Module 5: AI and Society¶
Module 5 of 10
What You Will Be Able to Do¶
- Give two examples of bias in artificial intelligence (AI) outside the clinic, and say what to check when an output affects a person.
- Describe what AI costs in energy and water, with sourced figures, and why the estimates disagree.
- Recognize AI-made health misinformation and synthetic media, and respond well when someone brings it to you.
The Core Idea¶
The chatbots in the modules so far are built from people's writing and pictures, trained with the help of people paid to judge their answers, and run on electricity and water in data centers, and they now shape what many people read about their health.
You do not need to settle the debates around them to use them well. Knowing what is at stake helps you judge an output that affects a person, weigh the costs of a habit, and answer a friend or a patient who asks about something they saw.
Bias Beyond the Clinic¶
Module 9 covers bias in patient care. It also shows up in the work of a student or a teacher:
- Dialect. Shown the same statements written in African American English or in Standardized American English, language models matched the writers of African American English to less prestigious jobs, convicted them more often in invented trials, and sentenced them to death more often, while what the models said openly about African Americans was more positive. Training with human feedback softened what the models said openly, but not the hidden prejudice (Hofmann et al., 2024).
- Detectors. Seven detectors of AI writing, tested in 2023, labeled on average 61% of 91 essays by non-native English writers as AI-generated, while classifying essays by United States eighth-graders almost perfectly (Liang et al., 2023).
- Images. Asked for "a software developer", an image generator showed a white person in 99% of its pictures, against 56% of software developers in the United States, and it drew every nurse, housekeeper, and flight attendant as a woman; adding words meant to counter a stereotype did not undo it (Bianchi et al., 2023).
When an output affects a person, such as a grade, a reference, or a summary of someone's work, ask what it would say about someone else with the same facts, check what it says at the source, and keep the decision with a person. The AI Responsible Use Policy asks for that attention to bias in assessments and decisions (Responsible Use, section D), and a detector's score is a reason to look again, not evidence on its own.
The Environmental Cost¶
Every answer runs on computers in data centers that draw electricity and often use water for cooling. Estimates of how much differ widely, and the reasons are worth knowing. Some count only the electricity one answer uses; others add the machines kept idle and ready, the building's cooling, the water used to generate the electricity, or the training of the model. They measure different models, in different places and years. And some figures come from companies measuring their own products, which no one outside has checked.
The figures, who measured them, and what each leaves out
About 485 terawatt-hours (TWh)
of electricity used by data centers worldwide in 2025, about 1.5% of the world's total, projected to reach about 950 TWh, around 3%, by 2030
Source: International Energy Agency (IEA), Key Questions on Energy and AI, April 2026. Leaves out: how much of it is AI, which cannot be measured directly, and the 2030 figure is a projection with wide uncertainty.
0.24 watt-hours (Wh) and 0.26 milliliters (mL) of water
for the median text prompt in Google's Gemini app in May 2025, about five drops of water, by the company's own measurement
Source: Google, Measuring the Environmental Impact of Delivering AI at Google Scale, August 2025. Leaves out: training the model and storing data; the water used to generate the electricity; checking by anyone outside Google.
A 500 mL bottle for roughly 10 to 50 answers
of medium length from GPT-3, an older model, counting the water evaporated to cool the servers and the water used to generate their electricity; training it took an estimated 5.4 million liters
Source: Li, Yang, Islam and Ren, Communications of the ACM, June 2025. Leaves out: newer models, which may use more or less; it is an estimate from published averages, not a measurement.
Reading these figures well is mostly a matter of asking who measured, what was counted, and what was left out.
The People Behind the Models¶
Chatbots learn to follow instructions partly from people. In a 2022 study of this training at OpenAI, about 40 contractors, hired through Upwork and Scale AI, wrote example answers and ranked the model's replies (Ouyang et al., 2022).
Some of this work has been done far from the companies, for little pay. In 2023, TIME reported that workers in Kenya labeling violent and abusive text for OpenAI, through the outsourcing firm Sama, took home about $1.32 to $2 an hour, while OpenAI paid Sama $12.50 an hour for the work; Sama said its workers could earn $1.46 to $3.74 an hour after taxes (Perrigo, 2023).
Copyright and Creators¶
Models are trained on large collections of text and images, including work by people who were never asked. Whether that training is lawful is not settled. In the United States, the Copyright Office concluded in 2025 that material made by AI can be protected by copyright only where a person made enough of the creative choices in it, and that a prompt alone is not enough; its report on training found that some uses of copyrighted works to train models may be fair use and others not, depending on the use (Copyright and Artificial Intelligence). Courts are deciding the cases one at a time, and their answers so far differ.
Where the questions stand
148 lawsuits
over copyright filed against AI companies in US courts since 2023, counting those since settled or dismissed
Source: ChatGPT Is Eating the World, a tracker kept by Edward Lee of Santa Clara University School of Law, October 2026. Leaves out: what each suit claims, since not all are about training, and how each ended; it counts filings, not rulings.
A $1.5 billion settlement
for authors whose pirated books were copied into a library used to train AI, approved in July 2026, after the judge ruled in June 2025 that the training itself was fair use
Source: Bartz v. Anthropic, US District Court for the Northern District of California, July 2026. Leaves out: claims about what a model produces, which the settlement does not cover; a district court's ruling binds no other court.
Fair use, for these authors
a judge ruled in June 2025 that Meta's training was fair use against 13 authors who had offered no evidence of harm to the market for their books, and wrote that the ruling does not mean the training is lawful
Source: Kadrey v. Meta, US District Court for the Northern District of California, June 2025. Leaves out: other authors' claims, which the ruling does not decide; the case continues on other claims.
For your own work, the policy's section on intellectual property covers what you put into a tool and what you take out of it. A classmate's draft or a colleague's slides are their work, so check they are happy for it to go into a tool, as you would want for yours.
Misinformation and Synthetic Media¶
Chatbots can be made into convincing sources of false health advice. Researchers set up five leading models, through instructions a user never sees, to answer health questions falsely in a formal, scientific tone: 88 of 100 answers were disinformation, and four of the five models gave it in every answer (Modi et al., 2025). Deepfake videos of well-known television doctors have been used on social media to sell remedies, including products claiming to fix high blood pressure and diabetes (Stokel-Walker, 2024). And many people now ask chatbots about their health.
What surveys of the public found
32% of US adults
used AI tools for information or advice about their physical or mental health in the past year
Source: KFF Tracking Poll on Health Information and Trust, March 2026. Leaves out: what people asked and whether the answers were right; it is what people report, in the United States only.
44% of US adults
are not confident they can tell true from false health information from AI tools or chatbots, while 56% are
Source: KFF Tracking Poll on Health Information and Trust, June 2026. Leaves out: whether that confidence matches anyone's skill; it is what people report, in the United States only.
Telling a real video from a made one by eye keeps getting harder. A standard called Content Credentials lets cameras and editing tools attach a record of how an image or video was made, and of any use of AI in making it (C2PA); one that is present can help. But adding one is optional, a credential can be removed, and many platforms strip them, so a missing credential proves nothing. What works is checking the source and the claim.
When someone brings you a claim they saw, start by asking where they saw it and what they hoped it would do. Then find the source, such as the person's own channel or site, the journal, or the regulator; check the claim where it would be published if it were true; and talk it through with them rather than at them.
When someone brings you a health claim they saw
Ask where they saw it
and what they hoped it would do
Find the source
the person's own channel or site, the journal, the regulator
Check the claim
where it would be published if it were true
Talk it through
with them, not at them
Self-Check¶
A detector marks a classmate's essay, written in their second language, as likely AI. What does that score tell you?
Very little on its own. In a 2023 test, seven detectors wrongly labeled, on average, 61% of essays by non-native English writers as AI-written, so the score may say more about the writing than about how it was produced. A judgment rests on the work itself and on hearing from the writer.
A friend says every chatbot question uses a bottle of water. What do the figures say, and why do estimates vary so much?
The bottle comes from an estimate for GPT-3, an older model, that spread one bottle over roughly 10 to 50 answers, counting the water used to generate the electricity as well as the water used for cooling. Google's own measurement for its Gemini app came to about five drops for a typical question, counting only the water used on site. Both can be right about what they measured; they differ in what was counted, which model, where, and who measured it.
A patient shows you a video of a well-known doctor selling a supplement that "reverses diabetes". What do you check, and how do you open the conversation?
Check whether the doctor has endorsed it on their own channels, since deepfaked endorsements of real doctors are documented, and whether the claim appears in guidelines or reviews. Start from the patient's concern rather than the video: ask what they hoped it would do, say what you found, and talk through what the evidence supports.
Going Deeper¶
Color key: GuideLessonGovernance
- Common Misconceptions Calibrating trust by task, including why a detector's flag is not evidence on its own.
- AI-Generated Images in Teaching Where generated images mislead, and what to use instead.
- AI Responsible Use Policy The sections on bias, intellectual property, and reporting output that is discriminatory or misleading.
Content last reviewed October 15, 2026. Review dates are maintained as described on the About page.