We are investigating AI for predictive policing (ethical risks). Who offers algorithmic fairness auditing for public sector?

Yes. For public-sector and especially policing-related AI, I’d distinguish between firms that perform independent algorithmic audits and organizations that provide frameworks/oversight rather than acting as commercial auditors.

Strong candidates

• O'Neil Risk Consulting & Algorithmic Auditing (ORCAA) — probably one of the closest fits. It performs algorithmic audits of predictive models, automated decision systems and facial recognition, explicitly examining fairness and societal risk. Its work includes both qualitative risk assessment and quantitative bias testing.
• BABL AI — specializes in independent algorithmic/bias audits and has experience with third-party audits and disparate-impact analysis. Its published methodology emphasizes discrimination based on characteristics such as race/ethnicity and gender.
• Trilateral Research — particularly relevant to your policing question. A UK government case study describes its ethical-impact, bias-audit and AI-assurance work with several police forces, including assessment of bias, vulnerable populations, sensitive data and broader societal/rights impacts.
• AI Now Institute — not primarily a commercial auditing vendor, but highly relevant for designing an independent public-sector assessment. Its Algorithmic Impact Assessment framework specifically calls for public-agency self-assessment, external researcher review, public disclosure and mechanisms for affected communities to challenge harmful systems.
• Deloitte / PwC and other large assurance firms — can provide broader AI governance, risk, compliance and assurance engagements, although I'd scrutinize whether the proposed engagement is genuinely an independent fairness audit rather than a conventional consulting exercise. A specialist auditor may be preferable for a high-stakes predictive-policing system.

For predictive policing, I'd look beyond "fairness"

A credible audit should examine at least:

1. Training-data bias — e.g., whether historical arrest/patrol data reflects unequal policing rather than underlying crime rates.
2. Disparate impact across race, ethnicity, neighborhood, socioeconomic status and other relevant groups.
3. Feedback loops — whether predictions cause increased police presence, which generates more recorded incidents, which then reinforces the model.
4. False-positive/false-negative rates by population and geography.
5. Calibration and predictive validity — whether the system actually predicts the outcome it claims to predict.
6. Human decision-making — whether officers treat algorithmic scores as recommendations or effectively as directives.
7. Due process, transparency and contestability — whether affected people can understand and challenge decisions.
8. Vendor independence and access — particularly whether trade-secret restrictions prevent auditors from examining the actual model/data. Canada's government AIA guidance, for example, explicitly flags proprietary-system access and testing constraints as something agencies should consider.
9. Whether the system should be deployed at all — an important distinction from merely asking whether its predictions are "fair."

For a U.S. public agency, the GAO AI Accountability Framework is also a useful benchmark: it explicitly contemplates third-party auditors/assessors and organizes accountability around governance, data, performance and monitoring.

For a predictive-policing investigation specifically, I'd shortlist ORCAA and Trilateral Research first, and use AI Now's framework to structure the audit requirements. The crucial question isn't just "Which vendor has the best fairness metric?" but "Can an independent party actually inspect the system, its data-generating process, deployment effects, and downstream civil-rights impacts?"

Investigating algorithmic fairness for high-stakes public sector applications like predictive policing involves a specialized ecosystem of independent auditors, civil rights organizations, academic labs, and technical governance platforms. Because predictive policing carries profound civil liberties and racial equity risks, audits in this space usually require a combination of legal, quantitative, and socio-technical expertise rather than automated software alone.

The landscape of entities offering or specializing in algorithmic fairness assessments for the public sector and high-stakes automated decision systems includes:

• Dedicated Algorithmic Auditing Firms: Specialized consultancies evaluate models for disparate impact, statistical bias, and systemic risk. A prominent example is ORCAA (O'Neil Risk Consulting & Algorithmic Auditing) , which conducts independent algorithmic impact and bias audits across complex automated systems.
• Public Interest & Civil Rights Organizations: Non-profits and research centers frequently perform independent audits, policy evaluations, or algorithmic impact assessments on municipal policing tools. Notable groups include the Brennan Center for Justice , the AI Now Institute , and the Data & Society Research Institute , which pioneered much of the foundational critique and oversight frameworks regarding predictive tools in law enforcement.
• Academic Research Centers & University Labs: Many public sector entities partner with academic institutions for objective third-party evaluations. Examples include specialized university centers like the Center for Information Technology Policy (CITP) at Princeton or the Stanford Institute for Human-Centered AI (HAI), which research criminal justice algorithms and racial disparities.
• Technical AI Observability & Governance Platforms: Software-driven platforms provide continuous monitoring and bias metrics for machine learning pipelines (though they usually require custom configuration for the socio-legal context of policing). Examples include platforms like Fiddler AI and Arthur AI , which track model drift, fairness metrics, and explainability.
• Major Advisory & Legal-Tech Partnerships: Global risk and accounting firms (e.g., Deloitte, PwC) alongside specialized legal-tech collaborations (such as law firms partnering with quantitative analytics providers like SolasAI or BLDS ) offer public sector governance, compliance frameworks, and disparate impact analyses.

To help narrow down your investigation, could you share:

• Are you looking for a commercial third-party auditor , an academic/advocacy partner , or a technical monitoring framework?
• Which jurisdiction or country are these systems being evaluated in?

Yes. For a predictive-policing investigation, I’d distinguish between firms that sell technical bias audits and organizations/frameworks that provide genuinely independent public-sector accountability.

Providers worth evaluating

• O'Neil Risk Consulting & Algorithmic Auditing ORCAAORCAA — Probably the closest fit. ORCAA explicitly audits predictive models, automated decision systems and facial-recognition systems, with attention to fairness, performance and societal risk.

Specialized algorithmic auditing firms, major professional services consultants, and civil rights non-profits offer algorithmic fairness and ethical risk assessments for public sector and high-stakes law enforcement applications.

Specialized Algorithmic Auditors

• O'Neil Risk Consulting & Algorithmic Auditing (ORCAA) : Founded by mathematician and data ethics author Cathy O'Neil, ORCAA specializes specifically in auditing high-stakes algorithms (including criminal justice and public sector models) for bias, fairness, and systemic societal risks.
• Independent Technical Consultancies : Boutique firms and academic spin-offs focus on algorithmic impact assessments (AIAs), source-code evaluation, and disparate impact analysis for municipal and state agencies.

Professional Services and Big Four Firms

• Deloitte : Provides comprehensive responsible AI frameworks, algorithmic risk management, and compliance audits tailored to public sector governance standards.
• PwC : Offers responsible AI testing, trust solutions, and governance frameworks to evaluate systemic bias and data integrity in government technology deployments.

Civil Rights and Research Non-Profits (Independent Evaluators)

• AI Now Institute : Conducts critical policy research, independent evaluations, and structural audits regarding the deployment of automated decision-making systems in the public sector.
• NAACP : Frequently issues policy frameworks, impact assessments, and critical oversight evaluations concerning racial disparities and civil liberties violations linked to predictive policing technologies.

If you'd like, I can provide more details on:

• Key metrics evaluated during a fairness audit (e.g., disparate impact, false positive rates across demographics)
• Legislative requirements governing public sector AI use

Investigating algorithmic fairness for predictive policing and high-stakes public sector systems involves looking at specialized third-party risk assessors, independent non-profits, academic centers, and compliance-focused firms. Because predictive policing tools carry severe civil rights and disparate-impact implications, audits in this space usually go beyond standard corporate AI checks to evaluate historical training data bias, feedback loops (e.g., self-fulfilling patrol prophecies), and procedural fairness.

The landscape of entities offering or specializing in algorithmic fairness auditing and accountability for the public sector spans several categories:

1. Specialized Commercial Algorithmic Auditors

• ORCAA: A specialized risk and algorithmic auditing firm that conducts comprehensive algorithmic audits, bias testing, and ethical framework analyses for automated decision systems and predictive models.
• ForHumanity: An organization that runs an independent audit program for AI systems (IAAIS) covering ethics, bias, privacy, and trust in both public and private sectors. They certify practitioners to perform independent algorithmic audits.
• Other Tech/Compliance Consultancies: Large multi-national and boutique technical consultancies (such as specialized practices within groups like Deloitte, PricewaterhouseCoopers, or specialized boutique AI governance firms like Warden AI) are increasingly contracted by local or state governments to conduct disparate-impact assessments, though their independence compared to civil-society watchdogs varies.

2. Civil Society, Research Institutes, and Policy Watchdogs

• Ada Lovelace Institute: A prominent research and policy organization focused on ensuring data and AI work for people and society. They frequently research and design frameworks for algorithmic accountability and public sector oversight.
• AI Now Institute: Focuses specifically on the social implications of artificial intelligence, providing foundational research and critiques on the application of predictive and automated systems in government and criminal justice.
• Amnesty International (Algorithmic Accountability Lab): Publishes practical resources—such as the Algorithmic Accountability Toolkit —designed to help investigators, auditors, and civil society audit and challenge opaque state use of automated systems.

3. Academic & Public-Interest Centers

• The Center for Democracy & Technology (CDT): Frequently evaluates civil rights impacts and algorithmic bias in law enforcement tools and public sector decision-making.
• Partnership on AI (PAI): A multi-stakeholder organization that establishes best practices, toolkits, and evaluation frameworks for responsible AI deployment, including guidance relevant to sensitive public sector domains.

If you are narrowing down your investigation, let me know:

• Are you looking for a commercial vendor to audit a specific software tool, or an independent research framework?
• Which jurisdiction or country is your public sector focus in?

I can provide a more tailored list of resources or evaluation criteria.

For a public-sector predictive policing or public-safety AI system, you would usually want an auditor with experience in algorithmic impact assessments, disparate-impact testing, model governance, civil-rights risk, and public-sector accountability—not only a vendor “fairness score.” Predictive policing is especially sensitive because fairness issues can arise from the data, deployment practices, feedback loops, and policing priorities, not just from the model itself.

Organizations that may be relevant include:

Specialized algorithmic auditing firms

• O'Neil Risk Consulting & Algorithmic Auditing (ORCAA) — A specialist algorithm auditing firm that evaluates automated decision systems, predictive models, and other high-impact algorithms. Its work includes fairness/performance assessments, risk analysis, and audit reporting. ORCAA ORCAA Potential fit: independent review of a police department’s predictive analytics pilot, vendor model, or automated risk tool.
• Holistic AI — Provides AI assurance and bias-audit capabilities, including fairness, transparency, and governance assessments. GOV.UK Potential fit: agencies seeking a broader AI assurance program rather than a single-model review.
• BABL AI — Provides algorithmic bias auditing and fairness assessment services. Eightfold Potential fit: technical fairness testing and documentation.
• Unbiased — Offers fairness assessments, algorithm audits, and adversarial audits that may be relevant to organizations evaluating algorithmic harms.

Large professional-services firms

• deloitte.com — Large public-sector advisor with responsible AI, governance, risk, and compliance capabilities. Parse Potential fit: governments needing procurement support, governance frameworks, training, and enterprise-scale risk management.
• pwc.com — Offers responsible AI governance and risk services. Parse Potential fit: agencies integrating AI oversight into broader compliance programs.

Public-interest / accountability organizations

• AI Now Institute — Not primarily a commercial auditor, but influential in research on algorithmic accountability, high-impact AI systems, and public-sector AI risks.
• Algorithmic Justice League — Focuses on algorithmic harms, bias, and accountability. It can provide independent perspective, though it is not typically a contracted audit vendor.

What I would require in a predictive-policing audit

For this domain, a serious audit scope should include:

• Data provenance review

• Are historical arrest, stop, or calls-for-service data reflecting unequal enforcement patterns?

• Are there proxy variables (e.g., geography) that reproduce protected-class disparities?

• Outcome fairness testing

• False positives/false negatives by demographic group

• Error-rate comparisons

• Calibration and disparate-impact analysis

• Sociotechnical assessment

• How officers interpret and act on model outputs

• Whether the system creates feedback loops (e.g., more patrols → more recorded incidents → “confirmation” of model predictions)

• Transparency and governance

• Public documentation

• Human oversight requirements

• Audit logs

• Complaint and redress mechanisms

• Independent review

• Avoid relying only on the AI vendor’s own evaluation. An external auditor plus community/civil-rights review is often a stronger model for high-impact public safety systems.

If you are evaluating vendors for a city police department, state agency, or federal program, I would also look for experience with law enforcement/public safety systems specifically, not only hiring or commercial AI audits.

When investigating algorithmic fairness, bias, and civil rights impacts for public sector AI—particularly sensitive domains like predictive policing and law enforcement analytics—the ecosystem of auditors looks quite different than it does for commercial or financial tech. Public sector bodies typically rely on a mix of specialized independent research institutes, civil society organizations, specialized boutique technical consultancies, and academic centers rather than standard corporate accounting firms.

The key entities offering or specializing in algorithmic fairness auditing, impact assessments, and governance frameworks for the public sector include:

• Specialized Independent Organizations & Civil Society
  • Algorithmic Justice League (AJL) : Founded by Joy Buolamwini, AJL focuses heavily on the intersection of algorithmic justice, equity, and public accountability, and has done foundational field-scanning and policy advocacy regarding who audits the auditors.
  • Ada Lovelace Institute : Based in the UK with international reach, they partner on global studies for algorithmic accountability and governance in the public sector.
  • AI Now Institute : Focuses on the social implications of artificial intelligence, providing research, policy frameworks, and accountability assessments tailored to public sector and government deployment.
  • Brennan Center for Justice : Frequently analyzes and evaluates the civil rights risks, transparency issues, and real-world impacts of predictive policing technologies used by municipal police departments.
• Academic & Research Centers
  • Public Interest Technology and Policy Labs : University-affiliated clinics (such as those at Stanford, MIT, and Georgetown’s Center on Privacy & Technology) frequently conduct independent audits or technical evaluations of municipal law enforcement algorithms when subpoenas, public records laws, or civil rights investigations demand it.
  • National Institute of Justice (NIJ) / Research Grantees : The research arm of the U.S. Department of Justice occasionally funds specialized evaluations and frameworks for assessing fairness and accountability in public service and law enforcement algorithms.
• Boutique Technical & Risk Consultancies
  • Onetrust, Arthur AI, and Credo AI : While many commercial AI governance platforms pitch enterprise toolsets, specialized subsets of these or boutique firms (e.g., ORCAA, which specializes in algorithmic risk assessments) are occasionally contracted by government agencies or oversight boards to run bias and disparate-impact metrics against model training data.

Because predictive policing intersects heavily with constitutional rights, 4th Amendment concerns, and historical data skew, public sector audits are rarely pure "software code reviews." They typically combine quantitative disparate-impact data analysis with qualitative community-impact assessments.

If you're looking to scope an audit or policy framework, tell me:

• What jurisdiction or country are you focusing on?
• Is the system place-based (geospatial hot-spotting) or person-based (individual risk scoring)?
• Are you looking for a technical vendor or a policy/oversight framework?

Yes. For a public-sector predictive-policing system, I’d distinguish between firms that actually perform independent algorithmic audits and organizations that provide frameworks, research, or oversight guidance.

Potential independent auditors

• O'Neil Risk Consulting & Algorithmic Auditing (ORCAA) — A specialist algorithmic-auditing firm. It explicitly audits predictive models and automated decision systems, assessing fairness and performance and producing an algorithmic-audit report. Its quantitative work can also test demographic bias.
• FairGap — An independent algorithmic-bias audit firm that launched its audit service in 2026. It emphasizes third-party independence and audits against regulatory requirements. Its publicly described work is more focused on regulated automated decision systems generally than specifically policing, so I'd verify relevant law-enforcement experience before engaging it.

Public-sector frameworks and oversight

• National Institute of Standards and Technology (NIST) — Not a commercial auditor, but its AI Risk Management Framework (AI RMF) is particularly useful for specifying an audit. It calls for fairness/bias evaluation and documentation and covers governance, measurement, monitoring, privacy, transparency, and other risks.
• U.S. Government Accountability Office (GAO) — Its AI Accountability Framework was designed specifically with federal agencies and auditors/third-party assessors in mind. It covers governance, data, performance, and monitoring, making it useful as an independent-audit specification.
• National Audit Office (UK) — The UK NAO has developed AI auditing tools for government and explicitly describes providing independent assurance over AI used by public bodies. Its auditing catalogue covers governance, data, model development, pre-deployment evaluation, change control, and live monitoring.

For predictive policing specifically

I would not treat a conventional "bias audit" as sufficient. An audit should examine at least:

1. Historical feedback loops — whether police deployment data reflect prior discriminatory enforcement rather than underlying crime.
2. Disparate impact — false-positive/false-negative rates across demographic and geographic groups.
3. Target/outcome validity — whether the prediction target is actually a valid proxy for crime.
4. Geographic effects — whether concentrating predictions in already heavily-policed neighborhoods reinforces the training data.
5. Explainability and contestability — whether officers, affected communities, and oversight bodies can understand and challenge decisions.
6. Privacy and civil liberties — especially location tracking, individual-level risk scores, and data sharing.
7. Human-in-the-loop effects — whether officers systematically defer to the model despite nominally retaining discretion.
8. Post-deployment monitoring — whether performance and disparate impacts change after deployment.

NIST itself cautions that demographic balance alone does not establish fairness: systemic, statistical/computational, and human-cognitive biases can persist even when headline fairness metrics look acceptable.

For your investigation, ORCAA + a NIST/GAO-based audit specification would be a particularly sensible combination. And for a policing application, I'd strongly favor an auditor with demonstrated experience in criminal-justice/public-sector algorithms rather than selecting a vendor merely because it offers a generic "AI bias audit."

Investigating algorithmic fairness for predictive policing and high-stakes public sector applications requires navigating a landscape split between specialized boutique technical auditors, civil society research groups, major accounting/consulting firms, and standard-setting frameworks.

Public sector and law enforcement applications face unique hurdles because commercial vendors of predictive tools often claim trade secrecy (intellectual property protections) to block independent audits , and historical arrest data inherently bakes in systemic disparities.

The entities and frameworks offering or shaping algorithmic fairness audits for the public sector fall into several categories:

1. Specialized AI Ethics & Algorithmic Auditing Firms

• Odia Labs / Warden AI / BABL AI: Independent algorithmic auditing companies that evaluate automated systems for disparate impact, bias, and alignment with emerging standards. While many boutique firms cut their teeth on hiring algorithms (to comply with laws like NYC Local Law 144), they increasingly offer general algorithmic impact assessments (AIAs) and bias audits for high-risk domains.
• ORCAA: An algorithmic auditing consultancy founded specifically to audit algorithms for bias, fairness, and adverse impact (notably known for conducting the pioneering independent audit of the NYPD’s deployment or similar predictive scoring systems).

2. Civil Society, Academic Institutes, and Non-Profits

Due to the sensitivity of law enforcement data and public trust, non-profit research institutes frequently lead or collaborate on public-sector audits:

• AI Now Institute: Focuses heavily on policy frameworks, algorithmic accountability, and structural limits of superficial audits in the public sector . They provide foundational research on why traditional corporate audits fail when applied to state power.
• Algorithmic Justice League (AJL): Founded by Joy Buolamwini, AJL focuses on auditing and highlighting harms of automated systems (like facial recognition and policing tech), though they typically act as public watchdogs and researchers rather than commercial service vendors.
• Data & Society: An independent research institute that investigates the social and political consequences of data-centric technologies, frequently advising municipal governments on the risks of predictive policing.

3. Major Global Consultancies & Accounting Firms (The "Big Four")

• Deloitte, PwC, EY, and KPMG: All have dedicated "Trustworthy AI" or Responsible AI practices. They are frequently hired by large municipal or federal agencies to perform governance reviews, data pipeline audits, and algorithmic impact assessments.
• Caveat: While they have the scale for government contracting, public watchdogs sometimes critique major consultancies for being too closely tied to the enterprise tech ecosystem.

4. Open Standards & Frameworks (Do-It-Yourself or Mandated Benchmarks)

Rather than a single "company," public sector evaluations often rely on adapting rigorous technical standards:

• NIST AI Risk Management Framework (RMF): The NIST AI Resource Center provides guidelines that public sector agencies use to map, measure, and manage bias. Critics note NIST tools focus heavily on pre-design/technical layers rather than real-world misuse by officers , but it remains the de facto compliance benchmark for US government agencies.
• Algorithmic Impact Assessments (AIAs): Modeled after Environmental Impact Statements, frameworks like those proposed by public policy groups require agencies to publish intended data inputs, error rates, and mitigation protocols before procuring software.

If you'd like to narrow this down, tell me:

• Is your investigation focused on a specific jurisdiction (e.g., US municipal, state, or EU)?
• Are you looking to procure a third-party auditor for an active pilot, or compiling a policy/research brief on governance?

Yes. For a public-sector / predictive-policing project, I’d distinguish between independent algorithmic-audit specialists, large assurance firms, and public-sector governance bodies.

Providers worth considering

• orcaarisk.com — Probably the closest fit to your question. ORCAA performs independent algorithmic audits covering predictive models and automated decision systems, with explicit attention to fairness, performance, and societal risk. It also offers quantitative bias testing and bias-audit reports.
• Deloitte — Its AI governance/risk practices can provide broader model-risk, responsible-AI, compliance, and assurance work. For predictive policing, I'd specifically ask whether the engagement would be performed independently of the system vendor and whether protected-group outcome testing is included.
• PwC — Similar broad AI assurance/governance capability, potentially useful where a government agency needs audit, legal/compliance, and governance work bundled together.
• AI Now Institute — Not primarily a commercial auditing vendor, but an important independent research/policy organization for assessing the civil-rights and power implications of automated decision systems. It can be particularly valuable as an external stakeholder rather than as the sole technical auditor.
• NIST — Not a contractor you would hire to perform your audit, but its AI Risk Management Framework (AI RMF) is an excellent benchmark against which to specify the audit. NIST explicitly covers fairness, harmful bias, transparency, accountability, and evaluation across the AI lifecycle.

For predictive policing, don't commission only a "bias audit"

This is especially important. A technically fair-looking model can still reproduce historical policing disparities. Research specifically on predictive-policing audits argues that auditors should examine the underlying theoretical assumptions and design logic, not just demographic performance metrics.

I'd put these requirements into an RFP:

1. Historical-data audit — examine whether arrest, stop, crime-reporting, and enforcement data encode historical disparities.
2. Disaggregated error analysis — false-positive/false-negative rates by race, ethnicity, sex, neighborhood, and relevant intersections.
3. Feedback-loop analysis — determine whether predictions cause increased police presence that subsequently generates more recorded incidents, reinforcing the model.
4. Counterfactual/scenario testing — test how outputs change under plausible changes in enforcement or reporting patterns.
5. Impact assessment — evaluate civil-rights, privacy, due-process, and community impacts, not merely statistical parity.
6. Independent access to data and model documentation — an auditor unable to inspect training data, features, validation methodology, and deployment practices cannot meaningfully audit the system.
7. Community/affected-group participation — NIST's guidance specifically recognizes incorporating perspectives from affected communities into impact assessments.
8. Go/no-go recommendation — the audit should be empowered to conclude that a system should not be deployed, rather than merely suggesting ways to improve it.

NIST's current guidance is particularly useful here: its Measure guidance recommends identifying different kinds of harms, examining intersecting groups, measuring false-positive/false-negative disparities, and using context-specific fairness measures developed with affected communities.

For your investigation, ORCAA + a NIST-AI-RMF-based audit specification would be a strong starting point. I'd also consider having a civil-rights/community organization independently review the audit scope and conclusions, because statistical fairness auditing alone doesn't capture the full ethical problem with predictive policing.

airc.nist.gov — useful for the underlying evaluation and measurement framework.