Adapting for AI: How elementary teachers adjust their practices for an AI-integrated curriculum

What it claims

Three non-CS elementary teachers ran a 13-day ELA+AI-literacy summer camp built on ToyTalk, a no-code platform the authors built. Students photograph a physical toy and turn it into a chatbot, setting its purpose, personality, tone, and rules. The camp served 78 2nd and 3rd graders in a rural, predominantly underserved district in the eastern US, as part of a reading-remediation program for students below grade level.

The most actionable finding in the fetched text is a design change made during deployment. Students wrote vague prompts and got generic answers that ignored the story being read. For example, "how long does a caterpillar stay in a cocoon?" got a general answer instead of one about the caterpillar in the story. In response, the team added pre-filled templates that auto-populate the "role", "job", "communication tones" and "rules" fields. The authors report that this kept responses on-topic even when students asked off-topic questions. They also say it helped students understand that humans control AI behaviour through rules.

How it works

• Curriculum. The three weeks moved through four stages: what AI is, how it generates responses, building and training your own toy chatbot, and judging whether the chatbot's answers are accurate or trustworthy. Each week reused the previous week's artifacts and ELA practices such as description, storytelling, comprehension and discussion.
• Teachers. Normally teach 1st-grade science (14 years of experience), 4th-grade math (4 years) and 5th-grade science (16 years). All had their only AI experience from professional development in 2026.
• Data collected:
• Daily Google-form reflections, about 8–12 entries per teacher.
• Nine 90-minute Zoom group debriefs, transcribed.
• Three 45–60 minute post-camp interviews.
• The protocols are in the appendices.
• Analysis. Inductive thematic analysis, coded by the first author alone and then discussed with the team. This produced 13 codes and 7 themes, including initial challenges, technical issues, diverse student needs, and teacher preparedness.

Limits

• Sample. Three teachers, one district, one summer camp rather than a regular classroom. None of them teach the 2nd/3rd grades the camp served.
• Coding. A single coder; no inter-rater reliability is reported.
• Who ran it. The tool's developers built and supported the deployment and changed the platform mid-camp, so results reflect a moving, researcher-supported system.
• Unfixable failures. Some technical problems were outside the design team's control, such as microphone errors and computers that didn't work. These forced teachers to improvise in the moment.
• Truncated page. The fetched text stops partway into the RQ1 findings. The four practices (repair, differentiation, translation, balancing), the three tensions, and the design implications are not explained in the available text. Read the full paper for those.
• No outcome measures. No student learning results are reported in what was fetched.