AIAI Agent Dev Labaiagentdev-lab.com
Sign in to save progressSign in
Menu

G検定 · LESSON 8

AI ethics and AI governance20%（当サイト設定） · 7 topics

Learn AI guidelines in Japan and abroad, privacy, fairness, safety and security, transparency and democracy, values such as environment and labor, and AI governance.

8.1Guidelines in Japan and abroad

Guidelines in Japan and abroad (AI principles, hard law and soft law, risk-based approach, EU AI Act, AI Guidelines for Business)

Concept

AI principles are basic ideas for developing and using AI, such as human-centricity, fairness, transparency, safety and privacy; examples include Japan's “Social Principles of Human-Centric AI” and the OECD AI Principles. Hard law is legally enforceable regulation; soft law is non-enforceable norms such as guidelines. The risk-based approach varies measures and the strength of regulation with the size of risk. The EU AI Act regulates by dividing AI into unacceptable risk (banned), high risk (strict obligations), limited risk (transparency obligations) and minimal risk. Japan's AI Guidelines for Business are soft law for developers, providers and users.

How it works / steps

  1. Check which risk category your AI falls into
  2. Check both domestic soft law and overseas hard law
  3. Bring guidelines into internal rules

Workplace example

A company offering recruitment support AI in the EU market falls under the EU AI Act's high risk, so it meets obligations such as risk management, data quality and human oversight.

Common confusions and exam tips

Mini quiz

Q1. Consider ways to govern AI. Which idea varies the strength of regulation and required measures according to the size of risk for each AI use?

  1. Agile governance
  2. Soft law
  3. Risk-based approach
  4. Hard law
Show answer and explanation

Answer: C Risk-based approach

The EU AI Act sorts AI into unacceptable risk (banned), high risk, limited risk and minimal risk. (Deciding point: regulation proportional to risk.)

Q2. Consider ways to govern AI. Which idea sets goals and keeps improving governance mechanisms by rapidly repeating evaluation and revision during operation?

  1. Multi-stakeholder approach
  2. Risk-based approach
  3. Soft law
  4. Agile governance
Show answer and explanation

Answer: D Agile governance

Agile governance is an approach for fast-changing technology, also described in the AI Guidelines for Business. (Deciding point: improving by repeated evaluation and revision.) [Other options] Soft law is non-binding norms, a risk-based approach scales measures to risk, and multi-stakeholder means involving diverse parties; none names the cycle of repeated evaluation and review.

Reference links

8.2Privacy

Privacy (privacy by design, camera image utilization guidebook)

Concept

Privacy by design builds privacy protection in from the planning and design stage. The camera image utilization guidebook shows businesses using camera images in stores and elsewhere considerations such as prior notice, stating the purpose and data management. Beyond complying with the APPI, considering users' expectations and concerns also matters.

How it works / steps

  1. Assess privacy impact at the planning stage (PIA)
  2. Collect only the minimum data needed and delete or anonymize it early
  3. Clearly announce that cameras are filming and why

Workplace example

In analyzing store visitors, don't store face images; aggregate only attributes (estimated age group and gender), discard images immediately, and post notice of filming and its purpose at the store.

Common confusions and exam tips

Mini quiz

Q1. Which combination best fits the values commonly discussed in AI principles and guidelines in Japan and abroad?

  1. Technical secrecy, developer anonymity, deregulation, data monopoly
  2. Maximum profit, shortest development time, minimum cost, maximum market share
  3. Promoting military use, stronger surveillance, full automation, excluding human involvement
  4. Human-centricity, fairness, privacy, safety, transparency and accountability
Show answer and explanation

Answer: D Human-centricity, fairness, privacy, safety, transparency and accountability

Common values appear in the OECD AI Principles, the Hiroshima AI Process, Japan's AI Guidelines for Business and others.

Q2. Consider privacy for AI using camera images. You are planning a new face recognition service and want to avoid noticing problems later and rebuilding. Which response is most appropriate?

  1. Set a procedure to promptly discard the original face images after estimating attributes
  2. Assess privacy impact from the planning and design stage and build in protections
  3. Post a notice at the entrance about filming, its purpose and privacy care
  4. Set rules against use beyond purpose or third-party sharing and show them to users
Show answer and explanation

Answer: B Assess privacy impact from the planning and design stage and build in protections

Privacy by design builds privacy protection in from the planning and design stage, with a privacy impact assessment (PIA). (Deciding point: building it in at design.) [Other options] Discarding images, usage rules and notices are valid protections, but to avoid rebuilding later, an impact assessment (PIA) at the design stage comes first.

Reference links

8.3Fairness

Fairness (algorithmic bias, sampling bias, sensitive attributes)

Concept

Algorithmic bias is AI decisions being skewed against particular groups. Causes include sampling bias, where training data was collected in a skewed way, and data reflecting past discriminatory decisions. Even removing sensitive attributes such as gender and race from features, bias can remain through strongly correlated proxy variables such as postal codes.

How it works / steps

  1. Check whether training data represents the target population
  2. Compare predictions and error rates by attribute
  3. If bias is found, review the data, features and thresholds

Workplace example

A document screening AI trained on past hiring results rated a particular gender unfavorably, so compare pass rates by attribute and review the training data and features.

Common confusions and exam tips

Mini quiz

Q1. An online store's AI was found inferring customers' health conditions from purchase history. Which describes the issue best?

  1. Biased training data makes inferences unfavorable to certain groups
  2. At the collection stage, purchase history was gathered without consent
  3. People can't correct the inference even if it's wrong
  4. At the inference stage, AI infers sensitive information the person never provided
Show answer and explanation

Answer: D At the inference stage, AI infers sensitive information the person never provided

As in a case where pregnancy was inferred from purchase history, inference results can invade privacy. [Other options] Collection without consent is a collection-stage issue. Bias is a fairness issue. Correction rights matter too, but inferring unprovided information is the point here.

Q2. Consider causes of AI fairness problems. The gender field was removed from inputs, but a “former school club” field strongly correlated with gender still produced gender differences. Which is the closest cause?

  1. Indirect discrimination via proxy variables
  2. Learning past discrimination (historical bias)
  3. Labeling bias
  4. Sampling bias (bias in how data was collected)
Show answer and explanation

Answer: A Indirect discrimination via proxy variables

Removing sensitive attributes isn't enough; watch for proxy variables too. (Deciding point: discrimination via a correlated variable.)

Reference links

8.4Safety, security and misuse

Safety and security (adversarial examples, data poisoning, model stealing, security by design) and misuse (deepfakes, fake news)

Concept

Attacks specific to AI include adversarial examples, which cause misrecognition with tiny noise humans can't see; data poisoning, which mixes improper data into training data; model stealing, which replicates a model through many API queries; and inferring training data (membership inference). Security by design builds security in from the design stage. Misuse includes fake images, video and audio (deepfakes) and spreading fake news.

How it works / steps

  1. Identify expected attacks and impacts at the design stage
  2. Manage the origin of training data to prevent poisoning
  3. Prevent model stealing with API query limits and monitoring

Workplace example

For traffic sign recognition in autonomous driving, prepare for adversarial examples where a small sticker causes misrecognition, adding robustness through adversarial training and multiple sensors.

Common confusions and exam tips

Mini quiz

Q1. Identify attacks on AI. Sending huge numbers of queries to a public API and replicating a nearly identical model from the input-output pairs. Which attack is this?

  1. Data poisoning
  2. Model stealing
  3. Adversarial examples
  4. Membership inference attack
Show answer and explanation

Answer: B Model stealing

Countermeasures to model stealing include limiting query counts and output information. (Deciding point: replicating a model from query results.)

Q2. Identify attacks on AI. Inferring from a model's output tendencies whether a person's data was used in training, invading privacy. Which attack is this?

  1. Model stealing
  2. Adversarial examples
  3. Data poisoning
  4. Membership inference attack
Show answer and explanation

Answer: D Membership inference attack

A kind of data theft, countered with techniques such as differential privacy. (Deciding point: inferring whether data was in the training set.)

Reference links

8.5Transparency and democracy

Transparency (explainability, data provenance, black boxes) and democracy (filter bubbles, echo chambers)

Concept

Transparency is being able to explain to stakeholders that AI is used and how it was built and makes decisions. Deep learning has a black box problem where reasons for decisions are hard to see, which explainability (XAI) techniques and records of data provenance (where data came from) help address. Effects on democracy include filter bubbles, where recommendation algorithms surround people with only information they like, and echo chambers, where similar opinions echo and strengthen.

How it works / steps

  1. Be able to explain AI use and the basis of decisions to users
  2. Record and manage the provenance of training data
  3. Design recommendation systems so users also encounter diverse information

Workplace example

A news app shows not only articles matching preferences but a set share of articles with different viewpoints, and shows why each was recommended.

Common confusions and exam tips

Mini quiz

Q1. Consider AI misuse and countermeasures. Scams using deepfake calls imitating the CEO's voice to order money transfers are rising. You want to review accounting procedures. Which countermeasure is most appropriate?

  1. Work with fact-checking organizations and systems, spreading the habit of verifying sources
  2. Define prohibited uses in the generative AI terms of use and suspend use when misuse is detected
  3. Build into identity verification a check through another channel, not relying on voice or video alone
  4. Attach digital watermarks and provenance information to generated content, combined with detection technology
Show answer and explanation

Answer: C Build into identity verification a check through another channel, not relying on voice or video alone

Prevent it through procedures such as calling back via another channel. (Deciding point: verification not relying on voice alone.)

Q2. Consider AI misuse and countermeasures. Your company's generative AI API is being used to mass-produce spoofing emails. You want to respond as the provider. Which countermeasure is most appropriate?

  1. Attach digital watermarks and provenance information to generated content, combined with detection technology
  2. Build into identity verification a check through another channel, not relying on voice or video alone
  3. Work with fact-checking organizations and systems, spreading the habit of verifying sources
  4. Define prohibited uses in the generative AI terms of use and suspend use when misuse is detected
Show answer and explanation

Answer: D Define prohibited uses in the generative AI terms of use and suspend use when misuse is detected

Providers handle misuse through terms of use and monitoring. (Deciding point: prohibition and suspension as the provider.)

Reference links

8.6Environment, labor and other values

Environmental protection, labor policy and other values (inclusion, military use)

Concept

Training and running large AI uses a lot of electricity, and environmental load is a challenge. AI automation changes job content, so labor policies such as reskilling employees matter. Inclusion is the idea that everyone, including people with disabilities and older people, can benefit from AI. Military use of AI, such as lethal autonomous weapon systems (LAWS), is also debated internationally.

How it works / steps

  1. Consider the electricity AI use consumes and ways to reduce it
  2. Plan reskilling for employees whose work is affected by automation
  3. Attend to accessibility so diverse users can use it

Workplace example

Alongside introducing AI for inquiry handling, reassign operators to managing AI response quality and handling complex cases, with training.

Common confusions and exam tips

Mini quiz

Q1. Which best describes the debate on AI and environmental protection?

  1. AI computation uses no electricity, so its environmental impact isn't discussed
  2. AI power use is unrelated to model size or the number of training runs
  3. Power consumption and CO2 emissions from training and running large models are concerns
  4. Using AI always increases environmental load, so it can never be used for climate measures
Show answer and explanation

Answer: C Power consumption and CO2 emissions from training and running large models are concerns

On the other hand, AI is also used for climate measures such as optimizing power supply and demand. Designing efficient models also matters.

Q2. Which best describes the debate on AI and labor?

  1. For work handed to AI, it's no problem if people lose the skills for it
  2. Job content changes, but collaboration with AI and new jobs also arise, making reskilling important
  3. AI will soon fully replace all jobs, so human work will disappear
  4. AI has no effect on work at all, so there is no need to rethink how we work
Show answer and explanation

Answer: B Job content changes, but collaboration with AI and new jobs also arise, making reskilling important

It can also offset labor shortages. Skill loss from over-reliance on AI is another issue.

Reference links

8.7AI governance

AI governance (AI policy, ethics assessment, stakeholder involvement, audits, traceability)

Concept

AI governance is an organization's mechanism for managing AI risks and maximizing benefits. Set an AI policy (basic principles), run ethics assessments (evaluating risks and impacts) at the planning and development stages, and incorporate the views of stakeholders such as users and affected people. Check operations with internal and external audits and ensure traceability (being able to trace the history of data, models and decisions).

How it works / steps

  1. Set an AI policy and a responsibility structure
  2. Run an ethics assessment for each new AI plan
  3. Keep records of data and model versions and of decisions so they can be audited

Workplace example

Before introducing a new credit AI, an ethics committee runs an assessment, hears outside experts, and decides to keep decision records for 5 years.

Common confusions and exam tips

Mini quiz

Q1. Choose an AI governance measure. The development team is skewed in gender, age and expertise and tends to overlook effects on unexpected users. Which is most appropriate?

  1. Set up a review structure with members of diverse backgrounds
  2. Run ethics assessments (impact assessments) before development and during operation
  3. Undergo an audit by an outside third party
  4. Draft and publish an AI policy
Show answer and explanation

Answer: A Set up a review structure with members of diverse backgrounds

A diverse structure brings in the viewpoints of varied stakeholders. (Deciding point: adding diverse perspectives.)

Q2. Choose an AI governance measure. Internal checks alone don't earn trust, so you want your AI management structure and operations verified from an independent position. Which is most appropriate?

  1. Set up a review structure with members of diverse backgrounds
  2. Undergo an audit by an outside third party
  3. Draft and publish an AI policy
  4. Run ethics assessments (impact assessments) before development and during operation
Show answer and explanation

Answer: B Undergo an audit by an outside third party

AI audits raise objectivity and trust with a third-party view. (Deciding point: verification from an independent position.)

Reference links

Lessons are written by this site based on public information and do not guarantee exam content. Product specifications change, so check the latest information at the reference links.

Notes (unofficial, original questions, no pass guarantee)