Digital Transformation, AI, and the Future of Workplace Safety

Introduction and Context

Workplace safety has traditionally depended on inspections, training, incident reports, and the professional judgment of supervisors and safety specialists. Digital transformation expands this system by making hazards more visible, continuously measurable, and potentially preventable before injury occurs. Wearable devices can identify heat exposure or proximity to moving equipment; computer-vision systems can detect missing personal protective equipment; robots can enter environments too dangerous for people; and digital twins can simulate operational changes before they are implemented.

These technologies do not, however, make safety automatically intelligent. They redistribute organizational power, generate new forms of surveillance, create cybersecurity dependencies, and may reproduce historical inequalities through biased data or poorly designed algorithms. The central managerial question is therefore not simply whether an organization can deploy an AI system, but whether the system produces a demonstrably safer, fairer, more resilient, and legally defensible workplace.

For HR professionals, occupational-health specialists, operations managers, and safety leaders, digital safety is a sociotechnical field. Technical performance must be evaluated together with work design, employee participation, privacy, collective bargaining, professional ethics, organizational culture, and regulatory obligations.

Learning Objectives

By the end of this chapter, readers should be able to:

1. Explain the purposes and limitations of wearables, sensors, computer vision, robotics, digital twins, predictive analytics, and AI-enabled risk detection.
2. Distinguish hazard detection, risk prediction, intervention, and accountability.
3. Evaluate digital safety systems using validity, reliability, fairness, usability, privacy, and cybersecurity criteria.
4. Analyze worker-surveillance, algorithmic-bias, ESG, climate-risk, and future-of-work implications.
5. Design a governance framework for responsible adoption of AI-enabled safety technologies.
6. Critically assess an applied implementation and recommend improvements.

1. From Reactive Safety to Cyber-Physical Safety Management

A reactive safety system learns primarily from injuries, near misses, complaints, and regulatory inspections. A proactive system identifies unsafe conditions before harm occurs. A predictive system estimates where and when risk may increase. A cyber-physical safety system connects digital observations with physical interventions: for example, a sensor detects excessive temperature, a platform alerts a supervisor, and the work schedule is changed.

A useful distinction is:

• Hazard detection: identifying a source or condition with the potential to cause harm.
• Exposure measurement: estimating the extent to which a worker, group, or asset encounters the hazard.
• Risk assessment: combining the likelihood and severity of harm under specified conditions.
• Prediction: estimating a future event or state from current and historical data.
• Prevention or mitigation: changing equipment, work organization, behavior, or environment to reduce risk.
• Assurance: demonstrating that controls work and that decisions are lawful, fair, and auditable.

Digital tools can improve the first four activities, but prevention still depends on management decisions and worker participation. An accurate alert that nobody can act upon has little protective value.

Risk is often represented conceptually as:

where R is risk, P(H) is the probability or likelihood of hazardous harm H, and S(H) is the severity of its consequence. This is a simplified model, not a universal physical law. It assumes that likelihood and severity can be meaningfully characterized and that their product is useful for comparison. In practice, risk matrices use ordinal categories and can conceal uncertainty; they should support, not replace, professional judgment and worker knowledge.

Digital transformation adds a temporal dimension. A system may identify a rising probability of collision, heat illness, fatigue, or musculoskeletal strain. Yet organizations should avoid treating a prediction as a fact. A prediction is an uncertain estimate that must be validated in the operational context.

2. The Digital Safety Technology Landscape

2.1 Wearables and Personal Sensors

Wearables include smart helmets, location tags, physiological monitors, exoskeletons, connected hearing protection, and devices measuring motion, temperature, heart rate, or fatigue-related indicators. Their legitimate safety applications include:

• exposure alerts for heat, noise, toxic gases, or radiation;
• geofencing around restricted or high-energy areas;
• lone-worker check-ins and fall detection;
• ergonomic feedback based on posture or repetitive motion;
• emergency location during evacuation or confined-space rescue.

A major limitation is construct validity: a measured variable may not accurately represent the safety construct of interest. Heart rate, for example, is affected by exertion, medication, emotion, temperature, and individual physiology. It should not automatically be interpreted as proof of fatigue or poor performance. Devices also differ in calibration, battery life, fit, connectivity, and tolerance across body types, skin tones, languages, and job tasks.

The least intrusive effective design is generally preferable. Organizations should first ask whether an environmental sensor, engineering control, or scheduling change can address the hazard without collecting individualized physiological data.

2.2 Environmental and Equipment Sensors

Internet-of-Things sensors can monitor vibration, pressure, gas concentration, temperature, humidity, noise, machine guards, vehicle movement, and structural conditions. Sensor networks can support preventive maintenance and real-time alerts. However, they introduce false positives, false negatives, missing data, calibration requirements, and dependence on network availability.

A sensor program should specify the measurement range, accuracy, sampling interval, calibration schedule, alert threshold, data owner, retention period, and response procedure. A technologically sophisticated installation without a defined response protocol is merely data accumulation.

2.3 Computer Vision

Computer vision applies cameras and machine-learning models to images or video. Safety uses include detecting entry into dangerous zones, vehicle-pedestrian proximity, falls, smoke, crowding, unsafe lifting, and missing protective equipment.

Vision systems are particularly sensitive to lighting, occlusion, camera angle, clothing, motion blur, and changing work environments. A model trained in one warehouse may perform poorly in another. Facial recognition is a distinct and substantially more intrusive application than anonymous hazard detection. Safety objectives do not provide automatic justification for identifying individual workers.

Good design favors event-based processing, edge computation where feasible, data minimization, masking or blurring, short retention, and human review for consequential decisions. A camera alert should normally trigger investigation or immediate hazard control—not automatic discipline.

2.4 Robotics and Autonomous Systems

Robots can remove workers from confined spaces, explosive environments, high elevations, contaminated areas, and repetitive heavy work. Collaborative robots may share workspaces with people, but their safety depends on speed and force limits, sensing, guarding, emergency stops, software assurance, maintenance, and human-machine interface design.

Automation can reduce physical exposure while creating new hazards: unexpected start-up, software failure, loss of situational awareness, automation complacency, maintenance exposure, and psychosocial stress associated with intensified performance monitoring. A task removed from one group may be transferred to technicians, cleaners, or subcontractors. Safety analysis must follow the entire work system and supply chain.

2.5 Digital Twins

A digital twin is a continuously updated digital representation of a physical asset, process, facility, or system. Unlike a static simulation, it is connected to operational data and may be used to test scenarios such as evacuation, ventilation changes, traffic flows, equipment placement, heat exposure, or emergency response.

Digital twins are valuable for “what-if” analysis, but their conclusions depend on model assumptions and data quality. A twin may omit informal workarounds, contractor behavior, degraded equipment, or social conditions. It should therefore be treated as a decision-support model, not a perfect replica of reality.

2.6 Predictive Analytics and AI-Enabled Risk Detection

Predictive systems use statistical or machine-learning methods to estimate outcomes such as incidents, equipment failure, absenteeism, or exposure exceedance. Common inputs include maintenance records, environmental measurements, work schedules, near-miss reports, weather, and operational conditions.

The distinction between correlation and causation is essential. If incidents are more frequent on a particular shift, the shift itself may not be the cause; staffing levels, production pressure, training, equipment condition, or reporting practices may be the underlying factors. A model that predicts risk without identifying modifiable causes can encourage ineffective or unfair interventions.

A binary classifier can be evaluated using a confusion matrix:

Important measures include:

Precision is the proportion of positive alerts that are correct; recall is the proportion of actual positive cases detected. These measures assume clearly defined labels and a relevant reference standard. A safety system may prioritize recall when missing a dangerous event is catastrophic, but excessive false alarms can create alert fatigue. Threshold selection must therefore reflect harm, resources, and the consequences of error.

Accuracy, defined as (TP + TN) ÷ (TP + TN + FP + FN), may be misleading when dangerous events are rare. For example, a system that never predicts an incident could appear highly accurate in a dataset where incidents are uncommon, while providing no preventive value.

3. A Critical Evaluation Framework

A graduate-level evaluation should examine at least eight dimensions.

Validity and Reliability

Validity concerns whether a system measures the intended hazard or predicts the relevant outcome. Reliability concerns consistency across time and conditions. A device may be reliable but invalid—for example, consistently measuring a variable that is only weakly related to injury risk. Both are necessary.

Validation should occur before and after deployment. Pre-deployment testing should use representative tasks, environments, devices, and worker populations. Post-deployment monitoring should assess performance drift caused by changed processes, seasons, equipment, or workforce composition.

Human-in-the-Loop and Human-on-the-Loop

A human-in-the-loop reviews or authorizes a decision before action. A human-on-the-loop supervises an automated system that may act independently but can intervene. The appropriate arrangement depends on the severity and reversibility of the decision. Automatic shutdown of a dangerous machine may be justified when delay creates immediate danger; automatic disciplinary action based on a probabilistic alert generally requires much stronger safeguards and should not be treated as equivalent.

Data Quality and Missingness

Safety data are socially produced. Near misses may be underreported where workers fear blame; injuries may appear higher in a department with a stronger reporting culture; contractor data may be incomplete; and historical records may reflect past discrimination. Missingness is therefore not simply a technical inconvenience. It can reveal who is excluded from the measurement system.

4. Privacy, Surveillance, and Worker Autonomy

Digital safety systems can blur the boundary between protection and performance control. Location data collected for emergency response may later be used to rank workers. Physiological data collected for heat prevention may be used to infer productivity or health status. Video collected for vehicle safety may become continuous behavioral surveillance.

Responsible governance begins with purpose limitation: define a specific safety purpose and prohibit incompatible secondary uses unless separately justified and authorized. Other principles include:

1. Necessity: collect only data needed for a legitimate safety objective.
2. Proportionality: use the least intrusive effective method.
3. Transparency: explain what is collected, why, by whom, and for how long.
4. Participation: involve workers, safety representatives, occupational-health professionals, and relevant unions.
5. Choice where feasible: do not make consent nominal when refusal threatens employment.
6. Access and correction: allow individuals to inspect and challenge personal data.
7. Retention limits: delete identifiable data when the safety purpose expires.
8. Separation: keep occupational-health information separate from routine managerial performance records.

Privacy law varies by jurisdiction. Organizations must assess applicable employment, data-protection, occupational-health, disability, biometric, and algorithmic-accountability requirements rather than assuming that a general notice is sufficient. Legal compliance is a minimum; ethical legitimacy also requires trust and meaningful worker voice.

5. Bias, Equity, and Algorithmic Accountability

Bias can enter through problem definition, data collection, labels, model design, deployment, or organizational response. Examples include a vision model that performs less accurately under particular lighting or skin tones, a fatigue model that treats disability-related variation as unsafe behavior, or an incident model trained on reports that overrepresent departments with stronger reporting systems.

Fairness cannot be reduced to one statistical metric. Relevant questions include:

• Are false negatives higher for a particular group?
• Are alerts more likely to trigger scrutiny for certain occupations or demographic groups?
• Does the model use proxies for protected characteristics, such as location, shift, language, or employment status?
• Are workers able to challenge an output?
• Does the intervention address a system hazard or merely burden the predicted individual?

Different fairness criteria can conflict, especially when base rates differ across groups. Consequently, organizations should combine quantitative subgroup testing with contextual investigation, worker consultation, and impact assessment. A fairness review should examine outcomes, not only model code.

Algorithmic accountability requires a documented chain of responsibility: the system owner, data steward, safety authority, privacy officer, security lead, vendor, and final decision-maker. Contracts should provide audit rights, incident notification, model-change disclosure, data deletion, and restrictions on vendor reuse. “The algorithm decided” is not an acceptable allocation of responsibility.

6. Cybersecurity and Operational Resilience

Connected safety technologies expand the attack surface of an organization. A compromised sensor may suppress an alarm; manipulated location data may misdirect emergency responders; ransomware may disable access-control or industrial systems; and stolen health data may cause severe personal harm.

A safety-oriented cybersecurity program should include asset inventory, identity and access management, encryption in transit and at rest, network segmentation, secure updates, vulnerability management, backups, logging, supplier assurance, and tested incident-response procedures. Security controls must not undermine emergency availability. For example, authentication requirements should not delay a necessary emergency shutdown.

Threat modeling should consider confidentiality, integrity, and availability. In safety-critical settings, integrity and availability may be as important as confidentiality. Organizations should also plan for graceful degradation: what happens when connectivity, cloud services, GPS, or the AI model is unavailable? Manual procedures, local alarms, and trained personnel remain essential.

7. ESG, Climate Risk, and the Future of Work

Digital safety is increasingly connected to environmental, social, and governance expectations. Environmentally, sensors can reduce energy use and detect leaks, while digital twins can model climate adaptation. Socially, organizations are expected to protect workers, contractors, communities, and vulnerable groups. From a governance perspective, transparent metrics, responsible data use, and board oversight are increasingly material to reputation and investment.

Climate change alters occupational risk through heat stress, wildfire smoke, flooding, storms, vector-borne disease, poor air quality, and supply-chain disruption. Predictive systems should integrate credible environmental data, but organizations must not rely on a dashboard instead of engineering controls, work-rest schedules, hydration, respiratory protection, evacuation planning, and emergency communication.

Future forms of work complicate responsibility. Remote and hybrid workers may face ergonomic, psychosocial, isolation, and domestic-environment risks that employers cannot monitor intrusively. Platform workers and contractors may be managed through opaque algorithms while lacking access to training, insurance, or meaningful appeal. Automated warehouses and AI-mediated scheduling may intensify pace and reduce discretion. A just transition requires reskilling, participation in automation design, and attention to who receives the benefits and who bears new risks.

8. Implementation and Governance Framework

A responsible implementation can follow this sequence:

Step 1: Define the safety problem

Specify the hazard, affected population, existing controls, baseline outcomes, and decision that the technology is intended to improve. Do not begin with a vendor solution.

Step 2: Apply the hierarchy of controls

Consider elimination, substitution, engineering controls, administrative controls, and personal protective equipment before relying on monitoring individual behavior. Digital surveillance should not compensate for an avoidable design failure.

Step 3: Conduct an impact assessment

Assess safety benefits, privacy, labor relations, disability inclusion, discrimination, cybersecurity, environmental effects, and foreseeable misuse. Identify stakeholders and consult them early.

Step 4: Establish data and model governance

Document data sources, lawful purpose, ownership, access, retention, quality, labeling, model version, thresholds, and validation evidence. Define prohibited uses.

Step 5: Pilot safely

Use a limited, representative pilot with worker participation. Do not use pilot data for discipline or employment decisions unless that purpose was explicitly assessed and authorized.

Step 6: Evaluate outcomes

Compare exposure, incident, near-miss, response-time, and worker-experience measures with a credible baseline. Examine subgroup performance and unintended consequences. Where possible, use a controlled before-and-after design, while recognizing that randomization may be impractical or unethical in safety-critical contexts.

Step 7: Scale with assurance

Train users, maintain fallback procedures, audit vendors, monitor drift, and establish periodic independent review. A system should have a retirement plan if it fails to produce net safety benefit.

A useful decision rule is that deployment should proceed only when expected safety benefit is material, evidence is adequate, residual risks are controlled, and affected workers have meaningful avenues for information and remedy.

Applied Case: Heat-Risk Analytics in a Distribution Centre

This is an illustrative fictional case.

Background

NorthStar Distribution operates a large warehouse where summer temperatures, high workloads, and limited ventilation have produced several heat-related near misses. Management proposes wrist-worn sensors that estimate physiological strain and an AI dashboard that predicts high-risk periods using temperature, workload, shift length, and historical incidents.

Problem

The proposal promises early warnings but raises concerns. Workers fear that heart-rate data will be used to identify “slow” employees. The historical incident data contain few reports from temporary workers. The vendor cannot explain how its fatigue score was validated across different ages, medications, body sizes, and job tasks. A technical review also finds that alerts depend on uninterrupted wireless connectivity.

Analysis

The safety problem is real, but the proposed solution risks individualizing a workplace exposure. Heat risk is influenced by ambient conditions, clothing, exertion, acclimatization, hydration, rest opportunities, and medical factors. A physiological signal cannot safely establish a worker’s capacity or blame a worker for an unsafe environment.

The organization therefore maps controls in order of priority: improve ventilation; schedule heavy tasks during cooler periods; provide hydration and shaded recovery; establish acclimatization procedures; train supervisors and workers; install environmental sensors; and use optional, privacy-preserving wearables only where a clear additional benefit remains.

Approach and Solution

NorthStar creates a joint worker-management safety committee. Environmental sensors are installed first, with local alarms that function without cloud connectivity. The wearable pilot is voluntary, collects short-lived exposure alerts rather than raw continuous physiological data, and prevents supervisors from viewing individual health measures. Data are retained only for aggregate safety analysis. The model is tested across shifts and worker groups, with recall and false-alert rates reported separately. Alerts trigger a welfare check and review of conditions, never automatic discipline.

The vendor contract requires documentation of model changes, security controls, breach notification, deletion, and independent audit. A manual heat plan remains in force if the system fails.

Outcome

After one summer, heat-exposure duration and delayed recovery periods decline. Workers report greater trust because the program changes environmental controls rather than ranking individuals. The model produces many false alerts during a particular high-exertion task, leading the organization to revise the task design rather than lower the threshold and ignore the problem.

Learning Points

1. The technology was useful only after higher-order controls were strengthened.
2. Data minimization improved legitimacy and reduced misuse risk.
3. Subgroup and task-specific validation exposed weaknesses hidden by overall accuracy.
4. Worker participation improved both design quality and adoption.
5. Resilience required local alarms and manual procedures.
6. A prediction should initiate investigation and care, not determine blame.

Challenges and Limitations

Digital safety systems face persistent limitations. Rare-event prediction is difficult because reliable labels are scarce. Models may drift as processes and workforces change. Sensors can create false reassurance when unmeasured hazards remain. Excessive alerts produce fatigue and normalization of deviance. Vendor secrecy may prevent meaningful audit. Data-rich organizations may still lack the capability to interpret outputs. Finally, quantification can marginalize experiential knowledge: workers often recognize hazards before they become measurable events.

There are also distributive questions. Who pays for devices? Who owns the data? Who benefits from increased productivity? Who is exposed to new monitoring? A system that lowers injury rates while increasing stress, autonomy loss, or discrimination cannot be judged successful solely by one outcome metric.

Best-Practice Checklist

• Begin with a clearly defined hazard and baseline.
• Apply the hierarchy of controls before individual monitoring.
• Involve workers and representatives in design and evaluation.
• Collect the minimum data necessary for the stated purpose.
• Validate accuracy, reliability, and subgroup performance in real conditions.
• Separate safety support from discipline and performance ranking.
• Require explainability appropriate to the decision’s consequences.
• Maintain human authority, appeal, and remedy.
• Secure systems and prepare for outages and cyber incidents.
• Audit vendors, model drift, unintended effects, and climate relevance.
• Publish meaningful aggregate results and retire ineffective systems.

Chapter Summary

Digital transformation can move safety management from episodic reaction toward continuous detection, analysis, and prevention. Wearables, sensors, computer vision, robotics, digital twins, predictive analytics, and AI each offer valuable applications, but each also creates technical, organizational, ethical, and legal risks. Effective evaluation requires attention to validity, reliability, false alarms, missing data, human factors, fairness, privacy, cybersecurity, and accountability.

The hierarchy of controls remains fundamental: digital monitoring should not substitute for eliminating hazards or improving engineering and work design. Predictions are uncertain estimates, not judgments of worker character or fitness. Responsible systems minimize data, involve workers, provide human review and appeal, protect against misuse, and retain manual safeguards. Climate change, platform work, automation, and hybrid employment make safety responsibility broader and more complex. The future of workplace safety will therefore depend not on automation alone, but on the quality of sociotechnical governance surrounding it.

Key Terms

Algorithmic accountability: Assignment of responsibility for the design, operation, effects, and correction of an algorithmic system.

Artificial intelligence (AI): Computational methods that perform tasks commonly associated with human intelligence, such as classification, prediction, or language processing.

Computer vision: Analysis of images or video by computational models.

Data minimization: Limiting collection and retention to data necessary for a specified purpose.

Digital twin: A data-connected digital representation of a physical asset, process, or environment.

Exposure measurement: Estimation of the intensity, duration, or frequency of contact with a hazard.

False negative: A hazardous or positive condition that a system fails to detect.

False positive: An alert indicating a condition that is not actually present.

Human-in-the-loop: A design in which a person reviews or authorizes an automated decision.

Internet of Things (IoT): Interconnected devices that sense, communicate, and exchange data.

Model drift: Decline in model performance as conditions or data distributions change.

Predictive analytics: Use of data and statistical or machine-learning methods to estimate future outcomes.

Privacy by design: Integrating privacy protections into system architecture and operations from the beginning.

Recall: The proportion of actual positive cases detected by a classifier.

Risk: The combination of likelihood and severity associated with potential harm.

Sensor: A device that detects and measures a physical, environmental, or biological variable.

Surveillance: Systematic observation or collection of information about people, behavior, or conditions.

Validity: The extent to which a measure or model represents what it is intended to represent.

Wearable technology: A body-worn digital device that measures, communicates, or supports a task.

Worker participation: Meaningful involvement of workers and their representatives in safety decisions.

Review Questions

Short Answer Questions

1. Distinguish hazard detection, risk prediction, and prevention.
2. Why can overall accuracy be misleading for rare safety events?
3. Identify two privacy risks associated with wearable devices.
4. What is model drift, and why does it matter in workplace safety?
5. Explain why the hierarchy of controls remains relevant to digital safety.

Long Answer Questions

1. Critically evaluate the benefits and limitations of computer vision for workplace safety.
2. Discuss how worker participation can improve the legitimacy and technical quality of AI-enabled safety systems.
3. Explain the relationship among privacy, surveillance, and algorithmic accountability.
4. Analyze the cybersecurity requirements of connected safety technologies.
5. Discuss how climate change and future forms of work alter digital safety priorities.

Analytical / Application Questions

1. An organization proposes using location data collected for emergency evacuation to rank worker productivity. Apply purpose limitation, proportionality, and accountability principles to assess the proposal.
2. A predictive model has high overall accuracy but substantially more false negatives for temporary workers. What investigations and corrective actions should the organization undertake?
3. Design a pilot plan for an AI-enabled safety system in a manufacturing facility. Specify the hazard, baseline, stakeholders, data safeguards, performance measures, human controls, and criteria for continuation or termination.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes prediction from prevention in digital safety?

A. Prediction automatically eliminates the hazard

B. Prediction estimates a future condition, while prevention changes conditions to reduce harm

C. Prediction is based only on injury records

D. Prevention requires no human or organizational action

Correct Answer: B. Prediction estimates a future condition, while prevention changes conditions to reduce harm

Question 2

Which measure is the proportion of actual positive cases detected by a classifier?

A. Precision

B. Accuracy

C. Recall

D. Specificity

Correct Answer: C. Recall

Question 3

Why may accuracy be misleading for a rare workplace incident?

A. It cannot include true negatives

B. A model can appear accurate by predicting the common negative condition while missing incidents

C. Accuracy measures only privacy compliance

D. Accuracy is calculated only for images

Correct Answer: B. A model can appear accurate by predicting the common negative condition while missing incidents

Question 4

Which practice best reflects data minimization?

A. Retaining raw physiological data indefinitely

B. Collecting all available worker data for possible future use

C. Collecting only the data necessary for a defined safety purpose

D. Sharing safety data with every organizational department

Correct Answer: C. Collecting only the data necessary for a defined safety purpose

Question 5

What is a principal limitation of a digital twin?

A. It cannot represent physical processes

B. Its conclusions depend on the completeness and quality of its assumptions and data

C. It always operates without real-time data

D. It eliminates the need for emergency procedures

Correct Answer: B. Its conclusions depend on the completeness and quality of its assumptions and data

Question 6

Which response is most appropriate when an AI safety alert identifies a potentially dangerous condition?

A. Automatically discipline the identified worker

B. Ignore the alert unless an injury occurs

C. Use the alert to initiate investigation and appropriate hazard control

D. Publish the worker’s personal data

Correct Answer: C. Use the alert to initiate investigation and appropriate hazard control

Question 7

Which is an example of algorithmic bias?

A. A model is tested under representative operating conditions

B. A model has different error rates across worker groups because its training data underrepresent some groups

C. A sensor has a documented calibration schedule

D. A system has a manual fallback procedure

Correct Answer: B. A model has different error rates across worker groups because its training data underrepresent some groups

Question 8

Why are local alarms and manual procedures important in connected safety systems?

A. They make cybersecurity unnecessary

B. They provide resilience when networks, cloud services, or models fail

C. They increase the amount of personal data collected

D. They eliminate the need for worker training

Correct Answer: B. They provide resilience when networks, cloud services, or models fail

Question 9

Which control generally has priority over monitoring individual worker behavior?

A. Eliminating or engineering out the hazard

B. Increasing surveillance intensity

C. Reducing the reporting of near misses

D. Using disciplinary thresholds

Correct Answer: A. Eliminating or engineering out the hazard

Question 10

A classifier has TP = 80 and FP = 20. Using Precision = TP ÷ (TP + FP), what is its precision?

A. 20%

B. here? 80%

C. 80%

D. 100%

Correct Answer: C. 80%