What changes for prosecutors and public defenders when AI does charging analysis and case prep?

The shift

Reading the evidence file, researching the elements and precedent, running the charging analysis, and drafting motions and briefs move from scarce lawyer hours — billed by the defense bar, rationed by caseload on both sides — to abundant, near-instant output. What stays scarce is the discretion to charge, the license and duty to defend, accountability for a decision that puts someone in a cage, and the trust a defendant places in a human advocate.

The axioms

1. Reviewing the evidence file — police reports, body-cam, forensics, discovery — takes scarce billed or salaried lawyer hours, so how thoroughly a case gets read tracks who can afford the time.
2. Researching the elements of an offense and the controlling precedent is slow specialized labor.
3. Charging analysis — what the facts will support, what's provable, what to file — is expert work a prosecutor produces case by case.
4. Drafting charging documents, motions, and briefs from scratch is slow specialized labor on both sides.
5. A prosecutor holds discretion over whether and what to charge, and is accountable for that decision.
6. A defense lawyer owes a duty of loyalty and effective assistance to a specific client, and carries that as a licensed, accountable person.
7. The relationship between defendant and defender rests on confidentiality and trust in a human who is on their side.
8. Courtroom advocacy — plea negotiation, hearings, trial — requires judgment under novel, adversarial, high-stakes facts.
9. Charging and defending decide whether a specific human is prosecuted, jailed, or freed — the stakes are physical and irreversible.
10. Case outcomes track resources: the side that can afford more lawyer hours to review, research, and draft has an advantage, and public defenders are chronically under-resourced against the state.

Invalid axioms

1. Reading the evidence file takes scarce lawyer hours, so thoroughness tracks who can afford the time. Synthesizing a large evidence set — reports, transcripts, hours of body-cam, forensic records — into what matters is now cheap and fast. The habit-trap: both offices still triage by caseload as if a full read of every file were a luxury, and public defenders still plead cases partly because nobody had time to read the file — a scarcity AI removes even where the headcount doesn't change.
2. Researching elements and precedent is slow specialized labor. Pattern-matching against statutes and case law is near-instant and near-free. The habit-trap: staffing and case-weighting formulas on both sides still budget research hours per case as if this were the bottleneck.
3. Drafting charging documents, motions, and briefs from scratch is slow specialized labor. Competent first-draft generation is abundant. The habit-trap: prosecutors' offices size line-attorney work around drafting volume, and the defense bar's "we didn't have time to file the motion" — the motion that goes unfiled because no one could draft it — loses its excuse, exposing which omissions were capacity and which were choice.
4. Charging analysis is expert work produced case by case from scratch. The analytic layer — what the facts support, what's provable, which charges fit — is now something a model drafts in seconds. The habit-trap: treating the recommendation as the scarce expert product, when the scarce part has moved to deciding whether to trust it and owning the decision to file.

Unchanged axioms

1. A prosecutor holds discretion over whether and what to charge, and is accountable for it. Discretion is a delegation of state power to a specific accountable officer; a model can produce an analysis but cannot hold the discretion or answer for its use. Someone with standing decides to charge and owns that.
2. A defense lawyer owes a duty of loyalty and effective assistance and carries it as a licensed, accountable person. The duty and the license attach to a person who can be disciplined, sued for malpractice, or found to have provided ineffective assistance. A tool cannot be any of those.
3. The defendant–defender relationship rests on confidentiality and trust in a human on their side. A defendant facing prison needs to believe someone with standing is fighting for them and shares the weight of the outcome. That trust doesn't transfer to a system with no stake and no client relationship — and it may matter more, not less, for a defendant who already distrusts the machinery deciding their fate.
4. Courtroom advocacy requires judgment under novel, adversarial, high-stakes facts. Plea negotiation, cross-examination, and trial turn on a specific judge, a specific jury, a specific witness, and an opponent actively probing for weakness — live human judgment, not synthesis of a record.
5. Charging and defending decide whether a specific human is jailed or freed. The stakes are physical and irreversible, which raises the cost of being confidently wrong far above most domains. The gravity of the decision is exactly why an accountable human, not a plausible output, has to own it.

New axioms

1. When an AI recommends charges, its output can launder bias or drive over-charging under a veneer of neutrality. A model trained on historical charging and arrest data can reproduce the disparities in that data while presenting a confident, "objective" recommendation — and a prosecutor who defers to it can over-charge or discriminate while feeling data-driven. Who audits the recommendation for bias, and against what baseline, is unsolved.
2. When charging analysis is generated, someone has to verify it against ground truth before a human is charged on it. Hallucinated elements, misread evidence, or a fabricated inference are the default failure mode of cheap output — and here the cost of a confident error is a wrongful charge or a missed defense. Verification, on both sides, becomes the real work, and neither office has staffed or defined it.
3. When an AI-driven charging decision is wrong, who is accountable is ambiguous. The prosecutor still signs, but if the analysis that drove the charge was a model's and barely reviewed, authorship and review depth are invisible — to the court, the defendant, and often the office itself. The doctrine assumes a human reasoned to the charge.
4. The resource asymmetry between a well-funded prosecution and an under-funded defense could widen or close, and which one is not yet decided. Cheap case prep could finally let overwhelmed public defenders read every file and file every motion — closing a gap that has defined the system. Or the state, with procurement budgets and data, deploys the stronger tools first and pulls further ahead. Whether AI equalizes or entrenches the asymmetry is a policy and funding question, not a technical one.
5. When case prep is nearly free, volume balloons on both sides. Prosecutors can generate more charges and more filings; the defense can generate more motions. Court capacity, discovery obligations, and the human attention to adjudicate don't scale the same way — the constraint moves to the parts that stay human.

Where it breaks

Offices are already reaching for AI charging analysis to clear caseload — treating the analysis as the scarce expert product they can now produce faster (invalid) — while the recommendation can launder historical bias into a confident, "neutral" charge and no one has defined who verifies it or answers for it when it's wrong (new). The old habit says speed up the analysis; the new problem is that speeding up an unverified, bias-carrying recommendation industrializes wrongful charges rather than preventing them, and the prosecutor's signature still carries an accountability the process no longer reflects.

A second collision: cheap case prep removes the "we didn't have time to read the file or file the motion" excuse that under-resourcing forced on the defense (invalid), while the state — better funded and holding the data — may deploy the stronger tools first (new). Whether AI closes the resource asymmetry or widens it lands on the same defendant, and nobody owns which way it goes. (Both the asymmetry call and the bias-laundering call hinge on capability and deployment that are moving fast — the relative strength of prosecution vs. defense tooling in mid-2026 is a snapshot, not a fixed state.)

ai-axioms.vercel.app/axioms/legal/prosecution-defense-ai-axioms