Deliberation analysis · Session 479D4D

Topic｜Generating and distributing AI content unsuitable for children and young people

The common ground is an adaptable policy mix—not a single ban

Among those who responded, the clearest directions were to involve children and young people, build evidence on concrete harms, and combine safeguards with support. Several apparent trade-offs were supported together.

71 participants with votes24 prompts1,028 responsesAggregate only
Strongest agreement91%

Formal role for children and young people · 65 people answered

Evidence first89%

Research into concrete harms · 47 people answered

85%Regular policy review · 34 answered (fewer responses)
Question needs revision43% chose neutral

Q24 is phrased as a question; the remaining responses should not be read as policy support or opposition · 23 responses (fewer responses)

01 · Common ground

Common ground is about how policy should work—not one preferred ban

Participants most clearly backed formal youth involvement and evidence on concrete harm. These are agreements about governing well, not a mandate for one regulatory instrument.

Five-point distribution for key statements Distribution of strongly agree, agree, neutral, disagree and strongly disagree responses across seven key statements and Q24, with the response count for each prompt. Strongly agree Agree Neutral Disagree Strongly disagree Formally involve children andyoung people in policymaking 35% 55% 8% n = 65 Urgently research specific AIharms to children and young people 40% 49% 9% n = 47 Age restrictions still provideuseful boundaries 23% 62% 9% n = 65 Provide annual parent–child AIsafety education 23% 54% 16% n = 61 Complete child impact assessmentsbefore launch 25% 49% 16% 10% n = 61 Prioritise victim support overrestricting generation 16% 57% 12% 16% n = 69 Prevent expanded surveillance inthe name of child protection 31% 35% 19% 15% n = 54 Can regulation keep pace with AI? 13% 26% 43% 17% n = 23 No ‘strongly disagree’ votes were recorded. Q24 is a question, so agreement cannot be read as support for a policy position.
Formal youth rolen = 65
Research concrete harmsn = 47
Age boundariesn = 65
Parent–child AI safetyn = 61
Pre-launch impact checksn = 61
Victim support firstn = 69
Prevent expanded surveillancen = 54
Can rules keep pace?n = 23

Ordered by agreement. Agree combines strongly agree and agree; response counts differ. Statement 24 is a question.

How to read: Each row is one prompt. Desktop bars show all five response points; mobile bars combine strong and ordinary agreement or disagreement. n is the response count, so denominators differ.Why it matters: Treat participation and evidence as policy-design requirements. Rewrite Q24 before using it to infer views on whether regulation can keep pace with AI.

02 · From principles to action

Three principles turn agreement into a practical governance test

The principles become useful when they are applied consistently across five action areas. This is an analytical framework—not a single vote on a complete package.

Three principles and five action areas for AI safety governance An analytical framework drawn from twenty-four prompts. Three strongly supported process principles connect to five action areas; percentages refer only to the example evidence named in each card. Three strongly supported principles Youth participation 91% n = 65 A formal role in policydesign and review Evidence on harm 89% n = 47 Use concrete evidence tocalibrate intervention Review and adapt 85% n = 34 · fewer responses Review regularlyas evidence changes Education 77% Parent–child safetyeducationn = 61 Build practical AIsafety skillsat home and school Age + context 67% Supported bothstatementsn = 36 Use age as abaseline, thenconsider content,context and risk Impactassessment 74% Pre-launch assessmentn = 61 Check impacts onchildrenbefore productslaunch Victimsupport 72% Victim support firstn = 69 Provide victimsupportand routes toredress Rightssafeguards 67% Prevent expandedsurveillancen = 54 Prevent expandedsurveillanceand protect privacy Each percentage refers to the named example statement; denominators differ. Age + context combines two statements.
Three strongly supported principles
Youth participationn = 6591%

A formal role in policy design and review

Evidence on harmn = 4789%

Use concrete evidence to calibrate intervention

Review and adaptn = 34 · fewer responses85%

Review regularly as evidence changes

  • Education77%

    Build practical AI safety skills at home and school · Parent–child safety education · n = 61

  • Age + context67%

    Use age as a baseline, but not the only criterion · Supported both statements · n = 36

  • Impact assessment74%

    Check impacts on children before products launch · Pre-launch assessment · n = 61

  • Victim support72%

    Provide victim support and routes to redress · Victim support first · n = 69

  • Rights safeguards67%

    Prevent expanded surveillance and protect privacy · Prevent expanded surveillance · n = 54

How to read: The top row shows three strongly supported principles; the lower row organises five action areas. Each percentage refers only to the named example evidence, and response counts differ.Why it matters: Use these three principles as shared checks for education, age safeguards, product assessment, redress and rights protection.

03 · Both-and priorities

Many people supported both sides of an apparent trade-off

The overlap shows where the same respondents supported two measures that are often presented as competing choices.

Venn diagrams of support for three paired policy priorities Three area-proportional Venn diagrams show how many joint respondents supported option A, option B, both options, or neither option. Prevention + redress 60 people answered both statements A · Impactassessment B · Victim support supported both 45% 27 people A only 28.3% B only 26.7% Everyone supported at least one Age + context 36 people answered both statements A · Age boundaries B · Contextmatters supported both 66.7% 24 people A only 25% B only 8.3% Everyone supported at least one Rights + creative freedom 26 people answered both statements A · Limitsurveillance B · Creativefreedom supported both 57.7% 15 people A only 7.7% B only 15.4% Outside both 19.2% · 5 people Within each panel, circle and overlap areas are proportional to people. Panel sizes should not be compared. Supportcombines agree and strongly agree.
Prevention and redress
A · Impact assessmentB · Victim support and redress
BOTH45%
A only 28.3%B only 26.7%

Everyone supported at least one option

Answered both: 60
Age boundaries and context
A · Age boundariesB · Context matters
BOTH66.7%
A only 25%B only 8.3%

Everyone supported at least one option

Answered both: 36
Limit surveillance and protect creative freedom
A · Limit surveillanceB · Creative freedom
BOTH57.7%
A only 7.7%B only 15.4%

Outside both: 19.2% · 5 people

Answered both: 26

Circle areas are proportional within each pair. Only people who answered both statements are counted.

How to read: Each Venn diagram includes only people who answered both statements. Circle A shows support for the first option, circle B the second, and the overlap shows support for both.Why it matters: Among joint respondents, 67% backed both age boundaries and contextual judgment; 45% backed both prevention and redress. The rights pairing is preliminary because only 26 people answered both statements.

04 · Where evidence is thinner

Later answers are thinner evidence—and Q24 needs rewriting

Response counts fell from 69 to 23. Before drawing substantive conclusions from later prompts, the response coverage and the wording of Q24 need attention.

Statement order and response counts Response counts declined substantially across 24 statements, from 69 responses to the first statement to 23 responses to the final statement. The first, middle and final groups of eight statements averaged 60.5, 38.8 and 29.3 responses. First 8 Average 60.5 responses Middle 8 Average 38.8 responses Final 8 Average 29.3 responses 20 30 40 50 60 70 1 4 8 12 16 20 24 Statement order Number of responses Statement 1: 69 responses Statement 2: 65 responses Statement 3: 65 responses Statement 4: 61 responses Statement 5: 61 responses Statement 6: 62 responses Statement 7: 54 responses Statement 8: 47 responses Statement 9: 44 responses Statement 10: 42 responses Statement 11: 42 responses Statement 12: 39 responses Statement 13: 37 responses Statement 14: 37 responses Statement 15: 34 responses Statement 16: 35 responses Statement 17: 33 responses Statement 18: 35 responses Statement 19: 32 responses Statement 20: 30 responses Statement 21: 29 responses Statement 22: 26 responses Statement 23: 26 responses Statement 24: 23 responses n = 69 n = 23 Later statements should be revisited; fewer responses do not mean participants cared less.
average responses
First 860.5
Middle 838.8
Final 829.3
How to read: The horizontal axis is statement order and the vertical axis is response count; the background separates the first, middle and final eight statements.Why it matters: Use the later prompts as an agenda for the next round, not as evidence of weaker support. Re-ask them with clearer wording and fuller participation.

Higher-neutral statements are priorities for follow-up

Neutral is not disagreement. It marks where wording, choices or context should be revisited; the reason was not measured.

Prompts with the highest neutral response shares A ranked lollipop chart of the five prompts with the highest neutral shares. Percentages use each prompt's respondents, and prompts with fewer than half of voters responding are marked as having fewer responses. Neutral among statement respondents 0% 25% 50% #24 Can regulation keep pace with AI?† n = 23 FEWER RESPONSES 43.5% #13 Majority-supported positions may contradict oneanother n = 37 29.7% #21 Establish an AI content rating system n = 29 FEWER RESPONSES 24.1% #16 Prioritise family education, not governmentintervention alone n = 35 FEWER RESPONSES 22.9% #19 Involve more government and industry representatives n = 32 FEWER RESPONSES 21.9% Neutral means the respondent selected the neutral option; it does not include missing responses. Percentages use each prompt’srespondents. Prompts with fewer than 36 responses are labelled ‘fewer responses’. † Q24 was phrased as a question, making its response meaning less clear.
#24 Can regulation keep pace with AI?†43.5%
n = 23FEWER RESPONSES
#13 Majority-supported positions may contradict one another29.7%
n = 37
#21 Establish an AI content rating system24.1%
n = 29FEWER RESPONSES
#16 Prioritise family education, not government intervention alone22.9%
n = 35FEWER RESPONSES
#19 Involve more government and industry representatives21.9%
n = 32FEWER RESPONSES

Reasons for neutral responses were not measured. † Statement 24 was phrased as a question.

How to read: A longer line means a higher neutral share among that statement’s respondents. “Fewer responses” means fewer than 36 people answered.Why it matters: This is not a disagreement ranking. It identifies statements that need clearer wording, response choices or background information.

05 · Opinion landscape

Three overlapping tendencies—not opposing camps

The clearest differences concern how to distribute attention: before and after harm, and across product rules, education, context and civil liberties. A participant can align with more than one tendency.

Two opinion dimensions and three overlapping tendencies An exploratory map with two response dimensions and three overlapping soft profiles: care and participation, upstream safeguards, and adaptive governance. People are not assigned to fixed camps. More product rules & checks More education, context & liberty More support & redress after harm More prevention before harm Care & participation Victim support · youth participation Upstream safeguards Impact checks · technical standards Adaptive governance Privacy & education · context-sensitive rules What each tendency emphasises Care & participation Victim support and redress Youth participation and harm research Upstream safeguards Pre-launch child-impact assessment Clear technical standards and ratings Adaptive governance Privacy, education and public dialogue Context-sensitive rules with creative space Shared across all three Formal youth participation and research into concrete harms
Care & participationUpstream safeguardsAdaptive governanceRules & checksEducation, context & libertySupport after harmPrevention before harm
Care & participation

Emphasises victim support, redress and formal youth participation.

Upstream safeguards

Emphasises impact assessment, technical standards and content ratings.

Adaptive governance

Emphasises privacy, education, dialogue and context-sensitive rules.

Shared coreAll three strongly support formal youth participation and research into concrete harms.

How to read: The two axes summarise the main response gradients. Shaded areas are overlapping soft profiles; they are a reading aid and do not assign anyone to a fixed group.Why it matters: All three tendencies share formal youth participation and research into concrete harms. Their differences concern policy emphasis, while several priorities remain compatible.

06 · What follows

The data points to a sequence of next moves

These are practical implications from this session—not a claim to represent everyone in Taiwan.

  1. 01

    Make youth participation structural

    91% agreed · n = 65

    Give children and young people a formal role in policy design, implementation and later review—not a one-off consultation.

  2. 02

    Build evidence and review into the policy cycle

    89% backed harm research; 85% backed regular review · n = 47 and 34

    Define the harms being addressed, connect intervention strength to evidence, and schedule regular reassessment as technology changes.

  3. 03

    Design policy packages, not forced choices

    67% backed both age boundaries and context; 45% backed both prevention and redress among joint respondents

    Combine age with context and risk, and pair upstream safeguards with victim support and routes to redress.

  4. 04

    Improve the evidence before deciding the edge cases

    Responses fell from 69 to 23; Q24 was 43% neutral

    Rewrite Q24, revisit the later prompts, and broaden participation before treating these results as a final policy mandate.

07 · Evidence

Explore the detail when needed

Open the sections below for the step-by-step methodology, complete prompt results and chart downloads.

How this report was analysed—step by step

Open this section when you want to see how the raw responses became each chart and conclusion.

  1. 01

    Prepare and validate the data

    We checked that each participant had at most one response to each prompt and that every response used a valid scale option. The final dataset contains 71 participants with votes, 24 prompts and 1,028 responses. Names and identifiers are excluded from the report.

    How to interpret itThis establishes what was analysed and prevents duplicate votes from changing the results.

  2. 02

    Calculate each prompt’s result

    Agreement combines ‘strongly agree’ and ‘agree’; disagreement combines the two disagreement options. Percentages use only the people who answered that prompt. An unanswered prompt remains missing and is never counted as neutral. Coverage is the number who answered divided by all 71 participants with votes.

    How to interpret itDifferent prompts have different denominators, so later low-response prompts carry less evidence.

  3. 03

    Organise statements into themes and principles

    We compared the wording and response patterns, then grouped related statements into principles and action areas. These labels are an analyst-created synthesis; they are not an additional vote or an automated topic model. Every percentage shown still belongs to its original prompt.

    How to interpret itThe themes help readers connect individual statements without inventing a new score.

  4. 04

    Measure support for paired priorities

    For each Venn diagram, we included only people who answered both prompts. A person enters the overlap when they agreed with both. Neutral and disagreement count as ‘not supported’ for this comparison. Circle areas are proportional to the four observed combinations: both, A only, B only and neither.

    How to interpret itThe overlap shows whether the same respondents combined two priorities; it does not show cause or intensity.

  5. 05

    Find the main opinion dimensions

    For this step, responses are encoded as agree (+1), neutral (0) and disagree (−1), while missing remains separate. PCA uses all 71 voters to find the two strongest response gradients. Missing cells use each prompt’s observed average only while fitting the axes; personal positions use answered prompts and are adjusted for response count. The two axes explain 28.1% of the observed variation.

    How to interpret itAxis names are plain-language interpretations of the statements with the strongest PCA contributions; the map is a summary of part of the variation.

  6. 06

    Describe overlapping opinion tendencies

    54 people who answered at least 7 prompts entered a three-profile fuzzy c-means analysis; 17 people did not meet that threshold. Each person receives weights across all three profiles rather than one fixed assignment. Profile positions are weighted centres, and the shaded regions show weighted spread. The regions are enlarged slightly for readability and are not confidence intervals.

    How to interpret itThree profiles were chosen to keep the public map readable. Their names describe relative emphasis in this session and should not be treated as permanent identities.

  7. 07

    Check uncertainty and limit the claims

    Separately, we tested fixed K-means partitions with two to five groups and resampled the prompts 100 times. Median repeatability was 0.22; 0 runs reached the report’s 0.70 publication guardrail. We therefore publish overlapping tendencies, not participant camps. The analysis is Pol.is-inspired offline work, not official Pol.is output.

    How to interpret itNo individual coordinates, profile weights or group assignments are published. Results describe this session and are not representative of everyone in Taiwan. Response counts fall from 69 to 23, no one selected ‘strongly disagree’, and Q24 is a question that does not fit the agreement scale, so those results require extra caution.

View all 24 prompts
No.Short labelnCoverageAgreeNeutralDisagree
1Prioritise victim support over restricting generation6997.2%50 (72.5%)8 (11.6%)11 (15.9%)
2Age restrictions still provide useful boundaries6591.5%55 (84.6%)4 (6.2%)6 (9.2%)
3Formally involve children and young people in policymaking6591.5%59 (90.8%)5 (7.7%)1 (1.5%)
4Provide annual parent–child AI safety education6185.9%47 (77%)10 (16.4%)4 (6.6%)
5Complete child impact assessments before launch6185.9%45 (73.8%)10 (16.4%)6 (9.8%)
6Treat cases with and without identifiable victims differently6287.3%44 (71%)12 (19.4%)6 (9.7%)
7Prevent expanded surveillance in the name of child protection5476.1%36 (66.7%)10 (18.5%)8 (14.8%)
8Urgently research specific AI harms to children and young people4766.2%42 (89.4%)4 (8.5%)1 (2.1%)
9Build public understanding before taking legislative action4462%30 (68.2%)9 (20.5%)5 (11.4%)
10No single standard fits people from different backgrounds4259.2%31 (73.8%)7 (16.7%)4 (9.5%)
11Prioritise education over restrictions4259.2%33 (78.6%)5 (11.9%)4 (9.5%)
12Redress matters more than prevention and enforcement3954.9%31 (79.5%)6 (15.4%)2 (5.1%)
13Majority-supported positions may contradict one another3752.1%22 (59.5%)11 (29.7%)4 (10.8%)
14Age should not be the only criterion3752.1%28 (75.7%)7 (18.9%)2 (5.4%)
15Revise policy as understanding develops3447.9%29 (85.3%)4 (11.8%)1 (2.9%)
16Prioritise family education, not government intervention alone3549.3%26 (74.3%)8 (22.9%)1 (2.9%)
17Establish clear technical standards3346.5%26 (78.8%)4 (12.1%)3 (9.1%)
18Emotional learning and interpersonal collaboration matter equally3549.3%33 (94.3%)1 (2.9%)1 (2.9%)
19Involve more government and industry representatives3245.1%25 (78.1%)7 (21.9%)0 (0%)
20Build shared understanding through education rather than compulsory rules3042.3%27 (90%)1 (3.3%)2 (6.7%)
21Establish an AI content rating system2940.8%20 (69%)7 (24.1%)2 (6.9%)
22Avoid letting political correctness constrain creative freedom2636.6%19 (73.1%)4 (15.4%)3 (11.5%)
23Verify AI-generated content, with extra care for younger children2636.6%18 (69.2%)5 (19.2%)3 (11.5%)
24Can regulation keep pace with AI?2332.4%9 (39.1%)10 (43.5%)4 (17.4%)
Download charts
Key response patternsSVG vector graphicNeutral follow-upSVG vector graphicPrinciples and action areasSVG vector graphicPaired prioritiesSVG vector graphicResponse coverageSVG vector graphicOpinion-tendency mapSVG vector graphic