• Foundation
  • About 15 min
  • 5 practice questions

AI safety, jobs and governance: human-in-the-loop and why people stay responsible

Explain AI safety and misuse, the effect of automation on jobs and the need for reskilling, the levels of human control, and the main approaches to governing AI around the world.

If something looks wrong, please tell us.

The first four lessons looked at specific problems: unfairness, privacy and fraud. This lesson steps back to three big questions. How do we keep AI systems safe? What happens to jobs? And who makes the rules? The answer to all three ends in the same place: however capable a machine becomes, a person or organisation must remain responsible for what it does.

AI safety and misuse

AI safety means making sure an AI system works as intended, does not cause harm, and can be controlled. Problems fall into two groups. Accidents are harms nobody intended: a system gives confident but wrong answers, fails on situations it was not trained for, or is tricked by unusual inputs. Misuse is deliberate: people use AI to clone voices for fraud, write convincing scam messages, spread disinformation or attack computer systems. Safety work therefore includes testing before launch, limiting what a system is allowed to do, watching it in use, and having a way to switch it off.

Two groups of AI safety problems

  • Accidents Harms nobody intended.
    • Confident but wrong answers
    • Fails on situations it was not trained for
    • Tricked by unusual inputs
  • Misuse Deliberate harm by people.
    • Cloning voices for fraud
    • Writing convincing scam messages
    • Spreading disinformation
    • Attacking computer systems
TypeExampleTypical safeguard
Accident: errorA chatbot states a wrong fact confidentlyCheck important facts against reliable sources
Accident: new situationA crop-advice tool trained on one region fails in anotherTest in each setting before use
Misuse: fraudVoice cloning to impersonate a bank officerVerification steps, awareness, law enforcement
Misuse: disinformationMass-produced fake news postsLabelling, platform action, fact-checking

Automation, jobs and reskilling

AI usually changes tasks within a job before it changes whole jobs. Routine, repetitive tasks - data entry, sorting documents, answering standard queries - are the most exposed. Tasks needing judgement, care, physical skill in varied settings or human trust change more slowly. New work also appears: building, checking and maintaining AI systems, and roles that use AI as a tool. The ethical response is not to deny change or to panic, but to help people adapt through reskilling (learning new skills for a different role) and upskilling (improving skills for the current role).

Levels of human control

LevelHow it worksSuitable for
Human-in-the-loopThe AI suggests; a person reviews and approves before anything happensSerious decisions: benefits, loans, medical treatment, legal matters
Human-on-the-loopThe AI acts on its own while a person monitors and can step inFast, lower-risk tasks such as filtering spam
Human-out-of-the-loopThe AI acts with no human involvementOnly very low-risk tasks

Human oversight works only if it is real. An officer who clicks 'approve' on every AI recommendation without reading it is a rubber stamp, not a safeguard. People tend to trust machine outputs too much - this is called automation bias - so reviewers need time, training and the authority to disagree.

How the world is approaching AI rules

  • Risk-based laws: the European Union has passed an AI Act that sorts AI uses by level of risk. Some uses with unacceptable risk are prohibited, high-risk uses face strict duties such as testing, documentation and human oversight, and lower-risk uses face lighter duties such as telling people they are dealing with AI.
  • Using existing laws plus guidance: many countries apply their existing laws on data, consumer protection, cyber crime and discrimination to AI, and add guidelines and advisories. India's IT Act, 2000 and DPDP Act, 2023 already apply to many AI harms, and policy discussion on further rules continues.
  • International principles: bodies such as the OECD and UNESCO have adopted principles or recommendations on trustworthy and ethical AI that member countries can follow.

Worked example 1: choose the level of control

Suppose a district office uses AI to sort scholarship applications. Working: sorting by completeness is low risk, so the AI can flag missing documents with a person monitoring (human-on-the-loop). Rejecting an application is high risk, so an officer must read the case and decide (human-in-the-loop), and the student must get reasons and a way to appeal.

Match the human control to the risk

  • Sorting by completeness
    • Low risk
    • AI flags missing documents
    • A person monitors: human-on-the-loop
  • Rejecting an application
    • High risk
    • An officer reads the case and decides: human-in-the-loop
    • The student gets reasons and a way to appeal

Worked example 2: who is responsible?

Suppose a clinic uses an AI tool to read scans, and a doctor signs a report without looking, missing a disease. Working: the AI cannot be punished or held to account; it has no legal personality. Responsibility lies with the people and organisation who chose, used and approved its output - here, the doctor and the clinic. The supplier may also share responsibility if the tool was faulty. Saying "the AI decided" does not move responsibility to the machine.

AI can recommend; a person decides and answers for the decision. The higher the risk, the closer the human must stay to the loop.

Key takeaways: AI safety covers both accidents and misuse. Automation changes routine tasks first, so reskilling matters. Serious decisions need a human in the loop who genuinely reviews. Governments use risk-based laws, existing laws with guidance, and international principles. In every approach, people remain accountable.

Check your understanding

  1. 1. What does 'human-in-the-loop' mean?

    1. The AI acts entirely on its own with no person involved
    2. A person reviews and approves the AI's suggestion before a decision takes effect
    3. A human builds the computer hardware
    4. The AI is switched off at night
    Show answer

    Answer: B. In a human-in-the-loop system the AI only suggests, and a person must review and approve before anything happens. A system acting with no person involved is human-out-of-the-loop. Building hardware and switching off at night are unrelated.

  2. 2. Which statement best describes the European Union's AI Act?

    1. It sorts AI uses by level of risk, with stricter duties for higher-risk uses and some uses prohibited
    2. It bans all use of AI in Europe
    3. It applies only to video games
    4. It requires every AI system to be open to the public free of charge
    Show answer

    Answer: A. The EU AI Act takes a risk-based approach: uses with unacceptable risk are prohibited, high-risk uses face strict obligations, and lower-risk uses have lighter duties. It does not ban all AI, is not limited to games and does not make all AI free.

  3. 3. AI tools take over much of the routine data entry in an office. Which response is most in line with responsible AI?

    1. Hide the change from employees
    2. Ban computers in the office
    3. Help employees reskill and upskill for new or changed tasks
    4. Let the AI decide who should be dismissed with no review
    Show answer

    Answer: C. Automation affects routine tasks first, and the ethical response is to help people adapt through reskilling and upskilling. Hiding the change breaks transparency, banning computers ignores reality, and letting AI dismiss people without review removes human oversight.

  4. 4. A government officer approves every AI recommendation to reject applications without reading them. A wrong rejection harms a family. Who is responsible?

    1. The AI software alone
    2. Nobody, because a machine made the decision
    3. The applicant, for applying
    4. The officer and the office that used and approved the AI's output
    Show answer

    Answer: D. AI has no legal personality and cannot be held to account. The people and organisation who chose to use the tool and approved its output remain responsible. Rubber-stamping recommendations is a failure of human oversight, not a transfer of blame.

  5. 5. Which of these is an example of deliberate AI misuse rather than an accidental error?

    1. A weather app that misses a light shower
    2. Using a voice clone to pretend to be a bank officer and ask for an OTP
    3. A translation tool that chooses a slightly awkward word
    4. A map app that suggests a longer route by mistake
    Show answer

    Answer: B. Misuse means someone intends harm; cloning a voice to cheat people is deliberate fraud. The other cases are honest errors by systems trying to do their job - accidents, not misuse.

Privacy · Terms · Refunds · Delivery