Prompting alone could not keep the gpt-3.5-turbo model on-topic, while the text-davinci-002 engine stayed on-topic in the deployed chatbot¶

Claim · All claims
Evidence · 1 study · design r2 · q1
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person · recheck by 2028: tool and model accuracy changes quickly

Subclaims¶

q1 i? Despite prompt instructions to refuse off-topic questions, gpt-3.5-turbo answered off-topic questions after repeated attempts, so the authors switched to text-davinci-002. → Saniya Vahedian Movahed and Fred Martin 2025

Evidence¶

Saniya Vahedian Movahed and Fred Martin 2025¶

Saniya Vahedian Movahed and Fred Martin. (2025). Ask Me Anything: Exploring children's attitudes toward an age-tailored AI-powered chatbot. arXiv. https://arxiv.org/abs/2502.14217 doi:10.48550/arxiv.2502.14217

q1 · i? · design · r2

Design-phase experiments with the ChatGPT API, including varied prompt language, temperature values of 0.2, 0.5, and 0.7, and token settings, showed that "the gpt-3.5-turbo model would answer off-topic questions after repeated attempts"; text-davinci-002 stayed on-topic and was used in the deployed system.

"We discovered that despite our multiple statements to stay on-topic, the gpt-3.5-turbo model would answer off-topic questions after repeated attempts."

Discussion¶

Related Claims¶

• Student-AI conversations under teacher-authored prompts largely stay on the intended task (71% fully on-track; fewer than 1% substantially off-track) — related
• Preference explanations favoring LearnLM more often cite staying on topic, challenging the learner, and not giving away answers — related
• LLM assessment conversations can be derailed off-topic by gaming attempts, and one successful prompt revision was not demonstrated to be replicable — related
• FSM-based evaluation discriminates interactive quality across models, with GPT-5 Mini scoring highest and most tiers significantly separated — related