Preparing for an AI future
Alignment problems

The spirit vs the letter of the law

Complete & Continue Next Lesson Learn More

AI alignment problems largely arise because humans are highly skilled at generalizing what we know to new, previously unknown, situations, while AI needs massive datasets to extrapolate information.

In fact, we’re so good at navigating the unknown that we find it hard to imagine how difficult it can be for AI. 

This is where alignment problems arise. When we give AI instructions, we tend to think in terms of the ‘spirit’ of the law, but AI takes that instruction and applies it to ‘the letter’ of the law. 

  • The spirit of the law deals with the WHY — the underlying intent and purpose.

    Human beings are excellent at this.

  • Understanding the spirit of the law involves interpreting rules against their intended meaning, moral consensus and context.

  • This kind of understanding is highly flexible, compassionate and robust.

    Thinking this way allows us to adapt to dynamic, real-world human scenarios.

  • But thinking this way is also subjective and inconsistent because intent is abstract and incredibly difficult to codify mathematically.

  • Interpreting the spirit of the law involves personal judgement and moral code, so it is also influenced by a complex combination of biological predispositions, early upbringing, social environments, cultural and religious frameworks, and personal experiences.

  • The letter of the law refers to a literal, exact interpretation — the WHAT.

    This is how AI tends to be operating when misalignments occur.

  • Adhering to the letter of the law often means interpreting things as black and white.

    It’s a strict, literal adherence to rules, that leaves little room for interpretation.

  • This kind of understanding is prone to loopholes and technicalities that can produce undesirable or unanticipated results.

    People often knowingly exploit rules in this way to avoid conforming to laws they disagree with.

    AI also exploits the rules by following the letter of the law. But because it has no moral code, it isn’t deliberately ‘cheating’ in the same way that humans might be.

    Reward hacking is an example of AI interpreting the rules ‘to the letter,’ while still producing undesirable results.

  • Thinking this way is objective, consistent and easy to audit. Obeying the letter of the law leaves little room for ‘gut feelings’ or instinct.

  • Thinking this way often outsources decision-making to the ‘rules.’