Explain two problems that arise in AI decisions and output: unfair decisions caused by skew in training data or design (algorithmic bias), and output of content that is not true (hallucination).
Explain how fake video and audio that synthesize a person's face or voice (deepfakes) can be misused for impersonation or for spreading false information.
Explain social issues around AI: who is responsible for harm, how to decide in an unavoidable accident, and ways to refuse the use of one's data for training.
Algorithmic bias (アルゴリズムのバイアス) is when skew in training data or design tilts an AI's decisions unfairly according to attributes such as gender or age.
Past data that contains skew
The AI learns that pattern
Decisions unfavorable to certain groups
It is not that the AI means harm; it reads habits of human judgment left in past records as if they were rules. Since the cause usually lies in the data, countermeasures focus on the stage of collecting and preparing the data used for training.
AnalogyIt is like judging national tastes from one region's shop sales alone: that region's preferences end up as the national standard.
How to rememberBias originally meant "slant"; cutting fabric on the diagonal is still called a "bias cut." Picture a decision that should run straight leaning to one side.
Exam points
ExamCountermeasures against bias (バイアス) overlap with the work of keeping data quality, so they cover all data preparation before training begins.
Hallucination (ハルシネーション) is when generative AI outputs content that is not true in natural-sounding text, as if it were correct.
The user asks a question
Pick a likely next word by probability
Build sentences without checking the facts
The user checks the facts
So before using the output, the user checks the facts against another source.
4 / 4
Generative AI builds sentences by picking likely next words from patterns in the text it learned, so the word-picking itself never checks whether the content is true. It may give nonexistent references or wrong figures in a confident tone. When the output is used for work, the user confirms the facts with another source.
AnalogyIt is like someone who half-remembers a route but gives directions without a trace of doubt. The surer they sound, the harder it is to question them.
How to rememberA hallucination is seeing something that is not there; here, the AI states facts that are not there.
Exam points
ExamHallucination (ハルシネーション) refers to errors in the content of output, not its form. Fluent text is no proof that the content is correct.
A deepfake (ディープフェイク) is fake video or audio that uses deep learning to synthesize a person's face or voice, making it look real that they said or did things they never did.
Because it is made by training AI on the features of the person's real video or voice, it is hard to tell apart by looks or sound alone. It can be misused for fraud by impersonating the person, or for spreading false information.
AnalogyIt is like someone else wearing a costume that looks exactly like a person and speaking as them. The closer the likeness, the harder it is to see through.
How to rememberDeep (deep learning) + fake (forgery). The name itself says how it is made and what it is for.
Exam points
Mix-upA deepfake (ディープフェイク) is a fabrication its maker knowingly creates as fake. What separates it from generative AI (生成AI) unintentionally producing wrong content is intent.
The AI service liability debate (AIサービスの責任論) asks who bears responsibility when an AI's decision or output causes harm or an accident.
Because AI makes its own decisions based on what it learned, the traditional view that whoever used or made the tool is responsible does not settle every case. For example, many parties are involved, such as the company that built the AI, the business offering the service and the user, so it is hard to pin down who is responsible.
AnalogyIt is like an apprentice botching a job: whether the fault lies with the teacher, the shop that handed over the work or the customer who gave the instructions is hard to settle in a word.
The trolley problem (トロッコ問題) is a thought experiment in ethics that asks, when someone will die whichever way you choose, whether it is right to sacrifice a few to save many.
← Scroll the figure sideways →
In the best-known setup, switching a runaway trolley's track kills 1 person, while leaving it kills 5. It has become an issue in AI design, because how a self-driving car should act in an unavoidable accident has to be decided in advance. There is no single right answer, and debate continues over how to build ethical judgment into AI.
AnalogyIt is like having to decide in advance who gets a place when a lifeboat can carry only so many people.
An AI service opt-out policy (AIサービスのオプトアウトポリシー) lets users, on request, choose to keep their input data and the like out of AI training.
Opt-out
Opt-in
Default
Data is used
Data is not used
User action
Ask to refuse
Ask to allow
If the user does nothing
Use continues
Not used
In AI services, text or images that users enter may be used to train the model so the service can improve. The policy is a user protection measure that answers concerns about data being used for training without the person knowing.
AnalogyIt is like a mailing list that starts with everyone on it, where only those who want off have to ask to be removed.
How to rememberOpt out = choose to step out. Everyone starts inside, and those who want to leave step out.
Key points
Countermeasures against algorithmic bias extend to all data preparation before training, including how training data is collected and cleaned up.
With hallucination, generative AI output can read naturally yet not match the facts, so the user checks them.
Deepfakes are fake video or audio that synthesize a person's face or voice, and can be misused for impersonation or for spreading false information.
Harm involving AI has many parties, so where responsibility lies is the question taken up by the AI service liability debate.
An AI service opt-out policy answers users' concerns about their input data being used for training.