b) the aggregate fiscal benefit or regulatory preference crosses an objective “neutrality baseline.” At that point, the burden shifts: the state must prove the ensemble serves secular rather than sectarian ends. Courts could enforce the presumption through two remedial tools already familiar in public-law litigation: (1) structural injunctions that task the state with unwinding the package while preserving democratic choice, and (2) sunset-plus-cap decrees that automatically lift once the neutral baseline is restored. These mechanisms trade precision for administrability yet still seal the loophole.

Part IV contends the aggregation problem is replicating elsewhere—voter-suppression micro-measures, fragmented speech restrictions, cumulative prison deprivations—but only Establishment jurisprudence lacks a safety valve. Cross-doctrinal borrowing is both feasible and urgent. Finally, the article confronts conservative objections that holistic review “penalizes success” or resurrects Lemon. Because the trigger relies on quantity and trajectory—not theological judgments—it fits within Kennedy’s historical-practice framework while respecting judicial restraint.

Absent such holistic scrutiny, religious favoritism will continue to advance not by open confrontation but by calculated accretion. The choice is no longer between perfect measurement and passive acquiescence; it is between acknowledging cumulative harm or licensing it.

A Draft Law Review Article

As part of our "full stack development" I now shift over to my current favorite AI for long-form output: Deep Research from Google Gemini. I am going to provide it with the last response from the Volokh my claim CustomGPT and ask it to draft a 15,000-18,000 word law review article (with references) that researches and fleshes out the ideas. I would ask for something longer, but I find that asking for really long outputs or giving Gemini tons of information in a Deep Research prompt leads it to go out to research land and never return.

Here's the research plan Gemini developed.

(1) Conduct foundational research on the evolution of the Supreme Court's Establishment Clause jurisprudence, analyzing the doctrinal shift from the *Lemon* test to the 'history and tradition' standard in *Kennedy v. Bremerton* to frame the 'cumulative establishment' problem.

(2) Flesh out the provided empirical table by investigating the specific state laws mentioned (e.g., Texas's universal voucher, Louisiana's Ten Commandments law). For each, find legislative history, news coverage, and legal analyses to create concrete, real-world examples of the 'death-by-a-thousand-cuts' phenomenon.

(3) Analyze the doctrinal reasons why existing legal mechanisms fail to address aggregate harm. Research the limitations of: (a) Standing doctrine, particularly for taxpayer and citizen suits. (b) The judicial inquiry into legislative purpose. (c) The doctrine of severability in constitutional challenges.

(4) Critically explore the proposed 'pattern-or-practice' framework and its bright-line trigger rule. Research the use of such frameworks in other areas like Title VII, and evaluate the arguments for and against applying a quantitative, fiscally-based test in constitutional law, noting its novelty.

(5) Investigate the proposed remedial tools. Find and analyze legal precedents and scholarly commentary on the use of structural injunctions and sunset provisions in public law litigation, such as voting rights and prison reform, to assess their applicability and administrability in this context.

(6) Perform a cross-doctrinal comparison by researching how other areas of law handle aggregate harms, including voter-suppression measures, fragmented speech restrictions, and cumulative prison conditions, to argue for the feasibility of applying similar principles to the Establishment Clause.

(7) Develop robust responses to the anticipated counterarguments. Research the legal and philosophical underpinnings of objections related to neutrality, judicial manageability, federalism, and the *Kennedy* historical test, and formulate rebuttals based on the article's thesis.

(8) Situate the article within the current academic debate by locating and analyzing the works cited in the literature map (Schwartzman, Gedicks, Garnett, Esbeck, Litman) and other key scholarship on the Establishment Clause. Synthesize these sources to highlight the article's unique contribution.

That's a lot but I hope for the best, press the "Start Research" button, and decide to clean the kitchen while I wait for AI to generate my draft law review article. 18 minutes later two good things have happened: (1) the kitchen is cleaner, and (2) I have this draft article of 14,000 words.

Here is the introduction section. It has a concrete hook, a thesis and a roadmap paragraph.

The Establishment Clause of the First Amendment, in Thomas Jefferson’s memorable phrase, was intended to erect a “wall of separation between church and State.” ¹ For much of modern constitutional history, the judiciary has policed this wall by scrutinizing discrete government actions—a single prayer in a public school, a solitary religious monument on public land, a specific subsidy to a parochial institution. This mode of analysis, focused on individual tiles, has left courts myopically unprepared for a new legislative strategy that threatens to dismantle the wall not with a single battering ram, but by the patient and deliberate assembly of a sectarian mosaic.

Consider a hypothetical but increasingly familiar state legislative session. On Monday, the legislature passes a bill authorizing volunteer chaplains to provide “support, services, and programs” to students in public schools. ³ On Wednesday, it enacts a law mandating that a poster-sized display of the Ten Commandments be placed “in a conspicuous place in each classroom.” ⁴ On Friday, as part of a major budget bill, it creates a “universal” school voucher program, allocating hundreds of millions of taxpayer dollars to accounts that can be used for tuition at private religious schools. ⁶ Each of these measures, when viewed in isolation, might survive a constitutional challenge under the Supreme Court’s current, fragmented Establishment Clause jurisprudence. The chaplain program is “voluntary.” ⁸ The Ten Commandments display is justified as serving the “historical,” not religious, purpose of showcasing a foundational legal document. ⁹ The voucher program is facially neutral, available to parents to direct funds to secular or religious schools of their choice. ¹¹ A court, examining each tile for its individual defects, may well find none. Yet in doing so, it would fail to see the larger picture being constructed: a de facto state establishment of religion, advanced not by open confrontation but by calculated accretion.

This Article argues that this phenomenon—what it terms “cumulative establishment”—has become the dominant front in the modern conflict over church-state separation. It exposes a critical blind spot in First Amendment doctrine, which obsessively dissects discrete government actions while ignoring how today’s piecemeal victories for religious preference combine into tomorrow’s state-sponsored faith. Because litigation is fragmented and standing rules are narrow, judges seldom confront the whole policy mosaic. This structural failure of judicial review is not accidental; it is a predictable consequence of the Supreme Court’s demolition of generalizable Establishment Clause tests and its turn toward a myopic historical analysis that is ill-suited to address systemic harms. Legislatures, in turn, have astutely exploited this doctrinal vacuum, pursuing a death-by-a-thousand-cuts strategy to entrench religious favoritism in ways that evade holistic scrutiny.

This Article proceeds in four parts. Part I traces the doctrinal arc of the Establishment Clause from the mid-twentieth century to the present. It recounts the Supreme Court’s journey from the sweeping separationism of Everson v. Board of Education ¹² to the formulaic test of

Lemon v. Kurtzman ¹³, and through the subsequent collapse of doctrinal clarity that culminated in

Kennedy v. Bremerton School District. ¹⁴ It argues that

Kennedy’s embrace of a “history and tradition” standard, while jettisoning the much-maligned Lemon test, unintentionally entrenched an atomized mode of review that is structurally incapable of recognizing aggregate harms. Part I then provides a comprehensive empirical survey of recent legislation in fifteen states, demonstrating how lawmakers are leveraging this doctrinal blind spot to advance a coordinated, multi-pronged agenda of religious establishment through a combination of funding, symbolism, personnel, and curriculum.

Part II provides an anatomy of this constitutional underenforcement. It explains why the classic doctrinal tools available to courts are inadequate to address the challenge of cumulative establishment. Inquiries into legislative purpose are easily thwarted by carefully drafted pretexts. The modern, restrictive rules of standing—particularly for taxpayers and so-called “offended observers”—prevent any single plaintiff from challenging an entire legislative package. And the doctrine of severability, designed for surgical excision, is a poor instrument for remedying a systemic condition, as it allows courts to remove a single unconstitutional provision while leaving the broader architecture of religious preference intact.

Part III moves from diagnosis to prescription, critically exploring a workable framework for judicial review of cumulative establishment claims. It argues for borrowing the “pattern-or-practice” evidentiary framework from Title VII of the Civil Rights Act of 1964, which has long been used to address systemic discrimination. This Article proposes, as a mechanism to operationalize this borrowing, a bright-line, quantitative trigger rule that would shift the burden of proof to the state once a certain threshold of faith-favoring legislation is crossed. While this article does not endorse this specific formula as a definitive constitutional standard, it presents it as a serious proposal and rigorously analyzes its strengths—objectivity, administrability, and avoidance of theological entanglement—against its significant weaknesses, including its potential for arbitrariness and its jurisprudential tension with the historical methodology of Kennedy. Part III then explores potential remedies suited for systemic harm, including structural injunctions, which have a long history in public-law litigation, and a novel proposal for “sunset-plus-cap” decrees. It concludes by outlining a viable theory of “plural-plaintiff aggregate injury” to overcome the standing bottleneck.

Part IV situates the problem of cumulative establishment within broader constitutional law and confronts the most potent objections to the proposed framework. It demonstrates that the problem of aggregated micro-harms is not unique, and that other constitutional domains, such as voting rights and the Eighth Amendment, have developed doctrines to assess the “totality of the circumstances.” The Establishment Clause, this Part argues, is a conspicuous outlier in its lack of such a safety valve. Finally, this Part directly addresses and rebuts conservative objections that a holistic review of cumulative establishment would “penalize success,” prove judicially unmanageable, resurrect Lemon, or violate principles of federalism.

Absent a new doctrinal lens capable of perceiving aggregate harm, state-sponsored religious favoritism will continue to advance not by dramatic conquest, but by quiet, relentless accumulation. The choice facing the judiciary is no longer between perfect measurement and passive acquiescence; it is between acknowledging cumulative harm or licensing it.

But of course that's just the introduction. What about the rest of the article? I then had a human (me) evaluate the result to see if we could really go from undeveloped idea to draft law review article in the space of a morning with most of the work being done by AI. The good news is that I still have a job but perhaps a different kind of one than I imagined back in the Pleistocene era when I started as a law professor. In many ways, the article was better than something I could produce, certainly in a week let alone a morning. Here are some insightful passages.

Insightful passage 1

For decades, the Supreme Court sought to articulate general, workable principles to govern the relationship between church and state. The eventual collapse of that project, culminating in the Court’s recent decision in Kennedy v. Bremerton School District, created a doctrinal vacuum. In place of broad standards that assessed a law’s purpose and effect, the Court substituted a myopic historical inquiry focused on discrete government acts. This shift from a panoramic to a microscopic view left the Establishment Clause vulnerable. State legislatures, recognizing this new doctrinal landscape, have responded with a sophisticated and coordinated strategy: advancing a pro-faith agenda through a volley of individually defensible measures that, in the aggregate, achieve a level of religious establishment that would be plainly unconstitutional if enacted as a single statute.

Insightful passage 2

The evidence for this coordinated legislative strategy is overwhelming and can be seen across the country. In the short period between 2023 and 2025, numerous states have enacted a flurry of legislation that, taken together, systematically tilts the public sphere toward a favored set of religious beliefs and institutions. This legislative push can be categorized along several vectors: direct and indirect funding for religious education, the promotion of religious symbolism in public spaces, the integration of religious personnel into public schools, and the infusion of religious content into public school curricula and schedules.

Insightful passage 3

This article proposes a theory of "plural-plaintiff aggregate injury." ... Under this theory, a court would permit a coalition of plaintiffs to bring a cumulative establishment claim, provided that each member of the coalition can demonstrate a concrete, particularized injury from at least one component of the challenged legislative package. For example, a coalition could consist of (1) a non-adherent student who is directly and regularly confronted with a mandatory Ten Commandments display, (2) a taxpayer in a district whose public schools are losing funding to a state voucher program, and (3) a secular teacher who objects to the introduction of school-sponsored chaplains. Individually, each plaintiff only has standing to challenge one piece of the puzzle. Collectively, however, they have standing to challenge the entire scheme.

Shortcomings

On the other hand, the article needs a lot of work before I would press send to the various law review evaluation platforms. It doesn't really understand footnotes and uses references to a bibliography instead. (Someone should coax an AI to write a program that performs the conversion). It doesn't engage enough with the scope of Supreme Court case law in the Establishment field. The severability analysis is important but only part of the picture. And I am still not sold on its proposal of having a semi-quantitative standard for determining when the cumulation of laws promoting religion turns into an Establishment clause violation.

I also had another AI (Grok3) evaluate the article. Here's its analysis. It generally liked it and, in response to my prompt, said it did not think the article was written by AI: "The article's style, depth, and originality are consistent with high-quality legal scholarship, and there are no definitive characteristics (e.g., repetitive phrasing, logical errors) that would indicate AI involvement." Oops! Like me, however, Grok3 thought the proposed rule seemed arbitrary and found the engagement with Supreme Court cases to be limited.

Conclusion

So where are we? It seems clear to me that AI can develop interesting legal ideas with a little bit of guidance from a knowledgable human. (Maybe you can test what happens if you just tell an AI: "Come up with an interesting legal theory and develop it into a law review article."). It can work interactively and iteratively to refine and improve that thesis, though occasionally needing to be knocked back on track by that expert human. And with Deep Research it can create plausible long-form law review style essays that at least form the basis for an article suitable for submission to law review. In 2025, we have full stack "vibe writing" in law but not yet automated production of legal scholarship. Will it be the case that traversing the remaining 30 or 40% will be as easy as getting to where we are now, with AI doing 60 - 70% of the needed work? I suspect not. AI may well obey a law of diminishing returns. But a year or two from now when I suspect AI will in fact get to 80 or 90% of the way there and the legal space can thus be filled with ever more scholarship with ever greater ease, we need to think very hard about what the job of a law professor is.

Postscript: Scholarship as Algorithm

This experiment in "full stack" AI scholarship forces a final, more speculative question. My conclusion suggests the job of a law professor is shifting toward that of an expert human guide, steering powerful but sometimes errant AI. But perhaps the change is more fundamental. What if AI alters not just the production of scholarship, but its very nature?

In mathematics, one does not define a vector space by listing all possible vectors; one provides an algorithm that can generate them. This post, in detailing its own creation, hints at a similar future for legal scholarship. Its most valuable output may not be the 14,000-word draft article—a static, reified idea—but the documented workflow that produced it. The sequence of prompts, the critical feedback that corrected the AI’s "absurd" quantitative proposals, and the strategic choice of tools together form a "generator function" for a legal theory.

Viewed this way, the idea of "cumulative establishment" is not merely what is written; it is the replicable algorithm that can explore it. The core intellectual work becomes designing this process—architecting the inquiry itself. This reframes the problem of AI's "diminishing returns". The goal is no longer to create a single, perfect artifact, but a robust, dynamic process that can generate endless variations of the argument on demand. The future of scholarship may lie less in publishing a fixed conclusion and more in providing the generative code for thinking about it.