Maryland HB 820 (2025 Md. Laws ch. 747; Ins. 15-10B-05.1, 15-10A-06)

Binding law — in force

AI utilization-review determinations must rest on the enrollee's own clinical history and circumstances, not solely a group dataset (Maryland HB 820)

Md. Code, Ins. 15-10B-05.1(c)(1) · official text · In force: applies since 1 Oct 2025 · Maryland (US-MD)

From 2025-10-01, an entity subject to Ins. 15-10B-05.1 must ensure that the artificial intelligence, algorithm or other software tool bases its determinations on the enrollee's medical or other clinical history, individual clinical circumstances as presented by a requesting provider, or other relevant clinical information in the enrollee's record ((c)(1)), and does not base its determinations solely on a group dataset ((c)(2)). Detect utilization-review model calls built without the member's clinical record or the provider's submission.

Who it applies to

• Duty falls on: insurer, organization
• Sectors: insurance, healthcare
• Carriers (insurers, nonprofit health service plans, HMOs, dental plan organizations and other persons providing State-regulated health benefit plans) that use, or contract with or work through an entity that uses, an artificial intelligence, algorithm or other software tool for utilization review, and pharmacy benefits managers and private review agents that contract with a carrier to provide utilization review and use such a tool, for Maryland enrollees. In force 2025-10-01.
• Whether it applies depends on facts outside the code; a person has to decide.

features = {'cpt': req.cpt, 'icd10': req.icd10, 'cohort_approval_rate': cohort_stats(req.icd10)}
reply = client.messages.create(model=MODEL, max_tokens=400,
    messages=[{'role': 'user', 'content': f'Prior authorization request: {features}'}])

record = member_clinical_record(req.member_id, fields=PA_CLINICAL_FIELDS)   # history, conditions, meds
submission = provider_clinical_submission(req.id)                            # notes, letter of medical necessity
if not record or not submission:
    return review_queue.enqueue(req.id, reason='individual clinical information missing')
reply = client.messages.create(model=MODEL, max_tokens=400, messages=[{'role': 'user', 'content':
    render('pa_review.txt', request=req, clinical_history=record, provider_submission=submission)}])
determinations.save(req.id, inputs={'clinical_history': record.ids, 'submission': submission.ids})

Related incidents

No guardrail sits on this exact control; these incidents are cited by guardrails on related controls.

• Meta's automated moderation over-enforced Arabic and under-enforced Hebrew content (BSR due diligence) (2021-05; disclosed by the operator). An independent human rights due diligence by BSR, commissioned and published by Meta on September 22, 2022, found that during the May 2021 Israel-Palestine escalation Arabic content saw greater over-enforcement per user than Hebrew content and Hebrew content greater under-enforcement. BSR attributes this in part to Meta having an Arabic hostile-speech classifier but no Hebrew one, and to Arabic classifiers likely being less accurate for Palestinian Arabic. BSR found no intentional bias but 'various instances of unintentional bias' with different impacts on Palestinian and Arabic-speaking users. Meta committed to implement 10 of BSR's 21 recommendations and said it had since launched a Hebrew hostile-speech classifier. Source: Meta (operator response, 2022-09-22) · evidence grade: primary · cited by Check AI ranking, pricing, moderation, and ad targeting for disparities when features can stand in for protected traits, and screen every served language
• Toxicity classifiers rate African American English as more offensive (2019; confirmed). University of Washington researchers reported at ACL 2019 that tweets in African American English and tweets by self-identified African Americans were up to two times more likely to be labelled offensive by hate-speech models trained on widely used datasets, and that Jigsaw's public Perspective API showed similar racial bias, rating AAE phrases as more toxic than non-AAE equivalents. Source: Sap et al., 'The Risk of Racial Bias in Hate Speech Detection', ACL 2019 (original researchers) · evidence grade: primary · cited by Check AI ranking, pricing, moderation, and ad targeting for disparities when features can stand in for protected traits, and screen every served language
• Ride-hailing fares higher in Chicago neighborhoods with more non-white residents (researcher audit) (2018-11; alleged (not proven)). George Washington University researchers analysing Chicago's public data on more than 100 million ride-hailing trips from November 2018 to September 2019 report that trips in neighborhoods with larger non-white populations, higher poverty, younger residents and more college-educated residents were significantly associated with higher fares. They attribute this to pricing algorithms learning from demand, supply and trip duration. The finding is an observational association from census-tract data; the pricing models themselves were not examined. Source: Pandey & Caliskan, 'Disparate Impact of Artificial Intelligence Bias in Ridehailing Economy's Price Discrimination Algorithms', AIES 2021 (original researchers) · evidence grade: primary · cited by Check AI ranking, pricing, moderation, and ad targeting for disparities when features can stand in for protected traits, and screen every served language
• Google ads suggesting arrest records served more often for Black-identifying names (2012; confirmed). Harvard researcher Latanya Sweeney searched 2,184 racially associated full names on google.com and reuters.com (a Google AdSense host) from September 24 to October 23, 2012 and found ads suggestive of an arrest record appeared more often for Black-identifying first names; on reuters.com a Black-identifying name was 25% more likely to get such an ad (statistically significant). Ads appeared regardless of whether the name had an arrest record in the advertiser's database. The paper does not determine whether the advertiser's templates or Google's click-based ad optimization caused the pattern; the advertiser, Instant Checkmate, told the author it gave Google the same ad text for groups of last names. Source: Sweeney, 'Discrimination in Online Ad Delivery' (original researcher, 2013-01-28) · evidence grade: primary · cited by Check AI ranking, pricing, moderation, and ad targeting for disparities when features can stand in for protected traits, and screen every served language

Rule id md-hb820.individual-clinical-data-basis · review status: primary source derived

Binding law — in force

An AI, algorithm or other software tool may not deny, delay or modify care or replace the physician's role in adverse decisions (Maryland HB 820)

Md. Code, Ins. 15-10B-05.1(d) · official text · In force: applies since 1 Oct 2025 · Maryland (US-MD)

From 2025-10-01, an artificial intelligence, algorithm or other software tool may not deny, delay or modify health care services (Ins. 15-10B-05.1(d)), and the entity must ensure that it does not replace the role of a health care provider in the determination process under 15-10B-07 ((c)(4)), under which all adverse decisions are made by a licensed physician, or a panel of health care service reviewers with at least one physician, board certified or eligible in the same specialty and knowledgeable about the service (dentists for dental services; 15-10B-07(a)), not compensated in a way that deters appropriate care ((b)); and that the tool does not directly or indirectly cause harm to an enrollee ((c)(11)). Detect automated output that sets a denial, delay or modification without a physician's decision.

Who it applies to

• Duty falls on: insurer, organization
• Sectors: insurance, healthcare
• Carriers (insurers, nonprofit health service plans, HMOs, dental plan organizations and other persons providing State-regulated health benefit plans) that use, or contract with or work through an entity that uses, an artificial intelligence, algorithm or other software tool for utilization review, and pharmacy benefits managers and private review agents that contract with a carrier to provide utilization review and use such a tool, for Maryland enrollees. In force 2025-10-01.
• Whether it applies depends on facts outside the code; a person has to decide.

result = client.chat.completions.create(model=MODEL, messages=build_pa_prompt(request)).choices[0].message.content
if json.loads(result)['decision'] == 'deny':
    prior_auth.update(request.id, status='denied')
    send_denial_letter(request)

result = json.loads(client.chat.completions.create(
    model=MODEL, messages=build_pa_prompt(request, record=member_clinical_record(request))).choices[0].message.content)
if result['decision'] == 'approve' and AUTO_APPROVE_ALLOWED:
    prior_auth.update(request.id, status='approved', ai_assisted=True)
else:                                   # any non-approval goes to a clinician
    review_queue.enqueue(request.id, queue='pending_clinical_review',
                         specialty=request.specialty, ai_recommendation=result)

@app.post('/reviews/{case_id}/decision')
def record_clinical_decision(case_id: str, body: Decision, reviewer=Depends(licensed_clinical_reviewer)):
    decision = clinical_decisions.create(case_id=case_id, reviewer_id=reviewer.id, licence=reviewer.licence,
                                         specialty=reviewer.specialty, documents_reviewed=body.documents,
                                         outcome=body.outcome, rationale=body.rationale)
    if body.outcome in ('denied', 'downgraded'):
        send_adverse_determination(case_id, decision=decision, signed_by=reviewer)

Related incidents

No guardrail sits on this exact control; these incidents are cited by guardrails on related controls.

• UnitedHealth nH Predict claim-denial litigation (2023-11; alleged (not proven)). A class action filed in November 2023 alleges that UnitedHealth's nH Predict model had a 90% error rate, measured by denials reversed on appeal, while only about 0.2% of members appealed. UnitedHealth disputes the allegations; the litigation is ongoing. Source: STAT News · evidence grade: primary · cited by Monitor how often adverse AI decisions are reversed, and suspend models that are usually wrong
• Cigna PXDX batch claim denials (reported) (2022; alleged (not proven)). ProPublica, citing internal Cigna records, reported that Cigna's PXDX system was used to reject more than 300,000 claims over two months in 2022, with physicians spending an average of 1.2 seconds on each. Cigna disputes the reporting; related lawsuits are ongoing. Source: ProPublica / The Capitol Forum · evidence grade: press of record · cited by Make human review of adverse AI decisions substantive, not nominal

Rule id md-hb820.no-ai-denial-provider-decides · review status: primary source derived

Binding law — in force

Patient data used by AI in utilization review may not be used beyond its intended and stated purpose (Maryland HB 820)

Md. Code, Ins. 15-10B-05.1(c)(10) · official text · In force: applies since 1 Oct 2025 · Maryland (US-MD)

From 2025-10-01, the entity must ensure that patient data is not used beyond its intended and stated purpose, consistent with HIPAA, as applicable (Ins. 15-10B-05.1(c)(10)). Detect utilization-review patient data flowing to model training or fine-tuning, marketing lists or product analytics.

Who it applies to

• Duty falls on: insurer, organization
• Sectors: insurance, healthcare
• Carriers (insurers, nonprofit health service plans, HMOs, dental plan organizations and other persons providing State-regulated health benefit plans) that use, or contract with or work through an entity that uses, an artificial intelligence, algorithm or other software tool for utilization review, and pharmacy benefits managers and private review agents that contract with a carrier to provide utilization review and use such a tool, for Maryland enrollees. In force 2025-10-01.
• Whether it applies depends on facts outside the code; a person has to decide.

client = OpenAI()
resp = client.chat.completions.create(model='gpt-4o', messages=[
    {'role': 'user', 'content': f'Summarize: {patient.clinical_note}'}])

VENDORS = load_yaml('vendors/ai_vendors.yaml')   # baa_signed, zero_data_retention per endpoint

def phi_client(name: str) -> tuple[AzureOpenAI, str]:
    v = VENDORS[name]
    if not (v['baa_signed'] and v['zero_data_retention']):
        raise PermissionError(f'{name} is not cleared for PHI')
    client = AzureOpenAI(azure_endpoint=v['endpoint'], api_key=os.environ['AZURE_OPENAI_KEY'],
                         api_version=v['api_version'])
    return client, v['deployment']

client, deployment = phi_client('azure-openai-hipaa')
resp = client.chat.completions.create(model=deployment, messages=[
    {'role': 'user', 'content': f'Summarize: {patient.clinical_note}'}])

Rule id md-hb820.patient-data-purpose-limit · review status: primary source derived

Binding law — in force

AI in utilization review must not unfairly discriminate, must be fairly applied, and must be reviewed and revised at least quarterly (Maryland HB 820)

Md. Code, Ins. 15-10B-05.1(c)(5) · official text · In force: applies since 1 Oct 2025 · Maryland (US-MD)

From 2025-10-01, the entity must ensure that the use of the artificial intelligence, algorithm or other software tool does not result in unfair discrimination (Ins. 15-10B-05.1(c)(5)), that it is fairly and equitably applied, including under applicable HHS regulations and guidance ((c)(6)), and that its performance, use and outcomes are reviewed and revised, if necessary and at least on a quarterly basis, to maximize accuracy and reliability ((c)(9)). Detect the absence of a quarterly accuracy, outcome and disparity review.

Who it applies to

• Duty falls on: insurer, organization
• Sectors: insurance, healthcare
• Carriers (insurers, nonprofit health service plans, HMOs, dental plan organizations and other persons providing State-regulated health benefit plans) that use, or contract with or work through an entity that uses, an artificial intelligence, algorithm or other software tool for utilization review, and pharmacy benefits managers and private review agents that contract with a carrier to provide utilization review and use such a tool, for Maryland enrollees. In force 2025-10-01.
• Whether it applies depends on facts outside the code; a person has to decide.

clf = LogisticRegression(max_iter=1000).fit(X_train, y_train)
joblib.dump(clf, 'models/credit_v4.joblib')

from fairlearn.metrics import MetricFrame, selection_rate, demographic_parity_difference
from sklearn.metrics import accuracy_score

clf = LogisticRegression(max_iter=1000).fit(X_train, y_train)
y_pred = clf.predict(X_test)
mf = MetricFrame(metrics={'accuracy': accuracy_score, 'selection_rate': selection_rate},
                 y_true=y_test, y_pred=y_pred, sensitive_features=A_test)
dpd = demographic_parity_difference(y_test, y_pred, sensitive_features=A_test)
write_validation_report('credit_v4', mf.by_group, dpd)
if dpd > MAX_DPD:
    raise SystemExit('bias gate failed: owner review required before release')
joblib.dump(clf, 'models/credit_v4.joblib')

Standards that recommend the same control

• AI should avoid unfair bias and be assessed for discrimination (EU ALTAI) (EU ALTAI · ALTAI / EU Ethics Guidelines — Requirement 5 (Diversity, Non-discrimination and Fairness))
• AI models in financial decisions should be regularly validated for accuracy and bias (MAS FEAT) (MAS FEAT Principles · MAS FEAT Principles — Fairness (Accuracy and Bias), Principles 3-4)
• Insurers adopting the NAIC model guidance should validate and bias-test AI and predictive models (NAIC AI Model Bulletin · NAIC Model Bulletin, adoption statement (title page))

Related incidents

No guardrail sits on this exact control; these incidents are cited by guardrails on related controls.

• Meta's automated moderation over-enforced Arabic and under-enforced Hebrew content (BSR due diligence) (2021-05; disclosed by the operator). An independent human rights due diligence by BSR, commissioned and published by Meta on September 22, 2022, found that during the May 2021 Israel-Palestine escalation Arabic content saw greater over-enforcement per user than Hebrew content and Hebrew content greater under-enforcement. BSR attributes this in part to Meta having an Arabic hostile-speech classifier but no Hebrew one, and to Arabic classifiers likely being less accurate for Palestinian Arabic. BSR found no intentional bias but 'various instances of unintentional bias' with different impacts on Palestinian and Arabic-speaking users. Meta committed to implement 10 of BSR's 21 recommendations and said it had since launched a Hebrew hostile-speech classifier. Source: Meta (operator response, 2022-09-22) · evidence grade: primary · cited by Check AI ranking, pricing, moderation, and ad targeting for disparities when features can stand in for protected traits, and screen every served language
• Toxicity classifiers rate African American English as more offensive (2019; confirmed). University of Washington researchers reported at ACL 2019 that tweets in African American English and tweets by self-identified African Americans were up to two times more likely to be labelled offensive by hate-speech models trained on widely used datasets, and that Jigsaw's public Perspective API showed similar racial bias, rating AAE phrases as more toxic than non-AAE equivalents. Source: Sap et al., 'The Risk of Racial Bias in Hate Speech Detection', ACL 2019 (original researchers) · evidence grade: primary · cited by Check AI ranking, pricing, moderation, and ad targeting for disparities when features can stand in for protected traits, and screen every served language
• Ride-hailing fares higher in Chicago neighborhoods with more non-white residents (researcher audit) (2018-11; alleged (not proven)). George Washington University researchers analysing Chicago's public data on more than 100 million ride-hailing trips from November 2018 to September 2019 report that trips in neighborhoods with larger non-white populations, higher poverty, younger residents and more college-educated residents were significantly associated with higher fares. They attribute this to pricing algorithms learning from demand, supply and trip duration. The finding is an observational association from census-tract data; the pricing models themselves were not examined. Source: Pandey & Caliskan, 'Disparate Impact of Artificial Intelligence Bias in Ridehailing Economy's Price Discrimination Algorithms', AIES 2021 (original researchers) · evidence grade: primary · cited by Check AI ranking, pricing, moderation, and ad targeting for disparities when features can stand in for protected traits, and screen every served language
• Google ads suggesting arrest records served more often for Black-identifying names (2012; confirmed). Harvard researcher Latanya Sweeney searched 2,184 racially associated full names on google.com and reuters.com (a Google AdSense host) from September 24 to October 23, 2012 and found ads suggestive of an arrest record appeared more often for Black-identifying first names; on reuters.com a Black-identifying name was 25% more likely to get such an ad (statistically significant). Ads appeared regardless of whether the name had an arrest record in the advertiser's database. The paper does not determine whether the advertiser's templates or Google's click-based ad optimization caused the pattern; the advertiser, Instant Checkmate, told the author it gave Google the same ad text for groups of last names. Source: Sweeney, 'Discrimination in Online Ad Delivery' (original researcher, 2013-01-28) · evidence grade: primary · cited by Check AI ranking, pricing, moderation, and ad targeting for disparities when features can stand in for protected traits, and screen every served language

Rule id md-hb820.quarterly-accuracy-and-fairness-review · review status: primary source derived

Binding law — in force

Carriers' quarterly reports must state whether an AI, algorithm or other software tool was used in making each adverse decision (Maryland HB 820)

Md. Code, Ins. 15-10A-06(a)(1) · official text · In force: applies since 1 Oct 2025 · Maryland (US-MD)

From 2025-10-01, each carrier's quarterly report to the Commissioner (on the Commissioner's form, aggregated by zip code as required) must describe the number of adverse decisions it issued under 15-10A-02(f), whether each involved a prior authorization or step therapy protocol, the type of service, and whether an artificial intelligence, algorithm or other software tool was used in making the adverse decision (Ins. 15-10A-06(a)(1)(iii)6, as added by Chapter 747). Detect adverse decisions recorded from a model or scoring call with no field recording AI use, and no quarterly report job.

Who it applies to

• Duty falls on: insurer
• Sectors: insurance, healthcare
• Each carrier subject to subtitle 10A of the Insurance Article (internal grievance process) that issues adverse decisions for Maryland members; the quarterly report states, for adverse decisions issued under 15-10A-02(f), whether an artificial intelligence, algorithm or other software tool was used. In force 2025-10-01.
• Whether it applies depends on facts outside the code; a person has to decide.

def decide(app):
    resp = client.chat.completions.create(model=MODEL, messages=eligibility_prompt(app))
    applications.set_status(app.id, parse_decision(resp))

log = structlog.get_logger()   # TimeStamper processor adds the timestamp

def decide(app):
    decision_id = str(uuid.uuid4())
    resp = client.chat.completions.create(model=MODEL, messages=eligibility_prompt(app))
    status = parse_decision(resp)
    log.info('ai_decision', decision_id=decision_id, model=resp.model,
             input_ref=f'applications/{app.id}', output=status, operator=current_operator())
    applications.set_status(app.id, status, decision_id=decision_id)

Related incidents

No guardrail sits on this exact control; these incidents are cited by guardrails on related controls.

• Court compelled discovery on how nH Predict works (2026-03-09; confirmed). A federal magistrate judge in the District of Minnesota ordered UnitedHealth to produce documents on how nH Predict works, including whether it was designed to supplant physician decision-making. Plaintiffs needed litigation discovery to learn how the model was designed and used. Source: U.S. District Court, D. Minn. (Order, Doc. 162) · evidence grade: primary · cited by Record enough at decision time to reproduce and explain every consequential AI decision

Rule id md-hb820.quarterly-report-ai-use-in-adverse-decisions · review status: primary source derived

Binding law — in force

The utilization plan must describe the AI tool's use and oversight, and the tool must be open to inspection (Maryland HB 820)

Md. Code, Ins. 15-10B-05.1(c)(8) · official text · In force: applies since 1 Oct 2025 · Maryland (US-MD)

From 2025-10-01, the entity must ensure that written policies and procedures, including how the artificial intelligence, algorithm or other software tool will be used and what oversight will be provided, are included in the utilization plan submitted under 15-10B-05 (Ins. 15-10B-05.1(c)(8)), that the criteria and guidelines for using the tool comply with the title ((c)(3)), and that the tool is open to inspection for audit or compliance reviews by the Commissioner ((c)(7)). Detect the absence of an AI-use and oversight policy in the utilization plan and an inspection record.

Who it applies to

• Duty falls on: insurer, organization
• Sectors: insurance, healthcare
• Carriers (insurers, nonprofit health service plans, HMOs, dental plan organizations and other persons providing State-regulated health benefit plans) that use, or contract with or work through an entity that uses, an artificial intelligence, algorithm or other software tool for utilization review, and pharmacy benefits managers and private review agents that contract with a carrier to provide utilization review and use such a tool, for Maryland enrollees. In force 2025-10-01.
• Whether it applies depends on facts outside the code; a person has to decide.

resp = client.chat.completions.create(model=MODEL, messages=claim_msgs)
claim_decision = parse_decision(resp.choices[0].message.content)
claims.update(claim_id, status=claim_decision)

AI_USE_NOTICE = ('An AI system helped evaluate your claim. '
                 'You can ask us how it was used and request review by a claims adjuster.')
resp = client.chat.completions.create(model=MODEL, messages=claim_msgs)
claim_decision = parse_decision(resp.choices[0].message.content)
claims.update(claim_id, status=claim_decision, ais_program_ref='AIS-012')
send_ai_notice(claimant, text=AI_USE_NOTICE)

Standards that recommend the same control

• Insurers adopting the NAIC model guidance should maintain an AIS Program and AI-use notice (NAIC AI Model Bulletin · NAIC Model Bulletin, AIS Program Guidelines 1.1-1.9)

Rule id md-hb820.utilization-plan-ai-policies-and-inspection · review status: primary source derived