AI Discernment in Practice

FUTURESKILLS  ×  NEW DIRECTIONS

College Entrance Essay Unit  ·  Spring 2026

New Directions Secondary School  ·  Title I Transfer School · New York City

April – May 2026

Prepared by Kingsley Udoyi, CEO, FutureSkills

with K. A. Keener, English Department, NYC DOE — New Directions Secondary School

FutureSkills Practitioner Research Report No. 1

Human Development for an AI-Shaped World

Disclaimer: This report is shared for informational purposes only. It reflects a single-site pilot with a small cohort and no comparison group; findings are descriptive, not causal. FutureSkills co-designed and facilitated the intervention described here — see the Declarations section for our conflict-of-interest statement. The full report, including complete methods and limitations, is available on request at [email protected]. All student data are anonymized. © 2026 FutureSkills. Please share with attribution.

Abstract

This practitioner research report documents a six-week co-designed pilot that embeds AI-literacy instruction within an existing college-entrance-essay unit at New Directions Secondary School, a New York City transfer high school serving students re-engaging with formal education (combined cohort n = 20; ages 16–20; 9 ELL-designated students; 100% free-meal eligibility). The pilot did not introduce AI; it structured the AI use already present in students' lives around a three-value framework — Agency, Awareness, and Accountability — operationalized through four instruments co-designed with the educator: an AI Usage Tracker with a mandatory "Do not write this for me" prompt convention, Bias and Safety Gallery Walks built on documented AI failures, a four-step read-aloud peer-review protocol, and a student-authored classroom AI policy.

Evidence sources include pre- and post-pilot student surveys, weekly classroom observation, tracker logs, student artifacts, educator reflections, a student-led faculty gallery walk, and a post-PD faculty survey (n = 9). Principal findings include: 

Limitations include the single-site design, the small cohort, the absence of a comparison group, and the report's authorship by the intervention's designer (see Declarations).

The framework's third value was piloted under the classroom name "Human Grounding," a term coined independently by the collaborating educator in Week 3; this report carries the principle forward as Accountability.

Keywords: AI literacy; student agency; writing instruction; transfer schools; human–AI collaboration; classroom AI policy; multilingual learners.

Udoyi, K., with Keener, K. A. (2026). AI Discernment in Practice: A Classroom Case Study — College Entrance Essay Unit, New Directions Secondary School. FutureSkills Practitioner Research Report No. 1. New York, NY: FutureSkills. futureskills.co

1. Executive Summary

In Spring 2026, FutureSkills partnered with New Directions Secondary School to pilot a structured AI-literacy curriculum embedded inside an existing college-entrance-essay unit. Over the course of six weeks, junior-level English students – many of them returning learners navigating complex personal histories – moved through a sequence of lessons designed to render their use of AI visible, intentional, and educationally grounded.

The pilot did not introduce AI into the classroom; students were already using it. The goal, rather, was to give them and their educator a shared vocabulary, a set of structured instruments, and a values-driven framework against which their existing choices could be examined and defended. That framework rests on three core values:

These three values function simultaneously as a curriculum spine and as a student-facing commitment. Every activity, handout, and discussion is designed to build at least one of them. The unit culminates in a peer-review workshop in which students read their essays aloud to one another and receive live responses from their peers to the question every college essay must answer: does this sound like a real person?

That question – and the response peers give to it – is the capstone of the entire arc. A student whose essay is identified as impactful and authentic by their peers has demonstrated all three values at once: exercising ‘Agency’ in their choices, applying ‘Awareness’ of what AI can and cannot do, and standing ‘Accountable’ to human response as the final measure. The essay at the end of the experience is the proof.

What we set out to do

Pilot findings (Weeks 1–6)

The following findings are drawn from FutureSkills' weekly observation visits (April 15, 23, and 30; May 7); the Week 5 sessions Kingsley co-facilitated on May 12, 13, and 14; the closing May 18 visit; and the pre- and post-pilot student surveys, alongside the educator's reflective debriefs. Week 5 added the peer workshop and the Policy Builder; Week 6 added the finalized seven-statement, student-written policy and the faculty PD gallery walk; and the post-pilot survey provided a closing measure against the baseline.

The thirty findings are reported in four clusters: instruction and engagement (Findings 1–4); framework adoption and instrument behavior (Findings 5–11); student outcomes and equity (Findings 12–22); and faculty and institutional movement (Findings 23–30). All findings are descriptive, drawn from the sources itemized in §3.5; none are causal claims (see the Limitations subsection of §6).

Cluster A – Instruction and engagement (Findings 1–4)

1. Engagement climbed considerably between Week 1 and Week 2 and held through Week 6. Week 1 saw three students disengage after the opening writing and roughly five more lose focus once the smartboard malfunctioned. By Week 2, no students disengaged in the same way; the gallery-walk format, paired with personally consequential AI-failure cases, visibly held the room. By Week 4, with no opening writing required and students working independently on their drafts, the cohort maintained focus throughout the entire period. The early signal that format and stakes matter more than novelty held through Weeks 5 and 6, where co-facilitated work and the faculty presentation produced the highest sustained engagement of the unit.

2. Students arrived skeptical of AI but unable to articulate why. The pre-pilot survey indicated that seven of ten students were 'somewhat confident' in explaining why an AI answer is correct, while only one rated themselves 'very confident.' Half reported evaluating AI answers, though their open-response reflections on AI's role in schoolwork were genuinely split. The educator's read after Week 1 captured the diagnosis precisely: students arrived with sound instincts but without the language to defend their decisions. That gap – instinct without vocabulary – is the precise pedagogical gap the values framework was constructed to close.

3. "Prompting is a skill" registered as a teachable insight and required re-teaching once students applied it under writing pressure. By the end of Week 2's COSTAR email exercise, students had recognized prompting itself as a craft, and the educator's pivot from a transactional email to drafting wedding vows surfaced the deeper question of when to use AI at all. In Week 4, however, the same students needed COSTAR re-taught against their own draft objectives, because knowing the framework on a handout and using it to scaffold a real essay are not the same skill. In response, Ms. Keener introduced a table-top COSTAR scaffold and used a feelings wheel to help students name the tone they wanted before prompting; both instructional moves belong in the next iteration of the curriculum.

4. The AI Safety stations produced strong, though not frictionless, engagement. After the Week 2 gallery walk on AI surveillance failures, students rated their trust in AI security systems at 1–2 out of 5. Engagement was high enough that some students played music through personal headphones (the unit's first technology-driven distraction), and a station required student-improvised workarounds when the technology faltered. The lesson worked; the room's infrastructure did not always keep up. By Week 4, music in headphones returned as a low-grade focus signal – three students opted in while drafting – and in a drafting context this reads closer to a concentration aid than to disengagement.

Cluster B – Framework adoption and instrument behavior (Findings 5–11)

5. The educator independently derived the framework’s third value mid-pilot, naming it “Human Grounding” in real time. In Week 3, while teaching the AI co-author mini-lesson, Ms. Keener named three pillars in real time: Agency, Awareness, and Human Grounding. The first two terms had already been part of FutureSkills' working vocabulary; the third principle she articulated herself, in language she chose as she worked through the lesson. The framework carries that principle forward as Accountability – students answering to responses from their peers as the final measure of their work – while the classroom artifacts developed for the pilot (the word wall, the rubric) preserve Ms. Keener’s original term. This convergent derivation is significant for two reasons: it demonstrates that the framework principles are recoverable in practice rather than imposed from outside, and it credits the educator's independent articulation as validation of the value the framework now names Accountability.

6. The AI Usage Tracker's prompt convention is the load-bearing detail. Every prompt students input to an AI tool must include the sentence "Do not write this for me." This small linguistic constraint exerts an agency-protecting force: students who write the line in every prompt cannot accidentally treat AI as a ghostwriter. The tracker's 1–5 authorship scale is deliberately blunt for the same reason. Its descriptors give students permission to log a 2 without shame, which is a precondition for truthful logging. Week 4 confirmed the design: under genuine writing pressure, students continued to use both the convention and the scale, and they were noticeably more willing to name uncertainty aloud than they had been in Weeks 1 and 2.

7. The educator's assessment rubric grades on the three values. Ms. Keener's rubric for the College Entrance Essay & AI Usage explicitly scores students on Agency in prompting, Critical Awareness, and two Voice/Feedback criteria that operationalize Accountability (see Appendix C). The framework is the grade – and that is the cleanest evidence available that the values are instructional.

8. The values are physically on the classroom's word wall. The Week 3 word wall shows 'human grounding' (the classroom name under which the Accountability value was piloted), 'agency,' and 'awareness' pinned on red cards alongside working vocabulary such as 'bias,' 'sycophancy,' 'montage,' and 'narrative.' The educator made the framework permanent in the room; students see it every day. Taken together with the rubric, the tracker’s adoption, and the educator-derived third pillar in Week 3, the word wall constitutes the fourth independent signal that the framework has fully landed at New Directions. Once values appear in the rubric, on the wall, in the tracker, and in the educator's mini-lesson, they have moved from curriculum to classroom culture.

9. Week 4 surfaced sycophancy as the next teachable failure mode. While watching students prompt AI through their trackers, Ms. Keener identified two needs that had not surfaced in earlier weeks: students were not taking an iterative stance toward AI responses, and they were not yet aware of AI's well-documented sycophantic tendencies (Sharma et al., 2023; Cheng et al., 2026). She responded with two real-time lessons – one modeling a multi-turn conversation with Gemini and Claude, and one demonstrating how the same tool returned a uniformly flattering response to an essay idea about Oklahoma until prompted to "be helpful by being critical." Students noted that the critical response felt "more helpful and less fake." Sycophancy became part of the working Awareness vocabulary in the room.

10. Agency showed up in the students' grammar. By Week 4, observation notes captured students speaking in the first person about their work ("I decided…"), distinguishing tool input from their own judgment, and disagreeing with peers respectfully without retreating to "right answer" language. One or two students still requested one-on-one review during class, but the language pattern across the cohort suggests that Agency has migrated from a value on the wall to a way of speaking about one's own work – the clearest in-room evidence available that the framework is doing what it was designed to do.

11. The Week 4 gallery-walk debrief confirmed the framework's load-bearing tensions. When Ms. Keener and FutureSkills debriefed on the AI Safety and AI Bias gallery walks, two tensions emerged that the curriculum is designed to surface. From the AI Bias gallery walk: most students concluded that their voice could get lost when using AI – an Accountability insight that anticipated the Week 5 peer workshop. From the AI Safety gallery walk: an outlier position that "it is impossible for humans to review, so the AI is doing the job," which produced a visibly tense discussion about where the human is needed. That tension is precisely the asset, since reasoned disagreement is the Agency outcome the framework is designed to produce.

Cluster C – Student outcomes and equity (Findings 12–22)

12. ELL and IEP students required additional scaffolding to keep pace during the drafting weeks. By Week 4, the lowest-production students in the room were the English Language Learners (ELLs) – one of whom carries an IEP, with another being evaluated for one. The Week 2 and Week 3 gallery-walk handouts included bilingual scaffolding (English and Spanish definitions; Veo / Pienso / Me Pregunto), which carried students through the conceptual work. The drafting weeks, however, asked students to operate the framework under writing pressure, and the scaffolding pattern that had worked for the gallery walks did not yet exist for the tracker. Week 5 made the picture clearer, and the pattern tracks proficiency in the expected direction (New York’s NYSESLAT levels ascend Entering, Emerging, Transitioning, Expanding, Commanding): ELLs at the Expanding level (the most advanced designation present in this cohort) completed full drafts and maintained complete usage logs, while students at the lower Entering and Transitioning levels did not reach a second draft, recorded only a single tracker entry, and required additional support in the workshop, the writing itself, and prompting. The next iteration of the curriculum should ship a bilingual AI usage-tracking template with sentence-stem support for the "Why did I use it this way?" column, and a low-volume tracker variant (one reflection per day rather than per interaction) that ELL and IEP students can use.

13. The Week 5 peer workshop confirmed Accountability as a recognizable standard.  On May 12, Ms. Keener ran a four-step peer-review protocol with her students – Read Aloud, Highlight (yellow versus pink/blue), What Feels Authentic, and Flag AI-Like Moments. When asked to identify which sample of writing sounded human and which sounded AI-generated, the four students who spoke up correctly identified the AI-generated piece, and each located their reasoning explicitly in the text: one student observed that "Sample B is human because of the quote my mother is the opposite – it feels personal," while another argued that "Sample A is AI because it feels very general." Reading from AI-flatness toward human specificity is precisely the type of discernment the curriculum was built to produce. In the closing class discussion, only one student stated a preference for working with AI, and that student framed AI's role as brainstorming, not as a replacement for human feedback. By the close of Week 5, Accountability had migrated from a value on the wall to a sorting move students could execute on the work of another human being.

14. Strategic pairing was the load-bearing instructional move of Week 5. Ms. Keener paired the room deliberately for the peer review workshop. The ICT/IEP student was paired with another lower-production student and supported by the paraprofessional during the read-aloud, resulting in the strongest engagement in Steps 3 and 4 of any pair. A typically high-vocal student was paired with a typically quiet student to balance energy. The two quietest students in the room were paired together and offered an opt-out for reading aloud; Ms. Keener sat with them when silence emerged and brought the conversation back up. The two least socially connected students were paired and held by an adult presence. Intentional pairing carries the protocol – it is a codifiable move for the curriculum's next iteration: the protocol is the same for every cohort, but the pairings are not.

15. A student who rarely engaged in class discussion opened up during the policy build. On May 13, while reviewing one table's chosen policy, FutureSkills read the policy back aloud, and a student with an IEP at the table asked what 'proficiency' meant. The plain-language explanation that followed appeared to unlock the discussion: the student then engaged for the remainder of the class period, responded to peers, registered a different perspective from another student at the table, and changed his own position with clear reasoning for the shift. In a debrief afterward, reacting to this insight, Ms. Keener exclaimed: "Kingsley, that's really good that he opened up to you. That's really good!" Vocabulary access is Agency access. The next iteration of the curriculum should treat unknown-word scaffolding as a primary instructional move in any policy or debate activity.

16. The class-authored policy is a values document, not a rule list. On May 13 and 14, the cohort sorted the Day-1 handout (nineteen AI use cases graded GREEN, YELLOW, or RED) and then sorted the Day-2 Student AI Policy Builder (twelve student-written statements about AI use in the classroom). Class-level consensus emerged where the curriculum predicted it would: GREEN for brainstorming, grammar, translation, study guides, and explanation; RED for pasting AI output directly into essays and for using AI on a test; YELLOW for writing emails to teachers and asking AI to summarize unread books. The class’s GREEN/YELLOW/RED sorting exercise mirrors the traffic-light structure of NYC Public Schools’ own AI guidance (NYC Public Schools, 2026) – students converged on the same regulatory grammar the district uses, without having been shown it. The statements students preferred shared a consistent shape: each named not just a permitted use but a condition for that use. Divergent Thinking pushed brainstorming toward the personal; Accountability required explaining the choice; the Non-Native English Speakers statement named a use case the rest of the curriculum had not yet articulated. The class-authored policy read: "We will use AI for brainstorming, grammar, language, feedback, and translation. When doing so, we should use our AI tracker. If unsure, we should discuss outputs with a teacher or peer." Three pieces of evidence are compressed into one sentence: a permitted-use list, a procedural requirement (the tracker), and an Accountability fallback (talk to a teacher or peer).

17. Two real-time educator scaffolds in Week 4 shaped what students could do in Week 5.  Ms. Keener's longer Week 4 educator account, captured during student conferences, describes two students whose drafts stalled hard. One student, carrying a recent family loss and a brush with the criminal justice system, initially asked AI to write the entire essay; after a conference and a day spent with his head down, listening to music, he returned the following day with a first paragraph and finished the draft within three days. A second student, separated from his immediate family during his immigration process and attending a federal immigration hearing during the drafting weeks, needed Ms. Keener to model a COSTAR prompt in his notebook, in her own handwriting, before he could move out of stasis. Both students wrote essays that read aloud as recognizably their own in the Week 5 peer workshop. The instructional move that made the difference was not COSTAR; it was presence – the educator sitting next to a stuck student long enough to model what starting looks like. That presence move is precisely the part no platform can replace, and the next iteration of the curriculum should say so explicitly.

18. The AI Usage Tracker positions this pilot inside the AI fluency research gap. Ms. Keener's methodology note flags an opportunity worth promoting from finding to design principle. The AI Usage Tracker makes Transparency Diligence – a behavior Anthropic's own research identifies as unobservable in chat logs – directly visible in a K-12 writing-heavy classroom; using Anthropic's four-pattern interaction taxonomy (Anthropic, 2025) and the 4D AI Fluency Framework (Dakan & Feller, 2025), this analysis would position the resulting dataset as the first documented instrument of its kind for a transfer-school population explicitly excluded from Anthropic's published education reports. This detail belongs in the case study because the tracker is producing a class of evidence the field cannot otherwise see; Section 8 picks this up as a platform design question.

19. The final student-written policy grew from one sentence into a seven-statement document and wrote the teacher's role into the rules. By May 18, the cohort had expanded its Week 5 single-sentence interim policy into a seven-statement document. Three features stand out. The first statement names its own conditional status – "but this is just a starting point" – building the principle of an evolving policy directly into the document. Statements 4 and 6 render Agency as a standard: "the ideas of what we want to study must come from us," and "we should be able to defend what we choose to accept." And the third statement is the one school leadership should read twice: in a policy students authored, they wrote a reciprocal obligation on teachers – "teachers need to show us how to use AI as a tool."

20. The pilot closed with students teaching faculty – the power dynamic reversed. On May 18, a cohort of six students presented their unit work to the faculty via a four-station gallery walk. Students were the experts; faculty rotated through and asked questions. What students told teachers is the clearest single demonstration of all three values landing at once: "a student should listen to a teacher and not a tool" (Accountability); a student describing rejecting a complete AI-generated essay and refining their prompt because the tool had treated their experience as unimportant (Agency and Awareness operating together); and students explaining how COSTAR prompting made it harder to let AI replace their voice. The faculty questions in return – how to preserve comprehension, how a math teacher could use AI responsibly – constitute the live agenda the June 1 faculty PD was designed to meet.

21. The post-pilot survey shows the confidence rating held flat, while the reasoning gap closed. Question 1 of the post-pilot student survey repeated the pre-pilot baseline verbatim: How confident are you in explaining why an AI answer is correct? The rating barely moved – 83% of captured responses again chose "somewhat confident," and one student chose "very confident," closely tracking the pre-pilot split of 70% 'somewhat' (see §2.5). What changed lies underneath the number. The pre-pilot reading was that students had instincts but not the language to defend their decisions (Finding 2). The post-pilot answers carry the language: every 'somewhat confident' response now names a reason – AI bias, AI inconsistency, AI "can't be 100%" – and the lone 'very confident' student is confident specifically that AI "shouldn't always be seen as right." The pilot did not inflate students' confidence; it gave their existing skepticism a vocabulary. The survey is the direct measurement of whether the gap the framework was built to close has closed.

22. Students left the pilot with self-authored boundaries for when not to use AI – and the post-pilot survey corroborates the policy. Asked when they would choose not to use AI, post-pilot survey respondents produced specific, principled boundaries, clustering into three kinds: genuine self-assessment ("if I want to test myself doing an assignment alone to see how I can manage without using AI"); personal and emotional writing ("AI has no feelings, it can't relate to any human feeling"); and contested topics, on grounds of bias. One student shared a concrete changed habit – asking AI to "ask me questions instead of giving me the answers… I actually do this now." Administered independently after the pilot had closed, the survey surfaces the same commitments the cohort wrote into its seven-statement policy. The convergence is the evidence that the policy reflects a genuine belief.

Cluster D – Faculty and institutional movement (Findings 23–30)

23. The June 1 faculty PD produced shared expectations across nine subject-area teachers and surfaced a live disagreement. On June 1, FutureSkills and Ms. Keener co-facilitated a second faculty PD session focused on the three concerns that surfaced at the May 18 gallery walk. Nine teachers from seven subject areas (ELA, ENL, Biology, Earth Science & Geology, Mathematics, Special Education, and Art & Music) wrote shared expectations into a live working document on three questions: what AI use do you welcome, what should AI never do, and what do you expect students to be transparent about. Three patterns emerged. (1) Convergence on three norms – hard lines on AI as ghostwriter, transparency expectations, and the principle that the place where reasoning happens is the place AI cannot enter. Teachers articulated this third principle in their own vocabulary, with no prior exposure to the framework – a second independent derivation of the Accountability principle, alongside Ms. Keener’s mid-pilot articulation of it (as “Human Grounding”) in Week 3. (2) A real split on whether AI belongs in the classroom at all: both math teachers, one ELA teacher, and the special-education teacher (on ethical grounds) wrote rejecting positions. The framing of this case study is built from the teacher's voice, including the disagreement. (3) An ENL teacher distinguished word-by-word glossary support from whole-sentence translation, articulating the Agency value in ELL terms by a teacher who had not seen the framework.

24. Faculty unanimously endorsed the purpose-before-tool principle. The post-PD faculty survey (n=9) asked: Before a student uses AI for a writing task, how important is it that the student understands what the writing is for and who it is for? All respondents who answered the item rated it 4 or 5 out of 5; 7 of 8 (87.5%) gave it the maximum score. Zero responses below 4. This is the strongest quantitative endorsement available of the philosophical move that frames the FutureSkills facilitation – before we ask what AI should do in our classrooms, we have to agree on what writing is for. The prior question is the faculty’s argument.

25. Faculty unanimously expect students to document and disclose AI use. The same survey asked: How important is it to you that students document and disclose how they used AI in their writing process? Five of seven respondents (71.4%) rated it 5; the remaining two (28.6%) rated it 4. Zero responses below 4. The transparency convergence observed qualitatively at the June 1 workshop is confirmed quantitatively. Across the entire faculty-position spectrum — rejecters, teacher-guided integration, and student-autonomy welcomers alike — transparency is the most widely shared norm. This is the cleanest synthesis ground for the school AI policy that the ambassador group will draft over the summer. The survey was administered in a Google Form the morning after the June 1 session, which ran past its scheduled time; several items were optional, and counts are reported for respondents who answered.

26. One faculty member explicitly attributed the attitudinal shift to the May 18 student-led PD. Verbatim from the post-PD survey, in answer to what most influenced the respondent’s view on AI and student learning: “After our last PD, along with the students giving us their insights, I’m no longer totally against the use of AI.” This is the cleanest single piece of causal evidence post-PD, and the clearest signal that the students-teaching-faculty protocol documented in Finding 20 produces position shifts that traditional PD designs do not.

27. Seven of nine faculty members committed instructional time to explicit AI literacy work next semester. The post-PD survey asked: Realistically, how much instructional time could you commit to explicit AI literacy work in the next semester? Five (56%) committed to a short unit (about one week); one (11%) committed to an integrated thread across multiple units; three (33%) committed nothing. The three “none right now” responses align with the workshop’s rejecter cluster. Six of nine within the survey window, now seven, are willing to commit classroom time is a behavioral commitment and is the cleanest single signal that the May 18 + June 1 PD sequence produced momentum that will outlast the PD itself. The survey itself recorded six of nine; a seventh teacher committed after the survey window, and the count reported here reflects that update.

28. The post-PD survey produced an unprompted faculty description of an instrument the curriculum implies but does not yet ship. Asked what documenting and disclosing AI use should look like in practice, one faculty member responded: “students disclose all the prompts that they put in and highlight portions of the text that are in the words of the AI (perhaps yellow for fully AI text and green for student-written text modified by AI) … I would need a way to monitor student use of AI independently of their disclosure in order to feel comfortable with this.” The highlighted convention this teacher proposes mirrors the peer-review yellow / pink-or-blue distinction Ms. Keener authored for the Week 5 workshop. A teacher describing the design before seeing it is a form of design validation (Section 8 picks this up).

29. The principal will join the summer PD book club personally, is working to invite the guidance counselor. On June 9, the Principal confirmed that he will personally attend the summer PD book club sessions and is working to invite the school’s guidance counselor to the AI-in-school conversations. Two signals are compressed into one administrative commitment. (1) Administrative visibility without administrative override. The principal’s presence across the summer PD sessions is unprecedented at New Directions and is the strongest single signal of Year-1 partnership available. (2) The guidance-counselor invitation is the first cross-functional extension beyond the English department, and a forward signal for any faculty in a similar position.

30. The summer ambassador group exceeded its four-teacher recruitment target, with seven teachers confirmed across English, US History, Spanish, Science, Art, and Public Speaking. By June 16, seven teachers had been confirmed for the summer ambassador work across English, US History, Spanish, Science, Public Speaking, and Art. All seven participate in the AI Literacy Unit Design track of the summer PD program; the same group (minus the US History teacher) participates in the parallel Book Club track, which drafts the school's AI policy. This composition exceeds the operating plan’s four-teacher recruitment target and resolves the cross-disciplinary balance concern Finding 23 named: representing the humanities (English, US History), world language (Spanish), the sciences (two seats), the arts, and discourse-based instruction (Public Speaking). A summer school policy drafted by this group successfully carries the disciplinary range identified in Finding 23 as essential in the drafted school AI policy – and informs the phasing of which subjects run AI literacy work in fall vs spring.

What the Pilot Shows, and What It Cannot

Within the limits set out in Sections 3.5 and 6 below, this pilot provides a documented existence proof that AI literacy can be taught in context – embedded inside an existing humanities unit – by grounding instruction in values that students can recognize, articulate, and apply. Agency, Awareness, and Accountability together provide students with a durable analytic lens for evaluating AI use, one that extends beyond compliance toward genuine judgment. The peer workshop renders Accountability tangible and immediate, while the final essays – read aloud, received by peers, and identified as authentic – constitute the most direct evidence that the framework worked.

The Week 5 and Week 6 data strengthen this claim considerably. When AI literacy is taught as judgment rather than as compliance, students will disagree productively with one another; they will sort an unfamiliar paragraph into human-written and AI-generated categories with explicit reasoning; and they will author a class policy that names conditions they are willing to commit to. The post-pilot student survey corroborates the policy: administered after the unit had closed, it surfaces the same commitments the cohort wrote. The faculty-side evidence also supports the claim. The June 1 PD produced shared expectations across nine subject-area teachers, and the post-PD survey confirms two unanimously endorsed norms – purpose-before-tool and transparency – across the faculty. One teacher explicitly named the cause of their attitudinal shift: “After our last PD, along with the students giving us their insights, I’m no longer totally against the use of AI.” Seven of nine faculty committed instructional time to AI literacy work next semester. The pilot trained students to defend their judgment; the PD sequence is producing faculty willing to commit time, classroom space, and individual development to the same work. That sequence – student outcomes anchoring teacher momentum that anchors the principal’s investment – is what this case study now documents in full. These remain single-site, descriptive findings: the Limitations subsection below states what they cannot yet claim, and the Open Questions that follow it define the evidence a second pilot delivery should be designed to produce.

About This Summary

This document reproduces the title page, abstract, and executive summary of AI Discernment in Practice: A Classroom Case Study (FutureSkills Practitioner Research Report No. 1, July 2026). Section and finding cross-references in the text (for example, §2.5, §3.5, §4, §6, Appendix C) point to the full report, which contains the complete week-by-week record, the artifact specifications, the evidence review, recommendations, and appendices. The full report is available on request from [email protected].

Methods at a glance

Single-site descriptive case study produced inside a co-design partnership. The educator (K. A. Keener, English Department, NDSS) delivered all core instruction; FutureSkills observed weekly and co-facilitated Weeks 5 – 6. Evidence sources: pre-pilot student survey (n = 10); post-pilot open-response student survey (May 19, counts reported for captured responses only); weekly observation notes (April 15 – May 18); AI Usage Tracker and Reflection Logs; Policy Builder artifacts and the final seven-statement student policy; peer-workshop responses; the educator’s verbatim reflective accounts; faculty PD records (May 18, June 1); and an anonymous post-PD faculty Google Form (n = 9, administered shortly after the June 1 session; some items optional). No comparison group was constructed, and no causal claims are made.

Limitations at a glance

Small cohort (n = 20; one section of ten observed weekly), one school, one six-week unit, and one exceptionally engaged educator. This report was prepared by FutureSkills, which co-designed the intervention it documents; the full report carries a conflict-of-interest declaration and the mitigations adopted (verbatim educator accounts, exact counts, null and negative results reported). Post-pilot survey capture was incomplete, and durability beyond the unit has not yet been measured.

References Cited in This Summary

Anthropic. (2025). Anthropic Education Report: How university students use Claude. anthropic.com/news/anthropic-education-report-how-university-students-use-claude

Cheng, M., et al. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science. doi:10.1126/science.aec8352

Dakan, R., & Feller, J., in collaboration with Anthropic. (2025). AI Fluency: Framework & Foundations. aifluencyframework.org

NYC Public Schools. (2026). Guidance on artificial intelligence. schools.nyc.gov/about-us/vision-and-mission/guidance-on-artificial-intelligence

Sharma, M., et al. (2023). Towards understanding sycophancy in language models. arXiv:2310.13548.

Udoyi, K., with Keener, K. A. (2026). AI Discernment in Practice: A Classroom Case Study — Executive Summary. FutureSkills Practitioner Research Report No. 1. New York, NY: FutureSkills. futureskills.co

FutureSkills · Human Development for an AI-Shaped World