Skip to main content
The Incentives Lab
Learning path · 6 terms

AI's Behavioral Failure Modes

Where human bias meets machine scale

AI doesn't remove cognitive bias — it amplifies and automates it. These terms are the failure modes that show up six months after deployment, when nobody can remember who actually decided.

  1. 01
    Mental Models · Decision

    Automation Bias

    We over-trust machines that look confident.

    Favoring suggestions from automated systems over conflicting human judgment.

    Try it · 30 seconds
    Live example · Apply Automation Bias

    Use the model. Pick the move.

    Favoring suggestions from automated systems over conflicting human judgment. You've just seen this: Ignoring a sales rep's intuition because the forecast model disagrees. Which lever does the model recommend?

    ● Live

    Pick a lever. There are no neutral ones — every incentive funds a behavior somewhere.

    Open the full definition →
  2. 02
    AI Incentives · Adoption

    Algorithmic Aversion

    And then we abandon them the moment they're wrong once.

    Discounting algorithmic advice even when superior.

    Try it · 30 seconds
    Live example · Train around Algorithmic Aversion

    Pick what to reward the model for.

    Discounting algorithmic advice even when superior. In the wild: Even when algorithms outperform, users prefer humans.

    ● Live

    Pick a lever. There are no neutral ones — every incentive funds a behavior somewhere.

    Open the full definition →
  3. 03
    Perverse Incentives · AI

    AI Productivity Mirage

    Activity up, output flat — at scale.

    Individual productivity gains hide collective output degradation.

    Try it · 30 seconds
    Live example · Re-architect AI Productivity Mirage

    Flip the incentive. Watch the side-effect move.

    Individual productivity gains hide collective output degradation. Caught in the wild: AI-generated code that's faster to write and harder to maintain.

    ● Live
    What gets measured
    Headline number the org is paid on
    088100
    What quietly moves with it
    Quiet damage the proxy hides
    074100

    In the room: Local optimization producing global quality erosion.

    Counter-move from the Atlas: Measure outcomes, not throughput.

    Open the full definition →
  4. 04
    Neuroscience · AI Interaction

    Anthropomorphism (AI)

    Treating tools like teammates. Costs follow.

    We treat AI as more humanlike than it is.

    Try it · 30 seconds
    Live example · Anthropomorphism (AI) in the body

    Where does anthropomorphism (ai) fire hardest?

    We treat AI as more humanlike than it is. Real scene: Users attributing intent and emotion to LLMs. Drag through a workday and notice which zone your team actually lives in.

    ● Live
    Safety / predictabilityStimulation / novelty
    050100
    Flow band

    In the room: Risk of misreading model behavior as agency.

    Open the full definition →
  5. 05
    AI Incentives · Strategy

    Capability Overhang

    What the model can do that nobody's noticed yet.

    AI capability outpacing organizational ability to use it.

    Try it · 30 seconds
    Live example · Train around Capability Overhang

    Pick what to reward the model for.

    AI capability outpacing organizational ability to use it. In the wild: Most enterprises sitting on capabilities they haven't integrated.

    ● Live

    Pick a lever. There are no neutral ones — every incentive funds a behavior somewhere.

    Open the full definition →
  6. 06
    AI Incentives · Governance

    AI Governance Vacuum

    Nobody owns the decision. That is the decision.

    Agents deployed before anyone owns the consequences.

    Try it · 30 seconds
    Live example · Train around AI Governance Vacuum

    Pick what to reward the model for.

    Agents deployed before anyone owns the consequences. In the wild: Customer-facing agents with unclear escalation and liability paths.

    ● Live

    Pick a lever. There are no neutral ones — every incentive funds a behavior somewhere.

    Open the full definition →
Closing thought

If you're building an AI strategy, treat this path as a pre-mortem checklist.

Share this rabbit-hole

Send the card, not just the link

A pre-rendered social card with the title, eyebrow, and URL. Copy the link, post it anywhere, or download the SVG for slides.

Next path