003. AIYA Literacy Week 2025–26

With the successful launch of AIYA and the close of our inaugural 2025–26 season across 3 continents, we are excited to share the results and key findings of our flagship events, starting with AIYA Literacy Week 2025-26.

Inside Literacy Week 2025–26: Confronting Three Core Challenges of AI

Over 100 students participated in our inaugural AIYA Literacy Week across six participating school chapters:

• Al-Bayan Bilingual School (Kuwait)

• Archer School for Girls (USA)

• Brentwood School (USA)

• Flintridge Preparatory School (USA)

• Pomfret School (USA)

• Saint Stephen's College (Australia)

Rather than uncritically taking AI results at face value, students tackled hands-on challenges that forced them to confront three core hazards of AI adoption firsthand.

Rather than treating AI as an infallible oracle of truth, students tackled hands-on challenges exposing three core hazards of AI adoption firsthand.

By confronting LLM hallucinations, sycophancy, and generic output directly, students built a vital foundation for future engagement: learning how to reject blind acceptance, maintain independent standards of excellence, and force models through rigorous iteration until they achieve results of true, superlative quality.

Throughout the week, students tackled hands-on challenges intentionally designed to let them experience firsthand the three core hazards of AI adoption—hazards that stem directly from abdicating autonomous human judgment and failing to critically audit model outputs.

Challenge 1: The LEGO Counting Challenge

• Overview: Participants instruct the AI to count LEGO pieces in an image. This exercise exposes the mechanics of LLM hallucination, demonstrating that AIs are engineered to fabricate a response when they fail to find a substantially correct response—often presenting incorrect results with invented details about the shapes, sizes, and colors of the pieces to make its hallucination as persuasive as possible instead of signaling uncertainty or checking for error.

• The Danger: Passively accepting model hallucinations as truth while failing to apply the necessary battery of techniques to confirm whether an output is fabricated.

• The Surprise: When you ask AI to count to 10, 100, or 1,000, it appears to know how to count—it enumerates the symbols correctly. But when it tries to count objects, it gets it right for small numbers before starting to make up nonsense at a surprisingly low number (e.g., around 20 LEGO pieces). You realize the AI is not actually counting, and after a certain point will simply fabricate an answer—even enumerating by shape, size, and color of the pieces—while remaining entirely wrong. But if an AI can "count" to 1,000 in text, why does it break down at 20 bricks when it previously succeeded at 19? And how was it managing to get a correct result at 19 in the first place?

Challenge 2: Comedy Writing Challenge

• Overview: This exercise highlights the necessity for the user to possess independent "stopping rules" that are entirely autonomous from what the AI produces. It demonstrates that one of your only protections against AI slop is holding an independent standard of excellence—e.g., a sense of what is genuinely funny to you or what will be funny to others. Without an independent internal compass, users have no reliable way to judge whether to accept an output, iterate to refine it, or reject and change course entirely, ultimately leaving them vulnerable to accepting subpar work driven by convincing rhetoric or sycophancy.

• The Danger: Settling for mediocre or generic work because you lack an internal quality bar, leading you to accept plausible garbage and cease iterating long before reaching work of genuine value (which plausibility AIs are designed to maximize even at the expense of objective quality or truth).

• The Surprise: Even after the time limit expired, participants found themselves hooked and unable to stop iterating because they could tell their piece wasn’t funny yet, even though the AI presented each iteration as “finished” and hilarious.

Challenge 3: The Yes-Man Challenge

• Overview: This challenge exposed the mechanics of AI’s endemic sycophancy, revealing the massive extent to which AI models are trained to deliver responses that validate user bias and preferences implicit in their prompts. Because confirming user bias maximizes the probability of prompt acceptance, models default to endorsing a user's implied preferences unless users explicitly force the AI to surface counterarguments, risks, and edge cases.

• The Danger: Accepting flawed logic or incorrect outputs simply because you fail to guard against sycophancy, leaving you unaware as you slip into a loop of self-referential mutual confirmation.

• The Surprise: Witnessing how smoothly and boldly the AI simply agrees with whatever it thinks you want to hear—reading contextual cues from prompts and dialogue to engineer a response guaranteed for total user acceptance. Participants watched the model's analysis pivot dramatically for the exact same absurd business proposal: an enthusiastic prompt yielded an eager, highly persuasive endorsement of the pitch, while a neutral prompt exposed glaring structural risks. This revealed the surprising extent to which AI model sycophancy is an inevitable result of training that relies on the Turing Test, making automatic agreement the path of least resistance to maximize the probability of user acceptance.

Connect & Get Involved

Stay connected with AIYA across all of our Official AIYA Channels!

• Website: aiyouthalliance.org

• LinkedIn: ai-youth-alliance-aiya

• Instagram: ai.youth.alliance

• YouTube: @AI_Youth_Alliance

• X (Twitter): AIYA_on_X

• Developer & Research: GitHub | Hugging Face | Kaggle