Ethics and Power

Deepfakes: seeing is no longer believing

You know from Lesson 3 that a diffusion model can conjure any image from noise, and voice models can clone a voice from seconds of audio. Combined, that is a deepfake: fabricated video or audio of a real person. Some uses are harmless fun. Others are scams ("your grandson needs bail money", in his exact voice), fake politician clips before elections, and fake celebrity endorsements for sketchy investments, all of which have already hit the Caribbean.

Your defence is a habit, not an app:

• Check the source, not the clip. Who posted it first? Is any real newsroom carrying it?
• Watch the incentives. Urgency + money + secrecy ("act now, tell no one") is the scam signature, deepfaked or not.
• Verify out-of-band. A "family emergency" call gets a callback on the number you already have.
• Artifacts help but do not rely on them. Weird hands, mismatched earrings, and glitchy audio betray lazy fakes. Good fakes have none. The source-checking habit is what scales.

Bias: the data decides who is seen

Generative models learn from whatever data was collected, and the internet over-represents some faces, accents, and stories while barely sampling others. The results are concrete: face tools that work worse on dark skin, speech systems that stumble on Patois and Creole, image models that render "a professional" as someone who does not look like your street. This is not the machine being evil; it is arithmetic faithfully amplifying an unbalanced dataset, the same way one biased column skewed your Week 5 models.

The fix requires people who OWN the missing data: Caribbean voices, Caribbean images, Caribbean records, collected and governed by Caribbean institutions. That is not charity work; it is infrastructure, and it is a career one of you reading this will have.

Where the bias physically lives: it is just counting

Bias is not a mood inside the machine. It is a number you could point to: how many times each kind of example appears in the training data. Remember Week 5: gradient descent nudges the weights a tiny bit for every training example, so whatever appears most often gets the most nudges, and the model gets very good at exactly that. The internet the big models trained on skews heavily toward English text from the US and Europe, because that is who has been posting the most, the longest, on the platforms that got scraped. Patois lives mostly in voice notes, dancehall lyrics, and conversations that never became web text at all. If it was never written down and scraped, then as far as the model is concerned, it never happened.

Why AI confidently makes things up

Ask a chatbot for the year a small Jamaican community group was founded, and it may answer "1987" in a calm, certain tone, when no such fact exists anywhere. Why? Go back to Lesson 1: the model is choosing the most probable next token, not looking facts up in a database. "The organisation was founded in" is almost always followed by a plausible year in its training data, so a plausible year is exactly what gets generated. The sentence is fluent because fluency is what the model was trained to produce. Nothing in next-token prediction ever asks "wait, is this TRUE?". Fluency and truth are two different things, and the model only optimises the first one.

Who actually holds the power

Strip away the hype and generative AI runs on three ingredients. Whoever controls them controls what these systems can see and say.

Right now, all three are concentrated in a handful of companies, mostly in the US and China. That is why "just use the app" is not the whole story: the app's defaults, blind spots, and prices were set far from here. But the three ingredients are not equally locked up. Compute is the hardest for a small region to match. Data is the opposite: nobody on Earth can collect Caribbean voices, stories, and records better than the people who live them. Open-weight models you can fine-tune are the bridge in the middle. That is the realistic play for the region, and for you: own the data layer, master the models, and rent the compute until we build our own.

The bill nobody shows you

Before you trust it or share it: the 60-second check

You now know two mechanisms that should change your habits: models hallucinate where data is thin, and deepfakes are cheap. So run this routine on any AI output (or viral clip) before it leaves your hands:

• 1. Flag the checkable claims. Names, dates, numbers, quotes, "studies show". These are the tokens most likely to be fluent guesses, especially about Caribbean topics.
• 2. Verify the ones that matter in an independent source. Not by asking the same AI "are you sure?" (that just re-rolls the dice). Search for a real source: a newsroom, an official site, a textbook.
• 3. Ask: is this a thin-data zone? A claim about your parish, a local person, or a small organisation deserves double suspicion. The model has seen the least about exactly the things closest to you.
• 4. Check what happens if you are wrong. A wrong movie recommendation costs nothing. A wrong "fact" in your SBA, a fake quote about a classmate, or a money decision costs plenty. Match your checking effort to the stakes.
• 5. Only then share, and say AI was involved. Once you forward it, your name is the source, not the model's.

The builder's code

This course's answer to all of the above is not "avoid AI". It is: be one of the people who builds and steers it. A short code to carry into the AI Solution Challenge and beyond:

• Disclose. If AI made it or helped, say so.
• Verify. Never ship an AI claim you have not checked, especially about people.
• Consent. No generating real people's faces or voices without their yes.
• Represent. When you build, feed the model YOUR context. Make the Caribbean visible in the data.
• Question. Ask who benefits, who pays, and who is missing from every AI system you meet.