Skip to content
EN

Lesson 7 of 8 Β· 6 min Β· +100 XP

Is AI dangerous, and can we keep it under control?

AI can be dangerous when systems don't do what people intend, from unfair decisions to misuse. AI safety is the work of preventing that: alignment (teaching AI to follow human goals and values), testing for dangerous abilities before release, human oversight and rules. Experts disagree about how big future risks are, so keep humans in charge of important decisions.

β–Ά Try it

Why it matters

AI safety sounds like science fiction, but most of it is careful engineering: checking that systems do what we meant, finding flaws before release and keeping people in charge. Some big questions are genuinely open, and it's fine to say so.

In 6 minutes you'll be able to

  • Alignment means AI pursues what we meant, not just what we typed.
  • Real harms exist today; bigger future risks are debated, not settled.
  • Support testing and oversight, and keep humans in charge of key decisions.

What you'll practise

  1. Safety words

    Tap each card to see what safety researchers mean.

    🎯 Alignment
    Making an AI pursue the goals people actually intend, not just the literal instructions or a score.
    🏁 Specification gaming
    When AI hits the target in a way nobody wanted. A game-playing AI once learned to spin in circles collecting points instead of finishing the race.
    🧨 Red teaming
    Experts try hard to make an AI misbehave before release, for example to get dangerous instructions out of it.
    πŸ§‘β€βœˆοΈ Human oversight
    People can monitor, correct or switch off an AI system, and are responsible for important decisions.
    πŸ”“ Jailbreak
    A trick prompt that gets around an AI's safety rules. Safeguards help, but none are perfect yet.

  2. Now, debated or myth?

    Sort each statement about AI risk.

    • Chatbots sometimes give confident, wrong answers β†’ Happening now
    • AI-made voices and videos are used in scams β†’ Happening now
    • AI can repeat bias from its training data β†’ Happening now
    • Very capable AI could pursue goals people didn't intend β†’ Debated future risk
    • Humans could lose meaningful control over advanced AI β†’ Debated future risk
    • Today's chatbots secretly hate humans and plot revenge β†’ Myth
    • AI safety is just marketing, there are no real risks β†’ Myth

  3. A balanced view

    No hype, no doom. What's the most accurate answer?

    1. Who checks whether powerful AI is safe?

      • The companies, government AI institutes and independent researchers (Answer)
      • Nobody at all

    2. Do experts agree on how risky future AI will be?

      • No, views differ widely, which is why evidence and testing matter (Answer)
      • Yes, they all agree AI will end the world
      • Yes, they all agree there's nothing to worry about

    3. Lumi refuses to explain how to make a dangerous chemical. What is that?

      • A safeguard added during safety training (Answer)
      • Lumi being moody

    4. What can you do to keep AI under control in daily life?

      • Keep a human in charge of important decisions and check output (Answer)
      • Nothing, it's only for governments

Remember

Safe AI doesn't happen by luck: it's tested, overseen and questioned. And some questions are still open.

Key facts

  • The first International AI Safety Report, led by Turing Award winner Yoshua Bengio and backed by 30 countries plus the UN, EU and OECD, was published in January 2025. Β· Source: International AI Safety Report, 2025
  • The second International AI Safety Report was published on 3 February 2026, written by over 100 independent experts. It sets out evidence on capabilities and risks without recommending specific policies. Β· Source: International AI Safety Report, 2026
  • A November 2025 update found the number of companies publishing frontier AI safety frameworks had more than doubled, but sophisticated attackers can often bypass current safeguards. Β· Source: International AI Safety Report, Key Update, November 2025
  • At the Bletchley Park summit in November 2023, 28 countries and the EU signed a declaration agreeing that risks from frontier AI need international cooperation. Β· Source: UK Government, The Bletchley Declaration, 2023
  • The UK's AI Safety Institute, set up in 2023 to test advanced AI, was renamed the AI Security Institute in February 2025. Β· Source: UK Department for Science, Innovation and Technology, 2025

Updated:

Questions people ask

What is AI alignment?

Alignment is making AI systems pursue the goals and values people actually intend. A misaligned system might follow instructions too literally or chase a score in ways nobody wanted.

What is AI safety?

AI safety is the field that works to make AI systems do what people intend without causing harm. It covers alignment research, testing before release, human oversight and regulation.

What is red teaming in AI?

Red teaming means experts deliberately try to make an AI misbehave, for example to produce harmful content, so that weaknesses are found and fixed before the public uses it.

Who regulates AI safety?

It depends on the country. The EU has the AI Act, while the UK and US rely mainly on existing regulators and government AI institutes that test advanced models.

Related terms