Rasa Legal Uses AI to Help Clear Criminal Records

Rasa Legal reached Google News after NPR profiled its attempt to use AI for criminal record clearance, but easier access does not guarantee correct legal outcomes.

The company was founded by Noella Sudbury, a former public defender and criminal justice reform advocate. Its service helps people determine whether records might qualify for sealing or expungement under state law.

That mission gives legal AI a very different test from drafting contracts or summarizing lawsuits. A mistake can affect someone’s employment, housing, education, licensing, and ability to move beyond an old case.

The central conflict is therefore not AI against lawyers. It is an automated starting point against a legal process that remains fragmented, expensive, and difficult to navigate without professional help.

Rasa says its technology can reduce that friction while its lawyers supervise the legal work. Critics of automated legal services still have a fair question: Can a system make help easier without hiding uncertainty or encouraging misplaced trust?

The Google News Story Is About Access, Not Automation Alone

Rasa Legal is applying AI to a narrow legal bottleneck, identifying possible record-clearance options and helping eligible users begin the required process.

Criminal record clearance is not a single nationwide procedure. Each state sets its own eligibility rules, waiting periods, filing requirements, exclusions, and definitions of sealing or expungement.

Those distinctions matter. Sealing commonly limits public access to a record, while expungement changes how government systems treat it under applicable law. Neither concept has one universal meaning across every jurisdiction.

Rasa’s mobile-friendly tool asks users for identifying and case information. According to the company, it then provides legal information about potential eligibility under the law of a supported state.

The company also offers attorney-supported services for cases that require a petition. That structure separates Rasa from a general chatbot offering unsupervised answers about any legal problem.

Sudbury founded Rasa after working as a public defender and on criminal justice policy in Salt Lake County. She had also participated in efforts surrounding Utah’s Clean Slate reforms.

Utah’s law created an automatic pathway for specified records. However, many cases remain outside that pathway and still require an individual petition, supporting documents, government processing, or judicial review.

That gap is the commercial and social opening Rasa is addressing. A right that exists only on paper provides little value when people cannot determine whether they qualify.

The NPR story surfaced through Google News as a human account of second chances. The underlying technology story is less sentimental and more consequential.

Rasa is trying to convert scattered records and complicated eligibility rules into a manageable workflow. It is using software to perform early sorting before lawyers handle decisions requiring legal judgment.

That approach can reduce repetitive work. It can also expose more people to legal information without requiring every potential client to begin with a full attorney consultation.

The company says its platform serves people in Utah, Arizona, and Pennsylvania. Its March 2026 announcement said it planned to use a new funding round to expand product development and enter more states.

That expansion brings a difficult engineering requirement. Every jurisdiction adds another changing collection of statutes, court rules, record systems, and administrative practices.

A legal eligibility engine cannot remain accurate through model training alone. It needs structured rules, reliable records, jurisdiction-specific updates, and a process for escalating ambiguous cases.

The record-clearing profile therefore captures only the visible part of the project. The harder work sits beneath the interface.

Rasa must interpret records, match them to current law, communicate uncertainty, and protect sensitive information. It must then help users complete steps that courts and agencies will accept.

The company’s promise is not that an AI model should decide who deserves relief. It is that technology can make an existing legal remedy easier to understand and pursue.

That narrower claim gives Rasa a more defensible position than many consumer legal bots. It also makes its accuracy and supervision easier to evaluate.

A Criminal Record Can Outlast the Sentence

Record clearance matters because the practical penalties associated with a case often continue long after the court-imposed sentence ends.

Employers commonly use criminal background checks. Landlords, professional licensing bodies, educational institutions, and government programs can also consider information connected to arrests or convictions.

These effects are called collateral consequences, meaning legal or practical restrictions that follow a criminal case beyond incarceration, probation, or a fine.

A record does not need to describe a serious conviction to create friction. Arrests, dismissed charges, outdated entries, and minor cases can still appear in databases used during screening.

The federal consequences inventory documents restrictions involving employment, housing, public benefits, education, licensing, and civic participation.

The exact impact depends on the record, jurisdiction, job, and screening system. It would be inaccurate to assume every record produces the same barrier.

However, the recurring problem is clear. People can remain subject to adverse assumptions after completing every formal requirement imposed by a court.

Background-check systems add another layer. Records may be copied by commercial providers, separated from later case updates, or presented without enough context for a reader.

A government database might eventually show that a case was dismissed or sealed. A private copy can remain available until the provider receives and processes corrected information.

The employment research summarized by the Urban Institute connects criminal background checks with the broader challenge of finding stable work after justice-system involvement.

Employment does not erase every risk or solve every reentry problem. Still, access to legitimate work is a central part of rebuilding financial stability.

Record clearance can therefore affect more than reputation. It can change whether an applicant reaches an interview, receives housing consideration, or qualifies for a particular occupational license.

That is why the burden of petition-based systems deserves attention. The process can require someone to collect records, classify cases, calculate waiting periods, obtain certificates, notify agencies, and file documents correctly.

Each step appears manageable in isolation. Together, they can exclude people who lack time, money, reliable transportation, legal knowledge, or confidence dealing with courts.

Traditional legal services can resolve that complexity, but they are not available to everyone. Free legal clinics face limited capacity, while private representation can remain beyond a household’s reach.

Automatic clean-slate laws address part of the problem by shifting work from individuals to government. The state identifies eligible records and processes them without requiring each person to hire counsel.

Utah’s experience shows both the value and limits of that model. Its Clean Slate law took effect in 2022, covering specified categories while leaving other records in a petition-based system.

The Utah Department of Public Safety reported receiving 474,480 records for clean-slate processing in its 2024 annual report. It had processed 297,531, with 176,012 remaining at the reported point.

Those figures describe records, not necessarily an equal number of people. One individual can have multiple cases or record entries.

The same report described thousands of ordinary expungement applications, certificates, court orders, and manually removed incidents. That workload illustrates why implementation matters as much as statutory authorization.

A person should not have to understand the architecture of court databases to benefit from a law designed for rehabilitation. Yet government automation does not cover every eligible case.

This is where Rasa’s model becomes relevant. It gives individuals a tool for assessing the remaining petition-based possibilities.

The goal is not merely faster paperwork. It is lowering the threshold between wondering whether relief exists and receiving a credible, jurisdiction-specific answer.

For knowledge workers evaluating personal information systems, the lesson is familiar. Finding a record is not the same as understanding its current meaning.

A reliable workflow needs provenance, status, applicable rules, and human review. That principle also underlies responsible knowledge management, although criminal records demand much stronger safeguards than ordinary workplace documents.

Rasa Legal Turns a Legal Maze Into a Supervised Workflow

The strongest case for Rasa is not that AI knows the law better than lawyers, but that software can handle structured screening while lawyers retain responsibility.

Rasa says its platform offers an instant eligibility tool and attorney-supported record-clearing services. Users can begin through a mobile interface instead of assembling the process independently.

The first task involves gathering enough information to locate and classify relevant records. The system must distinguish arrests, charges, dispositions, convictions, and completed sentences.

It must then apply jurisdiction-specific rules. These can include waiting periods, offense exclusions, outstanding obligations, limits on conviction counts, and pending-case restrictions.

Some rules look suitable for deterministic software, meaning conventional logic that produces the same result from the same verified inputs.

Other issues require interpretation. A case record can be incomplete, inconsistent, or unclear about the exact statute and final disposition.

Generative AI can assist with document extraction and plain-language explanations. It should not silently replace rule-based checks or an attorney’s analysis when eligibility is uncertain.

This distinction is essential because large language models generate probable text. They do not inherently guarantee that a citation exists or that a legal rule remains current.

Rasa’s supervised structure offers a response to that weakness. Lawyers can review results, handle petitions, and address cases that do not fit cleanly into automated categories.

That model also creates accountability. A user has a legal service provider responsible for the work, rather than an anonymous chatbot producing a confident paragraph.

Rasa has operated within Utah’s regulatory sandbox, a court-authorized framework that allowed approved organizations to test alternative legal service models under oversight.

The company received authorization focused on criminal expungement services. That defined scope matters because it limits what the system and associated professionals are permitted to offer.

The relevant court authorization describes the approved legal service category as criminal expungement only.

A sandbox does not certify that every output is correct. It creates a controlled setting where regulators can monitor service models that do not fit traditional law-firm structures.

Utah’s approach responded to a longstanding access problem. Legal professional rules protect consumers, but they can also restrict alternative delivery models that combine lawyers, nonlawyers, and software.

The challenge is preserving the protective function while permitting experimentation. Removing professional oversight entirely would not solve that problem.

Rasa’s expansion beyond Utah adds complexity because another state’s legal market may follow different rules. Authorization in one jurisdiction does not automatically transfer elsewhere.

The company must establish an appropriate service model in every supported state. It also needs mechanisms to update its eligibility logic when lawmakers or courts change the rules.

According to Rasa, its platform has already been used by thousands of people seeking eligibility information. Earlier reporting on its Arizona launch said more than 10,000 users had used the app by late 2023.

That reporting also said the company had helped about 400 people clear records at that time. Those were company-provided figures and should be read as such.

Rasa announced a $5 million late seed round in March 2026. Rethink Education led the financing, with participation from Halogen Ventures, Social Finance, and the Richard King Mellon Foundation.

The company said the money would support hiring, product development, and nationwide expansion. Funding is an input, however, not evidence that the service will scale safely.

Every new state increases maintenance demands. Laws change, court forms are revised, agency systems fail, and old records resist clean classification.

That makes Rasa less like a single AI application and more like legal infrastructure. Its value depends on connecting data, rules, people, documents, and government processes.

The funding announcement says Rasa’s service covers Pennsylvania, Utah, and Arizona. It also repeats the company’s ambition to expand nationally.

The expansion question is therefore operational. Can Rasa preserve local accuracy while increasing geographic reach?

A generic chatbot can enter every state instantly because it accepts no meaningful responsibility for completion. A supervised legal service must move more carefully.

That slower path can look less impressive in a product announcement. It is also more appropriate when a flawed answer can leave a person believing a damaging record has been cleared.

AI Reduces Friction but Does Not Remove Legal Risk

The principal risk is false confidence, especially when incomplete records or changing laws make a simple eligibility answer unreliable.

A false positive could tell someone that a record qualifies when it does not. The user might spend time, disclose sensitive information, or make decisions based on relief that remains unavailable.

A false negative creates a different harm. Someone who qualifies could abandon the process because the system incorrectly described the case as ineligible.

There is also a completion risk. A correct eligibility assessment does not mean every court, law-enforcement database, and commercial background provider has updated its records.

Users need clear status information. “Potentially eligible,” “petition filed,” “order granted,” and “record updated” represent different stages.

An interface that collapses those stages into one optimistic message would mislead users even if its initial legal analysis were correct.

Rasa should therefore be judged by more than speed. Useful performance measures include correction rates, attorney escalations, petition outcomes, processing time, and user comprehension.

The company should also explain which parts of its system use generative AI. “AI-powered” is too broad to tell users how an eligibility result was produced.

A rules engine applying current statutes carries different risks from a language model summarizing an uploaded case document. Both require testing, but the tests should not be identical.

Legal hallucinations remain a documented problem. Courts have sanctioned lawyers who filed briefs containing fabricated citations generated by AI systems.

Those incidents do not prove that every legal AI product is unsafe. They show why professionals must verify machine-produced claims before relying on them.

The American Bar Association’s ethics guidance says lawyers using generative AI must consider competence, confidentiality, communication, supervision, candor, and reasonable fees.

Competence includes understanding enough about a tool’s benefits and risks to use it responsibly. It does not require every lawyer to become a machine-learning engineer.

Confidentiality is particularly important for Rasa. Criminal records contain identifying details, case histories, and information that users might not want reused for unrelated purposes.

The service needs clear retention policies, access controls, security practices, and vendor restrictions. A record-clearing product should not create a new permanent profile of the person seeking help.

Data minimization offers a useful principle. The system should collect only information needed for the service and retain it only as long as legally and operationally necessary.

Human review must also be substantive. A lawyer approving every machine result without checking source records would provide supervision in name only.

The reviewer needs access to the underlying record, the applicable rule, and the reasoning connecting them. Otherwise, the interface simply transfers automation bias to a professional.

Automation bias occurs when people defer to a system because its result appears precise or neutral. Lawyers are not immune to it.

A well-designed product should make uncertainty visible. It should show missing information, conflicting entries, and reasons a case requires attorney review.

That approach may produce fewer instant answers. It would also make the system more trustworthy.

Consumers need boundaries as well. Rasa’s focused workflow should not imply that a successful state record-clearance process resolves immigration, federal, licensing, or private database consequences.

Some background checks may lawfully access information unavailable to ordinary members of the public. Certain government inquiries also follow specialized disclosure rules.

A cleared record may improve opportunities without erasing every historical trace. Marketing language should preserve that distinction.

The wider legal industry offers an instructive comparison. General-purpose tools promise assistance across research, drafting, discovery, and client communications.

Rasa is betting on depth instead. It focuses on a repeatable problem with structured data, identifiable laws, and a concrete outcome.

That narrower scope makes responsible automation more achievable. It does not make it automatic.

The tension is ultimately between access and assurance. A service that demands perfect data before helping anyone will reproduce the barriers it seeks to remove.

A service that treats every record as simple will produce avoidable mistakes. The credible route lies between those extremes.

What the Next Expansion Must Prove

Rasa’s next phase must demonstrate accurate state-by-state expansion, measurable case completion, and consumer protection that survives higher volume.

The first signal to watch is how the company enters additional states. A launch should include jurisdiction-specific legal support, updated rules, and a clear description of available services.

A marketing page alone would not establish readiness. The stronger evidence would be published authorization, named legal supervision, and documented workflows for local courts and agencies.

If Rasa expands while maintaining narrow, well-supported services, its supervised model gains credibility. If coverage grows faster than local legal operations, the core argument weakens.

The second signal is outcome reporting. User registrations and eligibility checks measure interest, but they do not show whether records were successfully cleared.

Rasa should report how many people received preliminary information, proceeded to representation, filed petitions, obtained orders, and confirmed database updates.

It should also disclose correction and escalation rates. Those figures would help regulators and researchers distinguish a functional legal service from an attractive intake tool.

Privacy rules can limit case-level disclosure. Aggregate reporting can still provide meaningful evidence without exposing individual clients.

The third signal is regulatory response. Utah’s legal-services sandbox gave Rasa room to test a model that blends software, attorneys, and alternative business structures.

Other jurisdictions may adopt different approaches. Some will focus on unauthorized practice concerns, while others may emphasize access benefits and supervised experimentation.

The Utah AI process illustrates one oversight model. It uses negotiated conditions and recurring evaluations to assess risks and benefits.

Rasa’s success would strengthen the case for carefully scoped legal innovation. Significant consumer harm or opaque AI practices would push regulators toward tighter restrictions.

Google News exposure can attract users and investors, but public attention also raises expectations. The company now needs evidence that its system works beyond a compelling profile.

The broader clean-slate movement will shape that evaluation. Governments are increasingly automating relief for cases that meet clear statutory conditions.

Private services should not become permanent substitutes for public implementation. Their strongest role may be serving people whose cases remain outside automatic systems.

That includes users with mixed records, older cases, complex dispositions, or petition-based options. These are precisely the situations where software requires meaningful professional review.

Rasa can also help reveal recurring administrative failures. Aggregated, privacy-protected data might show where users encounter missing records, inconsistent classifications, or delayed updates.

Such findings could improve public systems. They could also help lawmakers identify rules that create unnecessary barriers without advancing safety.

The company should resist overstating what its AI contributes. Better document processing and eligibility screening are valuable even when they do not resemble an autonomous lawyer.

Indeed, autonomy is the wrong measure here. The better question is whether the system helps qualified professionals serve more people without reducing accuracy or accountability.

For people carrying old records, the distinction is practical. They need a reliable answer, an understandable process, and confirmation when the work is complete.

For lawyers, Rasa is a test of whether technology can expand a practice area often constrained by limited client resources.

For regulators, it tests whether oversight can support new delivery models while preserving confidentiality, competence, and meaningful remedies for errors.

For AI product teams, the lesson extends beyond legal services. High-stakes automation must connect every recommendation to verified data, current rules, and a responsible human decision-maker.

The story discovered through Google News is therefore not evidence that AI has solved criminal record clearance. It is evidence that a narrow, supervised application has found a genuine access problem.

The next one to three months should clarify whether Rasa announces another state, publishes stronger outcome data, or receives new regulatory authorization.

Each development will sharpen the same judgment. Expansion supported by local expertise will strengthen the model, while scale without transparency will weaken it.

Second chances should not depend on mastering paperwork that even experienced professionals find difficult. They also should not depend on an unverified machine answer.

Readers should ask a direct question as Rasa grows: Does each new layer of automation make the legal process more understandable and accountable?

If the answer is yes, legal AI will have delivered something more useful than a dramatic courtroom robot. It will have helped people exercise rights they already possess.