CrosscheckSign in

Fight cognitive surrender in 10 minutes a week.

Cognitive surrender is what researchers call accepting an AI's answer with little or no scrutiny[1]. Crosscheck is a short weekly phone call that drills you on your own AI-assisted work, without the AI, so the thinking stays yours.

Get startedRead the research

Illustration: you work with your AI assistant as usual and it sends the work to Crosscheck. Later Crosscheck calls your phone and runs a short drill on that work, without the AI.

How Crosscheck works

  1. 1

    Work with your AI as usual

    Crosscheck picks up the work where your judgment mattered. No typing topics.

  2. 2

    Crosscheck calls you

    A voice coach runs two or three drills. You rate your confidence, then answer out loud, no AI.

  3. 3

    See where you stand

    A graded write-up and a judgment meter: where you were sharp, and where you were on autopilot.

Better, or just faster?

For two years the question has been how to get people using AI. Usage is up and output is up. Almost no one is measuring whether people are getting better, or just faster.

AI delivers performance instantly. It feels like progress, a sugar rush. But ability you borrow from a model isn't ability you own, and on a good day you can't tell the difference.

What the research shows

  • 48%

    higher scores for students practicing math with GPT-4…[2]

  • 17%

    …and lower than students who never had it, once the AI was taken away.[2]

  • 22%

    less likely to catch a flaw when a draft had AI polish. Entry-level reviewers were hit hardest; managers mostly weren't.[3]

  • 15 pts

    lower accuracy when the AI was wrong than when people worked with no AI at all.[1]

The students who leaned on GPT-4 also didn't notice they had learned less. At work, polish hides problems: that's how slop gets through, and people quietly stop thinking. In Atlassian's study, simply labeling a draft “early draft” largely undid the effect[3].

Everyone can buy the same models. The edge is the judgment on top.

If all the thinking goes into the model, your people stop being your advantage.

Effortful learning first. Then the gains compound.

Four ways to learn and perform with AILearning on the vertical axis, performance on the horizontal. Everyone starts at Floundering, low on both. Autopilot performance rises in performance but learning arcs up and falls back. Effortful learning climbs steeply in learning, then continues into Synergetic mastery, high on both. An illustration of a hypothesis, not data.PerformanceLearningFlounderingAutopilotperformanceEffortfullearningSynergeticmastery
Illustration of a hypothesis, not measured data. After Fan et al. (2026)[4].

  1. Floundering

    Researchers mapped how people learn and perform with AI[4]. Everyone starts somewhere new as a beginner: low performance, low learning.

    Without enough support you struggle and don't move. The question is which way you leave.

  2. Autopilot performance

    The easy exit. You offload both the thinking and the checking of your thinking to the AI. Output rises fast. Learning stalls, then slips.

    It's the 48% boost while students had GPT-4, and the 17% drop once it was gone[2]. It feels like progress the whole way.

  3. Effortful learning

    The harder exit. You do the thinking first and use the AI to check and stretch it. Slower and messier at the start, and it feels worse while you do it[6].

    But it's the path where what you can do on your own keeps climbing.

  4. Synergetic mastery

    Keep at it and the researchers expect effortful learning to grow into synergetic mastery: your skill and the AI's add up, instead of one replacing the other.

    That's what Crosscheck is for: 10 minutes of effortful learning on your own work, every week.

2013

FAA Safety Alert 13002

Aviation saw this first. Fly by hand.

After finding more manual-handling errors, the FAA urged airlines to give pilots regular time flying by hand, so the skill is there when the automation isn't[5]. Pilots cross-check their instruments. Crosscheck is the same habit for work you do with AI.

Five checks on your own work. Each with research behind it.

  • Predict

    Call the conclusion, number or outcome before you look.

    Trying to answer first, even when you get it wrong, makes the real answer stick better.[7]

  • Take the other side

    Make the strongest three-point case against where you landed.

    Deliberately considering the opposite corrects biased judgment better than being told to be fair.[8]

  • Spot the flaw

    Find the one error planted in an excerpt of your own AI output.

    Finding and fixing errors in worked examples builds understanding that holds up later.[9]

  • Explain it cold

    Explain the core reasoning in five sentences, no notes.

    People think they understand far more than they can explain. Explaining exposes the gap.[10]

  • Predict the reaction

    Say what the CFO, client or toughest critic will ask first.

    Getting inside the other side's head leads to better outcomes than empathy alone.[11]

What builds judgment feels worse while you do it.

In a randomized physics study, students in active classes learned more but felt they had learned less, because the extra effort felt like failing[6]. Every incentive pulls toward the easy path, especially when the easy path looks like performance. You have to eat your vegetables.

A good practice is

  • WeeklySmall and regular beats a yearly course.
  • On real workDrills come from what you actually did with AI this week.
  • Without the AIYou answer out loud, on a phone call.
  • Low stakesPrivate to you, so you can be honest.
  • Honest about where you standIt shows where you're sharp and where you're on autopilot.

Take your hands off autopilot.

Get started

Sources

  1. 1.Shaw & Nave (2026). Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender. Wharton working paper, SSRN.
  2. 2.Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS 122(26). Figures from the 2024 working paper.
  3. 3.Atlassian Teamwork Lab (2026). AI polish makes it harder to spot problems. But there's a quick fix.
  4. 4.Fan, Yan, Molenaar & Greiff (2026). A metacognitive approach to learning and performance in human-AI interaction. Communications Psychology 4, 150.
  5. 5.Federal Aviation Administration (2013). SAFO 13002: Manual Flight Operations.
  6. 6.Deslauriers, McCarty, Miller, Callaghan & Kestin (2019). Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom. PNAS 116(39).
  7. 7.Richland, Kornell & Kao (2009). The pretesting effect: Do unsuccessful retrieval attempts enhance learning? Journal of Experimental Psychology: Applied 15(3).
  8. 8.Lord, Lepper & Preston (1984). Considering the opposite: A corrective strategy for social judgment. Journal of Personality and Social Psychology 47(6).
  9. 9.Große & Renkl (2007). Finding and fixing errors in worked examples: Can this foster learning outcomes? Learning and Instruction 17(6).
  10. 10.Rozenblit & Keil (2002). The misunderstood limits of folk science: An illusion of explanatory depth. Cognitive Science 26(5).
  11. 11.Galinsky, Maddux, Gilin & White (2008). Why it pays to get inside the head of your opponent. Psychological Science 19(4).

The learning-path picture illustrates a published hypothesis and isn't measured data. Study results come from specific settings (high school math in one school, one workplace review study, lab reasoning tasks) and may not carry over exactly to your work.