To understand why 2023 and 2024 felt like a rupture, you need to understand how slowly this was building — and then how suddenly it accelerated.

Photography was invented in the 1830s. Within decades, photographers were already manipulating images. In 1865, a famous portrait of Abraham Lincoln circulating in the United States turned out to be Lincoln's head pasted onto the body of the Southern politician John C. Calhoun. Nobody noticed for nearly a century. The manipulation was discovered in 1961. For most of photography's history, faking an image required physical skill — darkroom chemistry, careful cutting, precise alignment of negatives. Only experts could do it convincingly.

Adobe Photoshop changed this in 1990. Suddenly, digital manipulation was possible without a darkroom. But it still required skill, time, and software training. A professional retoucher might spend days on a convincing composite image. The barrier wasn't gone — it had just moved. Detecting fakes became a specialized job: forensic image analysts who could spot clone-stamped textures, lighting inconsistencies, and metadata anomalies.

Then, between 2017 and 2022, something different happened. Researchers developed a new class of AI system — called a generative model — that didn't manipulate existing images. It created new ones from scratch, pixel by pixel, guided only by a text description. By 2022, tools like DALL-E 2, Midjourney, and Stable Diffusion were available to the general public. You typed a sentence. You got a photorealistic image. No skill required.

The same acceleration happened with audio. In 2023, tools like ElevenLabs allowed anyone to clone a voice from a 30-second audio sample. Video synthesis followed. By early 2024, companies were offering "talking head" video generation — realistic lip-synced video of a person saying words they never said — for subscription prices comparable to a Netflix account.

Here's something important that often gets lost in the panic: humans have always lied. Propaganda has existed for thousands of years. Staged photographs go back to the Crimean War in the 1850s, when photographer Roger Fenton rearranged cannonballs on a road to make a more dramatic image. The desire to deceive with images is not new.

What changed is the cost and the skill floor. Before 2022, creating a convincing fake required one of three things: money (hire a professional), time (learn the skills yourself), or access (work at a studio or media organization). Those barriers weren't perfect, but they filtered out casual deception. Most people who wanted to spread a fake image had to either use an obvious fake — the kind that falls apart under mild scrutiny — or spend significant resources on a good one.

Generative AI removed those barriers almost entirely. Creating a convincing fake image now takes seconds. A convincing fake voice takes minutes. A basic deepfake video takes hours but requires no professional equipment. The result is that the volume of synthetic media in the world is increasing faster than any detection system can handle. Researchers at the University of Washington and other institutions have repeatedly shown that humans — including trained professionals — can correctly identify AI-generated faces only about 48% of the time. That's basically a coin flip.

This is the world you are growing up in. Not a world where some sophisticated state actor occasionally produces a fake to deceive millions — though that still happens — but a world where any bored, curious, or malicious person can produce convincing fakes at scale, for free, in minutes.

You might expect the biggest danger from AI fakes to be specific incidents — a fake video of a president declaring war, a fabricated recording of a CEO committing fraud. Those things are real risks. But researchers who study information warfare argue that the more insidious danger is something subtler: generalized distrust.

If people believe that any image, video, or audio clip could be fake, they stop trusting evidence entirely. This has already started happening. After the New Hampshire Biden robocall case in January 2024, journalists noticed a new phenomenon in their comments sections: even real, verified recordings of public figures were being dismissed as "probably AI." The fake didn't need to fool everyone. It just needed to make everyone doubt everything.

Researchers call this the "liar's dividend" — the idea that the existence of AI fakes benefits liars not just by creating false evidence, but by giving real liars a new defense. Any time a real recording catches someone in a lie, they can simply claim it's AI-generated. This was already happening in court cases by 2024, with defendants' lawyers raising AI-fakery arguments against genuine video evidence.

This is the part that should genuinely concern you — not just as a future voter or consumer, but right now, as someone who shares content online. Every time you share something without checking whether it's real, you're part of this system. Understanding how synthetic media works is no longer optional if you want to be an honest participant in public life.

In October 2023, a group of researchers at the Massachusetts Institute of Technology published an AI-generated photo of a destroyed city to illustrate an article about climate disaster risk. The image was clearly labeled "AI-generated illustration." Their argument: the image conveyed the emotional reality of a possible future more vividly than any existing photograph. It told a true story about a thing that could happen, using a fake image of a thing that hadn't happened yet.

On the other side: critics argued that using synthetic images to represent possible futures, even labeled, trains readers to accept fabricated scenes as emotionally legitimate evidence. That once you've felt the reality of a fake disaster image, you can't fully "unsee" it. That the emotional response it produces is the same regardless of the label, and that emotion — not the label — is what drives belief and action.

Here is the ethical question, and it does not have a clean answer: Is it acceptable to use a fake image to tell a true story, if the image is labeled? What changes if the label is small? What changes if the story is genuinely important? What changes if there are no real photographs of the thing being described? There are serious, thoughtful people on both sides of this. You will have to decide where you stand — and that decision will matter for how you create and share content for the rest of your life.

Most modern AI image generators use a technology called a diffusion model. The name sounds technical, but the idea is surprisingly understandable once you have the right analogy.

Imagine you have a photograph. Now imagine you add random noise to it — static, like a TV without signal — until the original image is completely buried under the noise. Now imagine you train a neural network to reverse that process: given a noisy image, predict what the slightly less-noisy version looked like. Do this thousands of times, in smaller and smaller steps, and the network learns to "denoise" images all the way back to clarity. That's diffusion.

When you give a diffusion model a text prompt — "a photo of a city council member accepting cash in a parking lot at night" — it starts with pure noise and progressively denoises it toward an image that matches your description. It's not searching through stored photographs. It's synthesizing a new image by applying learned patterns about how photographic elements relate to each other in space. It knows that "parking lot at night" means certain colors, certain light sources, certain textures. It assembles those patterns.

This process is extraordinarily good at capturing average visual relationships — the things that appear most consistently across the billions of training images. A face, centered in frame, well-lit, in a common pose: the model has seen this configuration millions of times and can render it flawlessly. But it struggles at structured complexity — things where there are explicit rules about how elements relate to each other that aren't purely visual.

Human hands are one of the most structurally complex objects a generative model encounters. A hand has five fingers, each with three joints, arranged according to biological rules that are rigid: four fingers extend from the palm in a specific fan pattern; the thumb opposes them from the side; finger lengths follow a fixed proportion. The model doesn't know these rules explicitly — it only knows what hands tend to look like in the training images. And when the hand is partially obscured, at an unusual angle, or in motion, the model has seen fewer examples and has to extrapolate. The extrapolation produces extra fingers, melded knuckles, or fingers that bend in anatomically impossible directions.

Text inside an AI-generated image fails for a related reason. The model learned what text looks like — the visual shapes of letters — but not what text means. So it produces arrangements of letter-like shapes that follow the visual rhythm of text without encoding any actual words. Hold a Midjourney image up to the light and read the street signs, the newspapers, the storefront lettering: they're typically gibberish that looks like language without being any language.

Edge artifacts — faces that blur at the boundaries of a frame, backgrounds that become inconsistent near the edges of a subject — occur because diffusion models generate images holistically rather than layering objects in physical space. A real photograph has a lens, a focal plane, and a consistent physics of light. A generated image has learned to look like a photograph without having a lens. Near the edges where the training data provided less guidance, the model's confidence drops and artifacts appear.

Reflections are another reliable tell. Mirrors and reflective surfaces require exact geometric consistency — the reflected image must be the mirror-reverse of the real object at the correct angle. The model has seen reflections but doesn't understand the geometry. It produces plausible-looking reflections that, on inspection, reflect different objects than what's in the frame, or reflect from the wrong angle.

Here is where things get complicated, and where you need to hold a difficult idea in mind: the very artifacts you just learned to look for are becoming less reliable as indicators, because the generators are improving.

Midjourney v4, released in late 2022, produced six-fingered hands routinely. Midjourney v6, released in December 2023, handles hands significantly better. As detection researchers identify specific failure modes and publish their findings, the model developers train on that feedback and patch the failures. This is not a conspiracy — it's just how any technology improves. But it means that the visual tells that were reliable in 2022 are less reliable in 2024, and the tells that are reliable in 2024 will be less reliable in 2026.

This is why understanding the mechanism matters more than memorizing the tells. The specific artifacts will change. The fundamental logic — that these systems generate by pattern-matching rather than by understanding physical reality — will persist for the foreseeable future. A system that generates by pattern-matching will always struggle more with structured complexity than with common configurations. The specific places it struggles will shift, but the underlying reason it struggles will remain.

There is also a separate detection technology: AI systems trained specifically to identify generated images. Companies like Hive Moderation, Illuminarty, and AI or Not operate AI detectors. These work by identifying statistical patterns in how pixels are distributed in generated versus photographed images. They are useful but not reliable: as of 2024, detection accuracy drops significantly when the generated image has been compressed, cropped, or run through a filter — all things that happen routinely when images are shared on social media.

In December 2023, the government of Belarus published a series of photographs showing what it claimed were Ukrainian military atrocities — images of soldiers committing acts of violence against civilians. Independent researchers at Bellingcat analyzed the images and identified them as AI-generated composites. Belarus denied this. The images were used in domestic propaganda broadcasts.

Here is the question: should social media platforms automatically remove images that AI detectors flag as synthetic — even knowing that AI detectors have significant error rates and might remove real documentary photographs of genuine atrocities? What is worse: leaving AI-generated propaganda up, or mistakenly removing evidence of real violence? Who should make that decision, and on what timeline? These are decisions being made right now — by engineers at Meta, Google, and TikTok — and they affect what billions of people see. There is no version of this problem that doesn't harm someone.

The word "deepfake" was coined on Reddit in 2017, when a user began posting realistic face-swapped celebrity videos using a technique derived from deep learning research. The name stuck even as the technology evolved far beyond simple face swapping.

Modern deepfake video works through one of two main methods. The first is face replacement: taking a video of one person and mapping the face of a different person onto it, matching lighting, skin tone, and head movement. This requires a significant amount of source footage of the target person — at least several minutes of varied facial expressions and angles. Until 2022, this was the primary limitation: you could deepfake celebrities and politicians who had extensive public video footage, but not private individuals.

The second method is audio-driven face synthesis, sometimes called "talking head" generation. Here, you start with a single photograph of the target person and a new audio recording. The system synthesizes video of the person's face moving to match the audio — generating realistic lip movements, micro-expressions, and subtle head movements that match the emotional content of the speech. By 2024, tools like HeyGen and Synthesia were offering this capability commercially for entirely legitimate purposes — mostly corporate training videos and multilingual content. The same tools can be misused.

The tell-tale signs of deepfakes follow from these mechanisms. Face replacement often shows inconsistencies at the jaw and neck boundary — the seam where the swapped face meets the original body. Audio-driven synthesis struggles with tooth visibility (teeth are structurally complex and move in ways that are hard to predict from audio alone), with eye blinking patterns (blink rates become statistically abnormal), and with the subtle physics of how skin moves with underlying muscle.

Video analysis is harder than image analysis for an obvious reason: you're evaluating many frames per second rather than a single frame. But this also gives you more data. Inconsistencies that might not appear in any single frame become visible across time.

The most reliable indicators of deepfake video as of 2024 include: unnatural blinking patterns (either too frequent, too infrequent, or blinks that don't fully close); facial boundary issues, especially around the jaw, ears, and hairline where the face swap boundary is hardest to blend; inconsistent lighting, where the subject's face is lit differently from the rest of the scene or the lighting doesn't match the claimed location; and lip-sync imprecision, where the mouth movements don't quite match the consonant sounds in ways that become obvious if you watch the mouth while listening closely.

There is also a behavioral tell that doesn't require any technical knowledge: sudden high-stakes statements from powerful people appearing first on anonymous social media accounts rather than official channels. Real announcements by military officials, presidents, and executives are almost always first announced through official channels — press releases, verified accounts, press conferences. If a major statement by a named official appears first as a viral video clip from an anonymous account, that's not a technical artifact — it's a provenance red flag that any careful reader can notice.

This behavioral check is actually more reliable right now than technical visual analysis, because visual quality is improving faster than behavioral conventions are changing. Institutions still have communication norms. Deepfakes can't easily fake an institutional process.

Most public discussion of deepfakes focuses on political disinformation. But the majority of deepfakes that actually circulate online — by volume — are not political. According to research by Sensity AI published in 2023, over 96% of deepfake videos they catalogued were non-consensual intimate imagery: synthetic videos of real people, mostly women, generated without their knowledge or consent.

This is a different harm than political disinformation, but in some ways a more immediate one. In 2024, several high schools in New Jersey, Pennsylvania, and Spain reported incidents where students had generated non-consensual deepfake images of classmates using free tools available through standard app stores. The victims — almost exclusively girls — experienced documented psychological harm. Legislation in several US states moved to criminalize non-consensual deepfake imagery, but enforcement lagged significantly behind the technology's accessibility.

This matters for this course in a specific way: the ethical stakes of deepfake technology are not abstract or distant. They exist at the scale of your school. They affect people your age. Understanding how the technology works is not just about reading news more carefully — it's about understanding a capability that already exists in the hands of people in your immediate environment, and that has been used to cause specific, documented harm to teenagers.

In 2024, the Indian film industry used AI deepfake technology to restore the performance of actor Irrfan Khan, who died in 2020, for a film he had been scheduled to appear in. His family consented. The filmmakers argued they were honoring his legacy and fulfilling a creative collaboration he had agreed to. Critics argued that no one — not family members, not studio executives — can meaningfully consent on behalf of a dead person's likeness, and that normalizing posthumous AI performance creates a framework in which studios will eventually use deceased actors without family approval at all.

Here is the question: does consent from a living family member make AI resurrection of a deceased person's likeness ethically acceptable? What if the person left a will saying they wanted their likeness to be used this way? What if they said nothing? What if they actively said they didn't want it — but their estate, which controls the rights, disagrees? There is currently no legal consensus and no industry standard. These decisions are being made film by film, contract by contract. Where would you draw the line?

Visual tells change as technology improves. Detection software lags behind generation quality. No checklist of artifacts will remain valid for more than a couple of years. What remains stable is a set of questions about provenance — the origin and history of a piece of media — that apply regardless of what the technology looks like.

Question 1: Where did this first appear? Not who shared it with you, but where in the chain of distribution it originated. An image that first appeared on a verified news organization's account has a different credibility profile than one that first appeared on an anonymous account created three days ago. Most people never ask this question. Asking it makes you unusual.

Question 2: What is the claimed context, and can it be independently verified? An image of a fire means nothing without knowing where and when. A video of a public figure saying something means nothing without knowing when and where it was recorded. Can you find independent corroboration of the claimed context from a source that isn't just re-sharing the original?

Question 3: Who benefits if this is believed? This doesn't tell you whether it's real, but it tells you who has a motive to produce it. A deepfake of a political candidate saying something embarrassing right before an election benefits their opponent. Understanding motive doesn't prove fabrication — but it tells you where to look hardest.

Question 4: Have I seen the original? "Going to the source" sounds obvious but almost nobody does it. Many viral images and videos are screenshots of screenshots, compressed and cropped and re-uploaded until any metadata that might reveal their origin is long gone. Asking for the original file — and thinking about whether it's available — is a basic discipline that professional fact-checkers apply automatically.

Alongside the four questions, there are practical tools worth knowing — not because they're infallible, but because they're fast and often decisive.

Reverse image search — available through Google Images, TinEye, and Bing — lets you upload an image and find other places it has appeared online. This is the fastest way to catch misattributed real images: photographs from old events or different countries being relabeled as something current. In the Maui wildfires case, several "Lahaina fire" images were identified within minutes as Greek wildfire photographs from 2021 via reverse image search. The technique doesn't identify AI-generated images (which may not appear anywhere else online) — but it does catch the most common form of visual misinformation: real images misused out of context.

Image metadata (called EXIF data) can reveal the camera model, GPS coordinates, and timestamp of a photograph — but only if it hasn't been stripped. Most social media platforms automatically strip EXIF data when images are uploaded, so this is useful primarily for images received directly as files rather than screenshots shared from feeds.

Geolocation — cross-referencing visual details in an image (building styles, signage, vegetation, street layout) against satellite imagery — is a skill developed by professional OSINT (Open Source Intelligence) analysts. Organizations like Bellingcat have used geolocation to verify or debunk dozens of high-stakes images in conflict zones. For most everyday purposes, this level of analysis isn't necessary — but knowing it exists matters when stakes are high.

AI detection tools like Hive Moderation and Illuminarty are useful as a first-pass signal, not a verdict. Use them to flag an image for further investigation, not to conclude it's fake. As of 2024, their false-positive rates on compressed images are high enough that a positive detection result should increase your scrutiny, not end your inquiry.

Here is a framing that most media literacy courses avoid because it sounds preachy — but which is actually just accurate: when you share something, you are making an implicit claim that it's worth other people's attention. You are vouching for it, at least at the level of "this seemed interesting and plausible enough to pass on." In a world of mass social media, that vouching multiplies rapidly. A single share by someone with 200 followers, if re-shared by three people with larger followings, can contribute to something reaching tens of thousands of people by morning.

This doesn't mean you should never share anything unless you've verified it. That's an impossible standard. What it means is that there's a proportional responsibility: the more dramatic the claim, the more harmful the implications if it's wrong, and the more shareable the content, the more scrutiny you owe it before passing it on. A funny meme about a celebrity looking weird at an award show? Low stakes, share away. A video showing a politician committing a crime right before an election? That warrants the 30 seconds of checking that most people skip.

The people who get this right aren't the ones who are paralyzed by skepticism. They're the ones who've developed a fast, automatic habit: before sharing something that makes me feel something strong — outrage, shock, vindication — pause and ask the four questions. Strong emotional reactions to media are exactly the condition in which fabrications are most effective, because emotion and scrutiny don't coexist easily. Training yourself to apply scrutiny precisely when you feel the most certain is the core skill this entire course is trying to build.

In early 2024, a researcher at Stanford published a proposal arguing that all AI-generated media should be required by law to carry an invisible digital watermark — a code embedded in every pixel that identifies the content as synthetically generated, readable by any detection tool. The proposal had significant support from the AI safety research community. It also had significant opposition from civil liberties lawyers, who argued that mandating watermarks on all AI-generated content creates a government-controlled registry of who is producing synthetic media and when — a surveillance infrastructure that could be used to track political speech, artistic expression, and private communications.

Here is the question: should governments require all AI-generated images and videos to carry an identifying watermark, knowing that this creates a detection infrastructure that also enables surveillance of who creates what? What if watermarks can be stripped? What if only democratic governments implement it, and authoritarian ones don't — giving them an advantage in disinformation? What if the watermark system is controlled by a private company rather than a government? There are no easy answers here. These debates are happening in actual legislative chambers right now. Where would you weigh in?