Task 4.3: Describe the importance of responsible AI in business.

• ARedacting the member IDs from both of the datasets

  Incorrect

  Redaction removes the value entirely, so the datasets can no longer be joined by member and nothing can be recovered.

• BPseudonymization with a reversible, two-way token

  Correct

  Pseudonymization replaces values with cryptographic tokens; consistent tokens keep the datasets joinable, and a two-way token can be reversed with the same key by those allowed to use it.

• CMasking every member ID with asterisk characters

  Incorrect

  Character masking hides the value but destroys it as a join key and cannot be reversed.

• DGeneralizing member IDs into numeric ranges

  Incorrect

  Bucketing into ranges suits numbers such as ages; it breaks one-to-one matching of members and cannot be reversed.

Anonymizing transformations such as redaction and masking are irreversible; pseudonymization keeps data useful and, with two-way tokens, recoverable under control.

• AAnalyze the training data and compare results by subgroup

  Correct

  Models can amplify biases in their training data. Analyzing the data and comparing outcomes across subgroups finds where the skew comes from and whether a fix works.

• BRetrain a larger model with more parameters on the same history

  Incorrect

  A bigger model trained on the same skewed history can reproduce the same pattern.

• CLower the temperature so rankings become consistent

  Incorrect

  Consistency is not fairness; a consistently biased ranking is still biased.

• DAdd a content filter for harmful language

  Incorrect

  Content filters block categories such as hate speech in text; this bias shows up in scores, not in offensive words.

Fairness problems usually start in the data. Subgroup analysis is how you see them and how you check that a fix works.

• AA more accurate model trained on the same inputs

  Incorrect

  Better accuracy does not tell an applicant why the model decided as it did.

• BFeature attributions returned with each prediction

  Correct

  When you request explanations, Agent Platform returns predictions with feature attribution scores that show how much each input feature contributed to that result.

• CGrounding the model's answers with Google Search results

  Incorrect

  Grounding links generated answers to sources; it does not explain a credit model's prediction.

• DModel Monitoring dashboards for the model

  Incorrect

  Watching a model's health over time does not explain an individual decision.

Explainability means showing why a model produced a particular output; feature attributions do that for each prediction and can also expose bias.

• AA caption under each image stating that AI made it

  Incorrect

  A caption discloses AI use on the page, but it is not part of the file and cannot be verified once the image is shared elsewhere.

• BA public AI use policy page on the publisher's website

  Incorrect

  A policy page explains the publisher's practices in general; it does not let anyone check where a particular image came from.

• CModel Armor screening of each image before publishing

  Incorrect

  Model Armor protects against harms and attacks such as prompt injection; it does not add provenance information to the content.

• DSigned C2PA Content Credentials in each image file

  Correct

  Content Credentials add cryptographically signed C2PA metadata to images generated with Gemini 3 Pro Image, indicating that they are AI-generated and giving a verifiable history of their origin.

Transparency includes being open about AI-generated content, and provenance metadata that travels with the file makes that disclosure verifiable.

• AFeature attributions that explain each recommendation

  Incorrect

  Attributions show which inputs drove a recommendation, which helps explain it, but they do not make anyone responsible for reviewing it.

• BBias testing of recommendations across supplier types

  Incorrect

  Bias testing checks whether outcomes are fair across groups; it does not name who reviews and answers for each decision.

• CHuman oversight with clear owners for each decision

  Correct

  Google's AI Principles call for appropriate human oversight, due diligence and feedback mechanisms; named owners and review steps make someone accountable for each decision.

• DModel Monitoring alerts on the recommendation model

  Incorrect

  Monitoring alerts on drift in the model's inputs or predictions; it can flag a problem, but it leaves open who acts on each recommendation.

Accountability means people, not models, answer for outcomes. Explanations, fairness checks and monitoring support those people, but reviews and ownership have to be designed into the process.

• AHuman oversight and due diligence

  Incorrect

  Appropriate human oversight, due diligence and feedback mechanisms are a commitment listed under responsible development and deployment, not one of the three principles.

• BCollaborative progress, together

  Correct

  Google's approach to AI rests on three principles: bold innovation; responsible development and deployment; and collaborative progress, together.

• CPrivacy, security and intellectual property

  Incorrect

  Promoting privacy and security and respecting intellectual property is a commitment under responsible development and deployment, not a principle of its own.

• DRigorous design, testing and monitoring

  Incorrect

  Rigorous design, testing, monitoring and safeguards are a commitment under responsible development and deployment, not one of the three principles.

Bold innovation, responsible development and deployment, and collaborative progress together frame how Google builds and deploys AI; oversight, testing and privacy are commitments that sit under the second.

• ALimit access to the tuning bucket with IAM roles and tune as is

  Incorrect

  IAM controls who can read the files, but the model would still learn from the personal data, and generated content can leak training data.

• BAdd a system instruction not to repeat customer details

  Incorrect

  An instruction at answer time does not remove personal data from the model, and adversarial users may try to get the model to override its instructions.

• CEncrypt the bucket with customer-managed keys and tune as is

  Incorrect

  Encryption protects data at rest, but the model would still learn from the personal data.

• DDe-identify them first, for example with Sensitive Data Protection

  Correct

  Personal data the task does not need should not go into tuning; Sensitive Data Protection can find and de-identify it before the data is used.

Privacy by design means removing or de-identifying personal data the use case does not need, before it reaches a model. Access controls, encryption and instructions protect around the data but leave it in the model.

• ARely on the model's published multilingual benchmark scores

  Incorrect

  Google states that most of its benchmarks, including all fairness evaluations, are in English, so they do not show quality for each resident group.

• BTest thoroughly in English, then launch in every language

  Incorrect

  Quality in English says little about other languages; performance can be worse for non-English languages and less represented varieties.

• CGround answers with Google Search results in each language

  Incorrect

  Grounding adds current sources to answers, but it does not show whether the model understands and answers well in each dialect.

• DTest answer quality for each language and dialect it will serve

  Correct

  Google notes that models can give inconsistent quality across users, with weaker performance for some dialects, non-English languages and underrepresented varieties, so each one needs testing.

Fairness includes consistent quality across the people served; measure it per language and dialect rather than assuming it from benchmarks or from English.