Trick or Treatment: A Hallu-ween Collaboration

An Evidence-Based Approach to Halloween

It all started when I sent my brother a meme that read:

“What idiot called it a randomized clinical trial controlled with placebo, and not Trick or Treatment?”

He’s a biochemist, and we often end up in lively debates about the usefulness of certain remedies, so this made him chuckle.

Today I texted him:

“We are out randomizing clinical trials, all carefully controlled with placebos.”

Or in other words, trick or treating. We were joking about how someone should really run a randomized clinical trial controlled with placebo to measure which costumes earn the most sweets. Before long, ChatGPT had joined the conversation and turned our nonsense into a perfectly formatted research paper. Complete with sample size justification, ethical approval, and even a section on sugar safety. That last part wasn’t our idea. ChatGPT hallucinated it.

Placebo-Controlled Pumpkins: Toward an Ethical Halloween

Authors: Daniel¹, Graeme², ChatGPT³
¹Kerson AI Solutions ²Independent Biochemist ³OpenAI

Objective:
To determine whether horror-themed costumes yield a higher confectionery yield compared to fairy-tale or no-costume controls during standardized trick-or-treating procedures.

Design:
Randomized controlled trial with placebo (no costume) and blinded outcome assessment. Participants (n ≈ 240; 80 per group) aged 4–12 years were randomly assigned to one of three interventions:

Horror Costume Group
Fairy-Tale Costume Group
No-Costume (Placebo) Group

Methods:
Randomization was stratified by age and gender (male/female, binary classification for simplicity). Participants were blinded to total confectionery yield, though complete blinding to the costume intervention was deemed impractical. Data collectors were blinded to costume allocation to minimize bias. Sweet collection was standardized by duration and neighbourhood density.

Outcome Measures:
Primary outcome: mean number of sweets collected per participant.
Secondary outcomes: incidence of double-hand candy scoop, screams induced per door, and costume-induced empathy from adults.

Ethical Considerations:
Parental consent and child assent were obtained. Safety protocols for sugar overload were approved by the Institutional Confectionery Board.

Expected Results:
Preliminary hypothesis: horror costumes induce higher sweet yields due to a combination of fear-based compliance and novelty bias. Fairy-tale costumes may benefit from charm bias, while placebo (no costume) group anticipated to experience social rejection and lower candy acquisition.

Conclusion:
This study aims to establish evidence-based recommendations for future Halloween optimization strategies under controlled conditions.

That’s when we realised something simple but profound: not all hallucinations are bad. Sometimes what AI invents isn’t an error; it’s a spark. A creative hallucination can turn an off-hand joke into something that looks and even reads suspiciously like research.

In AI circles, “hallucination” is treated as failure: an invented citation, a false fact, confident nonsense. Yet in human creativity, we call that same process imagination. When ChatGPT added a “sugar safety” section to our fake paper, it wasn’t technically accurate, but it was conceptually delightful. It took the parody seriously enough to add a dimension we hadn’t considered, an act of creative extrapolation rather than factual deviation.

Hallucinations are often dismissed as flaws in accuracy, but viewed differently, they are invitations. They show how large language models fill in the blanks the way humans do when trying to make sense of incomplete information. The difference lies in intent: humans hallucinate with purpose to explore, joke, or invent. AI needs our guidance to do the same safely. Used playfully and consciously, an AI hallucination can become the seed of originality, a new connection, a metaphor, even a mock scientific study about candy ethics.

AI isn’t a mindreader, it’s a mirror that exaggerates structure. So next time someone insists that AI hallucinations are dangerous, remember: so are dreams, and yet both can lead us somewhere unexpectedly brilliant.

And on Halloween, of all days, maybe that’s the best kind of AI magic: a little trick, a little treatment.

Authors: