Industry Applications
Documented AI failures and rulings across regulated sectors, and where an Aethics assessment applies.
Standards-aligned evaluation supporting regulatory readiness and enterprise procurement.
Documented Cases
Real, publicly documented AI failures and rulings, each linked to its source. For every case we state what an Aethics assessment would and would not have examined. These are not Aethics client engagements.
Amazon's Experimental Recruiting Tool
What Happened
From 2014, Amazon built a machine-learning tool that scored job applicants' CVs from one to five stars. By 2015 it found the system was not rating candidates for software and other technical roles in a gender-neutral way: it penalized CVs containing the word "women's" and downgraded graduates of two all-women's colleges. Amazon abandoned the project.
Where an Aethics Assessment Applies
In scope for text models. A model that reads CVs is a text model, and gender bias in occupational contexts is exactly what the WinoBias benchmark measures.
What It Would Examine
- Gender and occupational bias testing (WinoBias)
- Stereotype benchmarks (StereoSet, CrowS-Pairs)
- Mapped to EU AI Act Article 10 — employment is an Annex III high-risk area
- Not covered: jurisdiction-specific bias audits such as NYC Local Law 144
Case Facts
Moffatt v. Air Canada
What Happened
In November 2022, Air Canada's website chatbot told a customer he could apply for a bereavement fare retroactively, which contradicted the airline's own policy. The airline argued the chatbot was responsible for its own statements. The British Columbia Civil Resolution Tribunal rejected that argument, held Air Canada liable for negligent misrepresentation, and awarded C$650.88 in damages plus interest and fees.
Where an Aethics Assessment Applies
Partially in scope. An assessment tests the chatbot's underlying model for harmful and unsafe outputs, and maps transparency requirements. It does not check answers against your own fares and policies.
What It Would Examine
- Harmful-output and safety tests on the underlying model
- Mapped transparency requirements (EU AI Act Article 13, OECD Principle 1.3)
- Not covered: verifying answers against company-specific policies
Case Facts
Italian Data Protection Authority v. OpenAI
What Happened
On 20 December 2024, Italy's data protection authority (the Garante) fined OpenAI €15 million over ChatGPT. It found no appropriate legal basis for processing personal data used for training, failures of transparency towards users, insufficient age verification, and a failure to notify a March 2023 data breach. It also ordered a six-month public information campaign. OpenAI said it would appeal.
Where an Aethics Assessment Applies
Partially in scope. An assessment tests model outputs for privacy leakage and maps the relevant GDPR requirements. The authority's core findings concern organisational controls that a model evaluation cannot assess.
What It Would Examine
- Privacy-leakage safety test on model outputs
- Mapped GDPR requirements (Articles 5(1)(a), 22, 35)
- Not covered: legal basis for training data, age verification, breach notification
Case Facts
Microsoft's Tay Chatbot
What Happened
Microsoft released Tay, a Twitter chatbot, on 23 March 2016. A coordinated effort by users led it to post abusive and offensive messages, and Microsoft took it offline within about 16 hours.
Where an Aethics Assessment Applies
In scope for the underlying model. Toxicity and harmful-output testing before release examines how a text model responds to abusive and manipulative prompts.
What It Would Examine
- Toxicity and harmful-output tests
- Refusal of unsafe requests
- Mapped safety requirements (NIST AI RMF MEASURE 2.2, OECD Principle 1.4)
- Not covered: live monitoring of a system after deployment
Racial Bias in a Population Health Algorithm
What Happened
A 2019 study in Science examined a widely used algorithm that decided which patients received extra care. Because it predicted health care costs as a proxy for health needs, Black patients were considerably sicker than White patients at the same risk score. Correcting the proxy would raise the share of Black patients receiving additional help from 17.7% to 46.5%.
Where an Aethics Assessment Applies
Outside today's scope. This was a tabular risk-prediction model; the platform currently evaluates text (NLP) models. The case shows why proxy variables in training data need scrutiny.
What It Would Examine
- Relevant mapped requirement: EU AI Act Article 10 on data governance and bias
- Not covered: tabular and risk-scoring models (text models only today)
Case Facts
Dutch Childcare Benefits Risk Model
What Happened
On 7 December 2021, the Dutch Data Protection Authority fined the Tax Administration €2.75 million for processing the (dual) nationality of childcare benefit applicants in an unlawful and discriminatory way, including using nationality as an indicator in a risk-classification model.
Where an Aethics Assessment Applies
Outside today's scope. This was a data-driven risk-classification system, not a text model. It is the kind of harm that fairness provisions in data protection law exist to prevent.
What It Would Examine
- Relevant mapped requirement: GDPR Article 5(1)(a) on fairness and transparency
- Not covered: rules-based and tabular risk-classification systems
Case Facts
Sector Coverage
Sectors with material AI governance requirements. Assessments today cover text-based (NLP) models used in these sectors; the regulations listed are sector context, not frameworks the platform assesses directly.
Healthcare & Life Sciences
Applications
- •Diagnostic AI systems
- •Clinical decision support
- •Patient risk stratification
- •Medical imaging analysis
Sector Regulations (Context)
Financial Services
Applications
- •Credit scoring models
- •Fraud detection systems
- •Algorithmic trading
- •Risk assessment models
Sector Regulations (Context)
Enterprise & HR
Applications
- •Recruitment screening
- •Performance evaluation
- •Workforce planning
- •Employee analytics
Sector Regulations (Context)
Critical Infrastructure
Applications
- •Threat detection
- •Operational monitoring
- •Predictive maintenance
- •Resource optimization
Sector Regulations (Context)
Education Technology
Applications
- •Adaptive learning
- •Student assessment
- •Learning analytics
- •Content recommendation
Sector Regulations (Context)
Legal & Professional Services
Applications
- •Document analysis
- •Contract review
- •Legal research
- •Due diligence support
Sector Regulations (Context)
Begin Assessment
Evaluate your text models against published bias and safety benchmarks, map the results to ten anchor frameworks, and document your governance readiness.
Built on an Analysis of 49 International Instruments