Ethics, Bias & Safety

The Lesson Where We Get Serious

This course has had fun with corgi astronauts and spam haiku. This lesson sets the jokes down for a bit, because the same technology that drafts your vacation reply also raises genuinely hard questions, and understanding them is part of being AI-literate. Not to scare you off. To make you the person in the room who understands what's actually at stake.

Bias: The Mirror Problem

Remember where models learn everything: from enormous amounts of human-written text and human-made images. That data carries our history, including our stereotypes, blind spots, and inequalities. A model trained on it absorbs those patterns just as faithfully as it absorbs grammar.

The results show up in concrete ways. Image models asked for "a CEO" have historically over-produced images of white men in suits. Language models can associate certain names or dialects with negative traits. A hiring tool trained on a company's past decisions can quietly learn to replicate that company's past discrimination, with a statistical accent that makes it look objective.

Two things make AI bias more troubling than the human kind it reflects:

• Scale. One biased hiring manager affects one team. One biased model deployed widely affects millions of decisions, identically and simultaneously.
• The neutrality costume. "The algorithm decided" sounds impartial. It isn't. It's the patterns of the training data, laundered through math. Treating model output as neutral is itself a bias.

Developers work to counteract this (curating data, testing outputs across groups, adjusting behavior), and it helps. But no one has fully solved it, which is why bias awareness belongs in the user's toolkit, not just the builder's.

Deepfakes and Misinformation: The Trust Problem

Lesson 6 celebrated generative images, audio, and video. Here is the other edge of that sword: the same tools can fabricate a politician's speech, a celebrity's endorsement, or an ordinary person's voice begging a relative for emergency money. Voice-cloning scams are not hypothetical; they are happening now, and they work precisely because the voice sounds right.

The deeper damage goes beyond individual fakes. As fabricated media gets easier, some people begin dismissing real evidence as fake; researchers call this the liar's dividend. When anything can be fabricated, the truth itself gets easier to deny. That corrodes something societies badly need: a shared baseline of facts.

Jobs: The Disruption Problem

Honest answer: generative AI is already reshaping work, and pretending otherwise helps no one. Roles heavy in routine drafting (boilerplate text, stock imagery, basic code, first-pass summaries) are being automated fastest. Lesson 9's pattern cuts both ways: if a job is mostly "produce a plausible draft," the draft machine is real competition.

History says technology shifts tend to transform work more than eliminate it, and new roles are already appearing. But "it worked out eventually" was cold comfort to workers displaced in every previous transition, and it's honest to say the same here. The most durable position, as Lesson 9 put it, is the accountability side of your craft: judgment, review, responsibility, and the human relationships around the work. People who understand these tools (which now includes you) tend to fare better than those who ignore them.

Privacy: The Data Problem

Two distinct concerns, both worth knowing:

Training data. Models learned from vast scraped datasets that include people's writing, art, and photos, largely without asking. Artists and authors have sued over it, and courts and lawmakers are still drawing the lines. However it resolves, "where did the training data come from?" is a fair and important question.

Your prompts. Whatever you type into an AI service is sent to that service. Policies on storage and reuse vary widely. The safe habit is simple: do not paste secrets (medical details, client data, credentials, your company's unreleased plans) into tools you haven't checked. If you wouldn't email it to a stranger, don't prompt it to one either. Workplaces increasingly have AI-use policies for exactly this reason; check yours.

Responsible-Use Habits

You don't control how models are trained. You fully control how you use them. The habits, gathered in one place:

VERIFY   Check factual claims before relying on or sharing them
(hallucinations, Lesson 7 - still undefeated)
QUESTION Treat outputs touching people - hiring, lending,
grading - as drafts needing human review, never verdicts
PROTECT  Keep sensitive personal and work data out of prompts
unless you know the tool's policy allows it
DISCLOSE Be honest when AI wrote something, especially where
trust matters - school, journalism, work products
REFUSE   Don't create deceptive media of real people or spread
unverified content, even as "just a joke"

Reflection Quiz