More ways to learn
AI tools can explain concepts in different ways, support language practice, and help people brainstorm or organize ideas. Accuracy and accessibility still need to be checked.
A critical-thinking project · Live educational website
Question the tool.
Examine the evidence.
AI can help solve real problems. It can also create serious ethical questions. Explore the benefits, investigate the costs, and decide what you think for yourself.
This is an educational project, not a definitive verdict on every AI system. Claims should be checked against the linked sources.
The question behind the question
Who benefits — and who bears the cost?
Technology is never just technology.01 / 05
01 / Potential benefits
What might AI make possible?
Potential benefits are worth examining. They are not guarantees, and they do not automatically settle questions about how a system was built or deployed.
Research support
Some AI methods help researchers analyze complex patterns or prioritize possibilities. Results require expert validation; a fluent answer is not proof of truth.
Reducing barriers
Speech, transcription, translation, and other assistive tools may help some people participate more fully. Design quality, cost, privacy, and user needs matter.
Critical question: Who can access these benefits, who controls the tools, and what evidence shows that a particular application actually helps?
02 / Ethical concerns
Progress for whom?
Ethical questions often concern the entire system around a tool: the data used, the people affected, the business model, and the consequences of deployment.
These are categories to investigate, not a claim that every AI product has the same impacts.
Who agreed?
Were personal data or creative works collected and used in ways people understood and consented to? Legal rules and ethical expectations do not always match.
Who gets paid?
Automation can change jobs and bargaining power. Ask who gains financially, whose work is displaced, and whether affected communities have a voice.
Who is misrepresented?
Systems can make uneven errors or reproduce patterns in their data. Testing must consider different groups and the real consequences of mistakes.
What is exposed?
Prompts and uploaded information may be sensitive. Understand a service's data practices before sharing personal, confidential, or identifying details.
What does it consume?
AI infrastructure uses electricity, hardware, and water across its lifecycle. Impacts vary by model, location, energy source, and how often it is used.
Who is responsible?
AI can generate plausible falsehoods and make mistakes at scale. Responsibility should not disappear behind claims that “the algorithm did it.”
03 / Interactive ethics lab
There isn't always a clean answer.
Choose a response to a fictional situation. This activity doesn't score your morality; it makes the trade-offs visible.
Try this: choose the option closest to what you would do, then consider what your choice leaves unresolved.
SCENARIO 01 · CREATIVE WORK
Your group needs artwork for a student project. An AI image tool can make something quickly, but you don't know exactly what training data was used.
What would you do?
Consider the trade-offs
Question to take with you: What evidence or principle would change your decision?
04 / Community perspective
What your peers think
Ethical disagreements can be useful when people explain their reasoning and listen to the evidence.
Loading approved opinions…
Remove your opinion
Have a privacy code from an earlier submission? Enter it here to permanently delete the matching opinion, whether pending or approved. This cannot be undone.
Only approved visitor submissions appear here. Fictional demonstration responses have been removed so they cannot be mistaken for real submissions. Because participation is voluntary and self-selected, these opinions are not a representative survey.
05 / Keep investigating
Don't take this website's word for it.
Start with institutional resources and primary documents. Read the source in context, check its date and scope, and look for credible disagreement.
A framework for identifying and managing risks in AI systems. Useful for thinking about governance, measurement, and accountability.
An international policy framework covering human rights, transparency, fairness, and environmental and social impacts.
Reports and policy materials on copyright and AI. Copyright questions are complex and can depend on the specific use and jurisdiction.
Analysis of AI-related energy demand and the broader relationship between data centres, electricity, and technology.
Principles for trustworthy AI, including human-centred values, transparency, robustness, and accountability.
Practical guidance and questions for organizations working to govern, map, measure, and manage AI risks.
These links are starting points, not proof that every claim on this page is settled. Before presenting the website for a grade, open each source, confirm it remains available, and add specific citations beside any factual claims you research further.
06 / Transparency
Put our own process under the microscope.
This website was built with AI assistance. We share what was done, what went wrong, and what still needs work—without asking visitors to read an entire technical transcript.
Who did what?
ChatGPT assisted with layout, writing, code, and suggested database checks. The human creator selected the project's ethical direction, tested the website, reviewed screenshots, and chose its privacy and publication policies. AI-generated suggestions were not accepted as automatically correct.
What did we discover?
An AI-suggested SQL query referred to a nonexistent column and needed correction. A privilege audit exposed an unnecessary TRUNCATE permission, which the creator revoked and verified. These mistakes illustrate why testing and human oversight matter.
How do visitor opinions work?
Submissions enter a pending moderation queue. The creator added a long-term publication notice and required acknowledgment checkbox. Browser testing showed that an unchecked form was blocked; a checked test response was stored as pending with acknowledgment true. A simulated anonymous database insert with acknowledgment false was rejected.
These are specific tests, not a comprehensive security certification. Controlled backend tests confirmed that a submission returned a one-time privacy code and receipt, and that the matching test opinion could be removed using that code. Live website tests subsequently confirmed submission with a privacy-code receipt, hiding of pending opinions, publication after manual approval, and permanent deletion using the removal code. Turnstile is active, and a database-backed five-attempts-per-hour fixed-window limit was tested for counting and rejection. A real HTTP 429 response and the trustworthiness of the IP header remain unverified; the review is not an independent security audit. The recorded tests used developer-created opinions.
06.1 / Development timeline
From prototype to ongoing project.
Key milestones and decisions, presented as an edited summary rather than a verbatim conversation.
Explore the ethics
Built the benefits, concerns, interactive dilemma, source links, and a demonstration opinions section.
Connect a database
Moved from browser-only demonstration responses to Supabase with pending/approved moderation.
Test, find, correct
Verified row-level security and role permissions, corrected a failed AI query, and removed an excessive TRUNCATE privilege.
Choose long-term disclosure
The creator decided to continue the site beyond the October 21, 2026 school presentation and removed a proposed October 28 submission deadline.
Verify the checkbox
Added a required acknowledgment, confirmed a successful pending submission, and verified that the database rejected a false acknowledgment.
Build submission receipts and removal
After several AI-generated revisions, the creator deployed and tested server-side opinion submission, private code hashing, and removal by code. The live receipt and deletion workflows subsequently passed end-to-end tests. Additional protections and review are documented in the later milestones.
Connect receipts and removal forms
The website was updated to call the two Edge Functions rather than allowing browser-side database inserts. A copyable, downloadable receipt and removal form were added. The creator tested the live website: a submitted opinion stayed pending, appeared publicly after manual approval, and disappeared from both the website and database after removal by privacy code.
Human judgment: the creator insisted on complete, paste-ready code rather than error-prone manual edits. An earlier AI draft incorrectly attempted to call a private database function through a public API; it was corrected to use a server-side database connection. These observed tests are not a comprehensive security certification.
Test the full moderation cycle
A developer-created opinion was submitted with Turnstile, stored as pending, manually approved, and shown on the public website. Its privacy code then removed it from both the website and the Supabase opinions table.
Count and reject excess attempts
Added a keyed visitor identifier and a database-backed limit of five verified submission attempts in a fixed one-hour window. A live submission incremented the counter; a rolled-back database test returned false on the sixth check. A live HTTP 429 response and IP-header trust remain unverified.
Inspect access and disclosures
Checked that row-level security was enabled, anonymous roles lacked direct read/write grants, and policies restricted approved-opinion reading. Searches of the published HTML found no matches for several private credential names or secret-key prefixes; a public Supabase publishable key was present as expected. A subsequent permissions review confirmed that neither anonymous nor authenticated visitors can execute the administrative rls_auto_enable function after its general EXECUTE permission was revoked. These checks are limited and do not prove the absence of all vulnerabilities.
About this record: this public page summarizes observed tests and decisions. The creator maintains more detailed technical notes privately. The summary is not a complete verbatim transcript and does not guarantee the website is secure.
07 / Human accountability
Meet the human developer.
AI assisted with drafting and coding, but people made the decisions about this project's purpose, content, safeguards, and publication.
Asterius D.
I created AI Under the Microscope to examine both the possibilities of AI and the ethical concerns surrounding how it is developed and used. I directed the project's approach, evaluated AI suggestions, tested its features, and chose how visitor opinions should be handled.
Who is accountable?
AI-generated output was treated as material to question, revise, and test—not as an authority. The creator remains responsible for the published content and moderation decisions.
Additional collaborators will be credited here only after they confirm their participation and preferred public credit.