Intro
L1
Β·
Quiz
Β·
Lab
L2
Β·
Quiz
Β·
Lab
L3
Β·
Quiz
Β·
Lab
L4
Β·
Quiz
Β·
Lab
Module Test
AI and Misinformation Β· Introduction

Every New Medium Rewrites What We Believe

This course exists because the machinery of falsehood has just received its most powerful upgrade in five centuries.

When Johannes Gutenberg's press began producing books around 1450, European authorities immediately grasped the danger: identical copies of the same text, reaching thousands of people simultaneously, with no single gatekeeper controlling what was printed. Within decades, pamphlets carrying theological arguments, political manifestos, and outright fabrications were circulating faster than any institution could rebut them. The Reformation, the witch-trial panic, and a century of religious warfare were all, in part, products of a medium that democratized publication before anyone had developed the critical vocabulary to evaluate printed claims. It took roughly two hundred years β€” the span from Gutenberg to Descartes β€” for skeptical reading practices to catch up with the printing press.

In 2023, a photograph of a white plume of smoke rising near the Pentagon appeared on Twitter and briefly moved financial markets before being identified as an AI-generated image produced in roughly thirty seconds. That same year, audio cloned from eleven seconds of a politician's voice was used to make robocalls suppressing voter turnout in a New Hampshire primary. A law professor at George Washington University was falsely named in a ChatGPT-generated summary as having committed sexual harassment β€” the cases cited did not exist. The underlying models producing these outputs had been publicly available for less than eighteen months.

This course examines how AI systems generate false content, why human perception is poorly equipped to detect it, how platforms and institutions are responding, and what individual verification practices are actually worth developing. It will not tell you that AI is uniquely evil or that the problem is unsolvable. It will give you precise language, documented cases, and practiced judgment. The goal is not alarm but literacy β€” the same literacy that, eventually, the printing press also demanded.

AI and Misinformation Β· Module 1 Β· Lesson 1

Deepfakes, Hallucinations, and Synthetic Text: A Taxonomy

AI-generated false content is not one thing β€” and misidentifying the type is the first mistake.
What exactly is AI-generated false content, and why does the category matter for how we respond to it?

On March 22, 2023, a photograph circulated on Twitter showing a large explosion near the Pentagon building in Arlington, Virginia. The image had visual hallmarks of authenticity: a realistic smoke cloud, accurate surrounding geography, the kind of slightly blurred quality typical of a smartphone taken at distance. Within minutes it had been retweeted thousands of times. The S&P 500 dipped briefly. The Arlington Fire Department and the Pentagon both issued denials. The image was traced to an AI image generator β€” most analysts identified artifacts consistent with Midjourney β€” and had been seeded by accounts later linked to coordinated inauthentic behavior. The entire episode, from posting to debunking, took under ninety minutes. The market moved anyway.

This incident illustrates a critical distinction: the image was not a manipulated photograph of a real explosion. It was entirely synthetic β€” generated from a text prompt with no underlying photographic source. That distinction matters enormously for both detection and legal response. The tools for spotting manipulated photographs β€” metadata forensics, JPEG artifact analysis β€” are largely useless against natively generated imagery. A different type of false content demands a different type of scrutiny.

Three Distinct Categories of AI False Content

AI-generated false content divides cleanly into three families, each with different production mechanisms, different detection signatures, and different downstream harms. Conflating them produces confused policy and ineffective personal defense.

Deepfakes are synthetic audiovisual media in which a real person appears to say or do something they did not. The term was coined on Reddit in November 2017 by a user who created face-swapped pornographic videos using publicly available TensorFlow code and celebrity imagery. Within a year, commercial apps had made the technique accessible to non-experts. The underlying method β€” generative adversarial networks, later replaced by diffusion models β€” pits two neural networks against each other until the forgery becomes statistically indistinguishable from real footage. By 2022, audio deepfakes had become a significant fraud vector: the UK's energy firm CEO was defrauded of €220,000 in 2019 after receiving a phone call from what appeared to be his parent company's chief executive β€” the voice was synthesized.

Hallucinations are confident factual errors produced by large language models. Unlike deepfakes, they require no malicious intent. They emerge from the statistical architecture of transformer-based models, which predict likely next tokens without a dedicated truth-verification mechanism. In 2023, attorneys Steven Schwartz and Peter LoDuca submitted a legal brief in Mata v. Avianca that cited six court cases generated by ChatGPT β€” none of which existed. Judge P. Kevin Castel fined the law firm $5,000. The cases had official-sounding names, docket numbers, and invented quotations from real judges. ChatGPT had not lied in any intentional sense; it had generated plausible-sounding legal text because plausible-sounding legal text was statistically probable.

Synthetic propaganda and coordinated text is the third category: intentionally produced AI-written content designed to shift opinion at scale. In 2023, researchers at NewsGuard identified 49 websites in 10 languages that appeared to be almost entirely AI-generated news sites, producing hundreds of articles per day with minimal human editorial involvement. Unlike hallucinations, this content need not be factually wrong in every detail β€” it can selectively emphasize true facts to create a misleading impression, a technique researchers call computational propaganda.

Why the Taxonomy Matters

Detection methods that work for one category often fail for another. Reverse image search catches recycled photos but not natively generated images. Provenance metadata helps with video manipulation but is absent in text. Voice biometrics can flag cloned audio but not well-crafted written impersonation. Knowing which type you are evaluating determines which tools to reach for.

The Production Cost Collapse

In 2017, producing a convincing deepfake video required a team, significant compute, and weeks of work. By 2023, ElevenLabs' voice cloning tool required eleven seconds of audio and returned a cloned voice in real time. Midjourney V5, released in March 2023, produced photorealistic imagery from text prompts in under a minute at essentially zero marginal cost. The significance is not merely that these tools exist β€” it is that the cost-per-fabrication has collapsed toward zero, fundamentally shifting the economics of misinformation production.

When printing a pamphlet cost money and required access to a press, the volume of false content was naturally limited by production costs. Digital social media removed distribution costs. Generative AI has now removed production costs. For the first time in history, a single individual with modest technical skill can produce thousands of convincing false artifacts β€” images, audio, video, articles β€” per day.

Deepfake
Synthetic audiovisual media depicting a real person doing or saying something fabricated, produced using generative neural networks. Distinguished from edited media by the absence of an authentic source clip.
Hallucination
Confident, plausible-sounding factual error generated by a large language model, arising from statistical text prediction without a grounded truth-verification mechanism.
Synthetic Propaganda
Intentionally produced AI-written content designed to shift opinion at scale, including websites, social posts, and comments that simulate organic human discourse.
Computational Propaganda
The strategic use of automation, algorithms, and big data to distribute politically motivated misinformation, often through coordinated inauthentic behavior on social platforms.
The Core Principle of Lesson 1

AI false content is a category error waiting to happen. Before evaluating any suspicious content, the first step is to identify which type of AI-generated material you may be looking at β€” because the production mechanism determines the evidence trail, and the evidence trail determines what verification is actually possible.

Lesson 1 Quiz

Five questions Β· Select the best answer Β· Instant feedback
1. The March 2023 image of an explosion near the Pentagon was significant primarily because it was:
Correct. The image had no real-world photographic source β€” it was generated entirely by an AI image model, likely Midjourney. This distinction matters because traditional photo forensics tools (JPEG artifact analysis, reverse image search) are largely ineffective against natively generated images.
Not quite. The key finding was that no source photograph existed at all β€” the image was entirely synthetic, generated from a text prompt. That distinguishes it from edited photographs and makes standard forensic tools largely irrelevant.
2. In the Mata v. Avianca legal case of 2023, what type of AI false content caused the $5,000 fine?
Correct. Attorneys Schwartz and LoDuca submitted a brief citing six ChatGPT-generated court cases that did not exist. The model had produced plausible-sounding legal citations β€” names, docket numbers, quoted passages β€” because such text was statistically probable, not because any underlying cases were real.
The harm here came from LLM hallucinations: ChatGPT generated citations to nonexistent court cases, complete with realistic-sounding docket numbers and invented judicial quotations. No deepfake or audio was involved β€” just statistically plausible text with no factual basis.
3. Which statement best describes why "hallucinations" occur in large language models?
Correct. Transformer-based language models generate text by predicting which tokens are statistically likely to follow given the context. They have no dedicated fact-checking module, so they can produce confident, detailed, entirely false statements whenever false text would have been statistically probable in their training distribution.
Hallucinations are an architectural feature, not a content problem. The model predicts likely next words based on statistical patterns β€” it has no mechanism to verify whether what it generates is factually grounded. This means plausible-sounding falsehoods can emerge even from models trained on largely accurate data.
4. The UK energy company CEO fraud of 2019 is an example of which category?
Correct. The CEO received a phone call in which the voice of his parent company's executive was synthesized β€” convincingly enough that he transferred €220,000. This is a deepfake applied as a fraud vector, distinct from political misinformation or text-based hallucination.
This was an audio deepfake: the attacker synthesized the voice of a known executive to authorize a fraudulent wire transfer. It's a financial crime enabled by voice cloning technology, which falls squarely in the deepfake category rather than propaganda or hallucination.
5. What does the term "computational propaganda" specifically describe?
Correct. Computational propaganda is defined by its combination of automation, algorithmic targeting, and coordination β€” the goal is to make manufactured political content look like organic human discourse at scale. Individual deepfakes or isolated hallucinations are not computational propaganda unless they are part of a coordinated distribution strategy.
Computational propaganda specifically refers to coordinated, automated distribution of politically motivated content β€” not just any computer-generated falsehood. The defining features are scale, coordination, and the simulation of organic discourse, not simply the use of AI to generate individual pieces of content.

Lab 1: Classifying AI False Content

Interactive AI discussion Β· Minimum 3 exchanges to complete

Your Task

You will discuss real documented cases of AI-generated false content with the lab assistant. Practice applying the three-category taxonomy from Lesson 1: deepfakes, hallucinations, and synthetic propaganda. The assistant will present cases and challenge your classifications.

Try: "Give me a case and I'll classify it." Or: "What are the detection differences between deepfakes and hallucinations?" Or: "Why does it matter whether the Pentagon image was generated vs. manipulated?"
AI Lab Assistant
Lesson 1 Β· Taxonomy
Welcome to Lab 1. We're going to practice the three-category taxonomy of AI false content: deepfakes, hallucinations, and synthetic propaganda. I'll present documented cases and you'll classify them β€” or you can ask me to explain the detection logic for any category. Where would you like to start?
AI and Misinformation Β· Module 1 Β· Lesson 2

How Generative Models Produce False Content

Understanding the mechanism is the precondition for understanding the failure mode.
What technical processes inside generative AI models make false content an inherent output risk rather than a design flaw to be patched?

In February 2023, a user on the AI art platform Midjourney prompted the model to generate an image of Pope Francis wearing a white puffer jacket. The resulting image β€” the pontiff in a stylish Balenciaga-style coat, photographed from below with the kind of candid framing typical of celebrity street photography β€” circulated on Reddit, Twitter, and Facebook before being identified as synthetic. Several verified accounts treated it as real. The image looked real not because Midjourney had access to a photograph of Francis in such a jacket, but because it had been trained on millions of images of people in jackets and millions of images of the Pope, and the probability distribution over those two datasets happened to produce a highly convincing composite. No one designed the model to deceive β€” the deception emerged from pattern completion.

Diffusion Models: Iterative Noise Removal

Modern AI image generators β€” Midjourney, Stable Diffusion, DALL-E 3 β€” are built on diffusion model architecture. The training process works in two stages. In the forward pass, the model takes real images and progressively adds Gaussian noise until the image is indistinguishable from static. In the reverse pass β€” the actual generation step β€” the model learns to remove noise from a random field, guided by a text prompt processed through a separate language model. What emerges is an image that occupies a high-probability region of the model's learned distribution of "what images matching this prompt look like." It is not a photograph; it is a statistical consensus of visual patterns associated with the prompt's concepts.

The practical consequence is that diffusion models have no concept of truth or historical record. When asked to generate "a photograph of [real person] doing [action]," the model does not know whether that event occurred β€” it simply finds the region of visual space where images of that person and images of that action overlap. Photorealism is an emergent property of training scale, not an indicator of documentary authenticity.

Large Language Models: Next-Token Prediction

Text-generating models like GPT-4, Claude, and Gemini are trained on next-token prediction: given a sequence of tokens (roughly, words and word-parts), predict the most likely next token. During training on internet-scale text corpora, these models absorb the statistical regularities of human writing β€” the patterns of how facts, citations, names, and causal claims are structured. The result is a system that can produce fluent, well-structured, citation-formatted text that nonetheless has no reliable correspondence to reality.

The Mata v. Avianca hallucinations are a textbook example. Legal briefs have a highly distinctive format: case name, reporter citation, page number, quoted language from the opinion. ChatGPT had seen thousands of real legal briefs during training. When asked to find cases supporting a legal argument, it generated text in perfect legal-brief format, because that format was statistically probable β€” and populated the format with invented case names that sounded plausible because real case names follow recognizable patterns. The model was doing exactly what it was designed to do. The design does not include fact-checking.

Why Retrieval-Augmented Generation Helps But Doesn't Solve It

Retrieval-Augmented Generation (RAG) systems attach a search step to LLM generation, pulling real documents into context before generating a response. This reduces hallucination frequency by grounding the model in retrieved text β€” but it doesn't eliminate it. Models can still misquote retrieved documents, blend retrieved content with fabricated content, or retrieve the wrong documents when source material is ambiguous. RAG is a mitigation, not a solution.

Voice Cloning: Speaker Modeling

Voice synthesis systems like ElevenLabs, Microsoft VALL-E, and OpenAI's Voice Engine work by creating a statistical model of a speaker's vocal characteristics from a short audio sample β€” pitch, timbre, cadence, formant frequencies, prosodic patterns. Once encoded, that model can be conditioned on arbitrary text, generating speech that sounds like the target speaker saying anything. VALL-E, described in a January 2023 Microsoft Research paper, demonstrated this capability from three-second audio clips. The system cannot verify the speaker's consent and has no inherent limit on what text it can synthesize in the cloned voice.

The New Hampshire voter suppression robocalls of January 2024 β€” in which an AI-generated voice mimicking President Biden told Democratic primary voters not to vote β€” reached over 20,000 households. The call was traced to political consultant Steve Kramer, working for a rival campaign. Kramer later publicly admitted commissioning the calls to "start a conversation" about AI in elections. The FCC responded by clarifying that the Telephone Consumer Protection Act covers AI-generated voice calls, but the content had already reached its target audience before any regulatory action was possible.

Diffusion Model
An image generation architecture that learns to reverse a noise-addition process. Generates images by iteratively denoising a random field, guided by a text prompt, to produce outputs in high-probability regions of the model's learned visual distribution.
Next-Token Prediction
The training objective of large language models: given a sequence of tokens, predict the most statistically likely next token. Produces fluent text without any mechanism for verifying factual accuracy.
Speaker Model
A statistical encoding of a person's vocal characteristics, derived from an audio sample, used to condition speech synthesis systems to produce arbitrary text in a target speaker's voice.
The Shared Architecture of Failure

Across image diffusion, text generation, and voice cloning, the same pattern holds: the model produces outputs that occupy high-probability regions of a learned distribution. Photorealism, fluency, and vocal fidelity are all measures of distribution-fit, not measures of truth. This is why the false content these models produce can be so convincing β€” and why no amount of stylistic polish is a reliable indicator of factual accuracy.

Lesson 2 Quiz

Five questions Β· Select the best answer Β· Instant feedback
1. Why did the AI-generated image of Pope Francis in a puffer jacket fool many viewers?
Correct. The model had been trained on vast numbers of images of people in jackets and images of the Pope, so a "Pope in a puffer jacket" prompt landed in a high-probability region of the visual distribution β€” no actual source photograph required. Photorealism is a property of distribution-fit, not of documentary evidence.
The image was entirely synthetic. Diffusion models generate imagery by finding high-probability regions in their learned visual distributions β€” the overlap between "images of Pope Francis" and "images of stylish jackets" was sufficient to produce a convincing output with no real-world source photograph.
2. In the diffusion model architecture, what guides the noise-removal process toward a specific image?
Correct. The text prompt is encoded by a language model (CLIP or similar) and used to condition each step of the reverse diffusion process β€” steering the emerging image toward the region of visual space associated with the prompt's concepts. The model is not retrieving or modifying any specific training image.
The denoising process in diffusion models is conditioned on the text prompt, which is encoded by a language model component. This conditioning steers the iterative noise removal toward a region of visual space associated with the prompt β€” not toward any specific stored image or human-reviewed output.
3. Why did ChatGPT generate convincing-looking legal citations in the Mata v. Avianca case?
Correct. The model produced text in perfect legal-citation format β€” case name, reporter, page number, quoted language β€” because that format was statistically familiar from training data. It populated the format with invented case names that sounded plausible because real case names follow recognizable patterns. No deliberate deception, no database error, no bug β€” just next-token prediction without fact verification.
This was next-token prediction at work. The model had absorbed the format of thousands of legal briefs during training. When asked to find supporting cases, it generated text in that format β€” which meant producing plausible-sounding case names, docket numbers, and quotes β€” because that was what statistically followed the context. There is no legal module and no database lookup during generation.
4. Microsoft's VALL-E paper (January 2023) demonstrated voice cloning from how much source audio?
Correct. VALL-E demonstrated speaker modeling from three-second audio clips, a threshold that represents a dramatic reduction in the data required for convincing voice cloning. This level of efficiency means that any publicly available audio of a person β€” a short video clip, a voicemail, a conference recording β€” is potentially sufficient source material.
VALL-E demonstrated voice cloning from just three seconds of audio β€” not minutes. This is significant because it means the bar for obtaining sufficient source audio is extremely low: any short publicly available recording of a target speaker could serve as input to such a system.
5. What does Retrieval-Augmented Generation (RAG) do, and why does it not fully solve hallucination?
Correct. RAG reduces hallucination by providing real retrieved documents as context β€” but the model still generates text based on those documents using next-token prediction. It can misquote retrieved content, blend retrieved facts with confabulated details, or retrieve documents that don't actually address the query. RAG is a significant mitigation, not a solution.
RAG attaches a retrieval step to the LLM, pulling real documents into context to ground generation. This significantly reduces hallucination β€” but doesn't eliminate it. The model still generates text using next-token prediction and can misquote, blend, or misapply retrieved content. It is a mitigation, not a complete solution.

Lab 2: Probing Model Architecture

Interactive AI discussion Β· Minimum 3 exchanges to complete

Your Task

Discuss the technical mechanisms behind AI false content with the lab assistant. Focus on connecting architectural features (diffusion, next-token prediction, speaker modeling) to the specific failure modes they produce. The assistant will ask you to reason from mechanism to consequence.

Try: "Explain why photorealism in a diffusion output tells us nothing about accuracy." Or: "If I wanted to reduce LLM hallucination in a high-stakes application, what would I do?" Or: "Walk me through why VALL-E's three-second threshold matters."
AI Lab Assistant
Lesson 2 Β· Architecture
Welcome to Lab 2. We're going to connect the technical architecture of generative models to the specific false-content failure modes they produce. I'll ask you to reason from mechanism to consequence β€” and I'll push back if you conflate different types. What aspect of the architecture would you like to explore first?
AI and Misinformation Β· Module 1 Β· Lesson 3

Detection: What Works, What Doesn't, and Why

Every detection method has a corresponding evasion. Understanding both sides is what makes you useful.
Which detection methods for AI-generated false content have demonstrated real-world validity, and which are providing false confidence?

In January 2023, the AI detection company Originality.ai published a study claiming its tool achieved 94% accuracy at identifying AI-written text. Within three months, researchers at the University of Maryland demonstrated that simple paraphrasing β€” asking ChatGPT to rewrite its own output in a slightly different register β€” reduced Originality.ai's detection rate to near chance. The detector had been trained on a static corpus of AI outputs. The models had continued evolving. The arms race between generation and detection is not a technical problem awaiting a solution: it is a structural feature of the domain, because any detection method that becomes publicly known can be trained against.

This does not mean detection is useless. It means the reliability ceiling of automated detection is fundamentally different from the reliability ceiling of, say, a DNA test β€” and treating AI detectors with forensic-evidence confidence is itself a misinformation risk.

Image Forensics: Real Tools, Real Limits

C2PA Content Credentials (Coalition for Content Provenance and Authenticity) is the most promising structural approach: cryptographic metadata attached to an image at capture or generation that records provenance β€” camera make, GPS, timestamp, and for AI-generated content, the model used. Adobe, Microsoft, Sony, Nikon, and Leica are among the signatories. The BBC and Reuters have piloted C2PA on photojournalism workflows. The limitation is adoption: C2PA only works if the camera or AI tool attaches the credential and the platform displays it. Images can be re-saved as JPEGs, stripping metadata, before upload.

GAN-specific forensics β€” detecting the periodic artifacts left in images generated by older Generative Adversarial Networks β€” worked well from 2019 to 2021, when GAN-based tools like early StyleGAN produced detectable frequency-domain signatures. Diffusion models do not produce the same artifacts, rendering GAN detectors largely obsolete against current-generation tools.

Reverse image search (Google Images, TinEye) identifies recycled photographs used with false captions β€” a common tactic entirely distinct from AI generation. It is a useful tool in the misinformation toolkit but does not detect natively generated imagery, which has no prior appearance in any searchable index.

Text Detection: A Compromised Category

AI text detectors β€” GPTZero, Turnitin's AI detector, Originality.ai β€” typically work by measuring the statistical predictability of text (called perplexity) and the consistency of that predictability across passages (called burstiness). Human writing tends to have variable perplexity: some sentences are predictable, others surprising. LLM output tends to be consistently low-perplexity. This heuristic produces above-chance accuracy on unmodified LLM output β€” but it degrades rapidly under paraphrasing, it produces false positives on non-native English writing (which tends to be lexically predictable), and it cannot account for human-AI hybrid workflows.

In 2023, a University of California, Berkeley study found that GPTZero falsely flagged 24% of a sample of human-written op-eds as AI-generated. The risk of false accusation β€” with real consequences for students accused of academic dishonesty β€” is not theoretical. Several school districts temporarily banned AI detectors from disciplinary proceedings after documented false positives.

OpenAI released its own text classifier in January 2023, acknowledging only 26% true-positive detection rates on AI-written text, with a 9% false-positive rate on human text. OpenAI discontinued the tool in July 2023.

The Watermarking Approach

OpenAI, Google DeepMind, and others have developed statistical watermarking β€” invisibly biasing the token selection distribution during LLM generation in ways that can be detected with the right key, without noticeably affecting text quality. Google DeepMind's SynthID system, announced in 2023, applies this to both text and images. The limitation: watermarks can be removed by paraphrasing, translation, or fine-tuning on un-watermarked data. Open-source models can be released without watermarks, and watermarks tell us only about provenance, not about truth value.

Voice and Video: The Biometric Frontier

Voice deepfake detection systems β€” Resemble AI's Detect, Pindrop, ElevenLabs' own detection tool β€” analyze spectral features, micro-pause patterns, and formant transitions that tend to differ between natural speech and synthesized speech. These tools were achieving around 80–90% accuracy on 2022-era voice clones in controlled conditions. As synthesis quality improves, the acoustic artifacts that classifiers relied on progressively disappear.

Video deepfake detection has similar dynamics. DARPA's Media Forensics (MediFor) program, which ran from 2017 to 2022, produced classifiers that performed well on its benchmark datasets β€” and poorly when evaluated on newer-generation deepfakes. A 2023 study in the journal Nature Communications found that state-of-the-art video deepfake detectors degraded significantly when tested on compressed video (the format used by WhatsApp, Telegram, and most social platforms), falling to accuracy levels near a coin flip in some compression conditions.

C2PA
Coalition for Content Provenance and Authenticity. An industry standard for cryptographic content credentials embedded in images and videos at capture or generation, recording provenance information. Effective only where the full chain of custody preserves the credential.
Perplexity (in text detection)
A measure of how statistically predictable a piece of text is to a language model. LLM-generated text tends toward low, consistent perplexity; human writing tends toward more variable perplexity. The heuristic is exploitable by paraphrasing.
Statistical Watermarking
A technique for marking AI-generated content by introducing undetectable statistical biases in the generation process (e.g., token selection probabilities) that can later be identified by an authorized detector holding the key.
The Honest Summary of Detection

No single automated detection tool for AI-generated content is reliable enough to be used as sole evidence in high-stakes decisions. Provenance-based approaches (C2PA) are the most structurally sound but depend on unbroken adoption chains. Statistical classifiers offer probabilistic signals, not verdicts. The most reliable current practice is triangulation: multiple independent signals, combined with contextual and source evaluation, rather than reliance on any single detector score.

Lesson 3 Quiz

Five questions Β· Select the best answer Β· Instant feedback
1. Why did the University of Maryland demonstration undermine AI text detectors like Originality.ai?
Correct. When researchers simply asked ChatGPT to paraphrase its own output, Originality.ai's detection rate fell to near chance β€” even though the content was still AI-generated. This illustrates the structural problem: detectors are trained on snapshots of AI output, but models continue evolving and adapting, making the arms race asymmetric in favor of generation.
The Maryland study showed that paraphrasing AI output β€” a trivial modification β€” was sufficient to defeat the detector. This is not about fraudulent claims; it reveals a structural limitation. Any detection method that becomes publicly known can be worked around, because generation systems can be adapted to evade it.
2. What is the primary limitation of C2PA content credentials as a deepfake detection approach?
Correct. C2PA is a sound structural approach but depends on unbroken chain-of-custody. The moment an image is re-saved, cropped, or uploaded to a platform that strips metadata, the credential is lost. The system proves provenance where the chain is intact β€” it cannot detect AI generation in images that have been re-processed.
C2PA's limitation is the chain-of-custody problem. A cryptographic credential attached at generation is meaningless if the image is re-saved as a clean JPEG before distribution β€” a trivial operation. C2PA is not a watermark in the visual sense; it's metadata, and metadata is easily stripped.
3. OpenAI's own AI text classifier, released January 2023 and discontinued July 2023, had what detection performance?
Correct. OpenAI's own classifier achieved only 26% true positive detection on AI-written text β€” meaning it missed roughly three-quarters of AI-generated content β€” while falsely flagging 9% of human-written text as AI-generated. OpenAI acknowledged these limitations in the release documentation and discontinued the tool six months later.
OpenAI's classifier had a 26% true positive rate on AI-written text and a 9% false positive rate on human text. The company disclosed these figures at launch. A tool that misses 74% of AI content while wrongly accusing 9% of human writers is not operationally useful for high-stakes decisions β€” hence its discontinuation.
4. Why are GAN-specific forensic detection tools largely obsolete for detecting current AI-generated images?
Correct. GAN-based image generation (StyleGAN, ProGAN) produced characteristic periodic artifacts in the frequency domain that classifiers could detect. Diffusion models β€” which power Midjourney, Stable Diffusion, and DALL-E 3 β€” have a fundamentally different generation process and do not produce the same spectral signatures. GAN detectors are simply looking for the wrong type of evidence.
GAN detectors were trained to identify artifacts specific to the GAN architecture β€” periodic frequency-domain signatures introduced by convolutional upsampling. Diffusion models work differently and don't produce those artifacts. It's not a patch or a legal ruling; it's a fundamental architectural shift that rendered an entire class of detectors inapplicable.
5. What does a 2023 Nature Communications study reveal about video deepfake detectors and social media compression?
Correct. The Nature Communications study found that the compression algorithms used by major messaging and social platforms significantly degraded deepfake detector accuracy β€” in some conditions, to levels barely above chance. This is a critical practical problem: the conditions under which deepfakes are most likely to spread (mobile messaging platforms with aggressive compression) are precisely the conditions under which current detectors are least reliable.
The study found the opposite: social media compression degraded detector performance toward coin-flip accuracy. This is because detectors rely on subtle visual artifacts that compression removes. The practical implication is that deepfake detectors tested under clean lab conditions may perform dramatically worse on real-world platform-compressed video.

Lab 3: Evaluating Detection Claims

Interactive AI discussion Β· Minimum 3 exchanges to complete

Your Task

Practice evaluating detection tool claims with the lab assistant. You'll be asked to assess real-world detection scenarios, explain why particular tools succeed or fail in specific contexts, and reason about when automated detection is and isn't appropriate for high-stakes decisions.

Try: "A school administrator wants to use GPTZero to discipline students. What should they know?" Or: "Under what conditions is C2PA actually reliable?" Or: "Why does social media compression matter for video deepfake detection?"
AI Lab Assistant
Lesson 3 Β· Detection
Welcome to Lab 3. We're going to pressure-test detection claims. I'll present scenarios where someone wants to use an AI detection tool for a high-stakes purpose, and you'll evaluate whether that's appropriate and why. Ready to evaluate some real-world scenarios?
AI and Misinformation Β· Module 1 Β· Lesson 4

Harms, Scale, and the Institutional Response

Individual cases matter β€” but the systemic response determines whether the problem is manageable.
What documented harms has AI-generated false content caused, and how are platforms, governments, and civil society actually responding?

In October 2023, the Stanford Internet Observatory and the University of Washington's Center for an Informed Public published research identifying over 1,000 AI-generated news websites operating across 60 countries, producing content in 52 languages. The sites β€” tracked under the label "Pink Slime 2.0," referencing earlier waves of low-quality local news sites β€” were generating articles at volumes impossible for human editorial teams, in some cases publishing hundreds of pieces per day with minimal human involvement. Many carried advertising from legitimate brands, unaware of the content context. Unlike the misinformation of the 2016 election cycle β€” which required teams of human writers in places like Veles, North Macedonia β€” the 2023 wave required almost no human labor after initial setup.

Documented Harm Categories

Electoral interference. Beyond the New Hampshire robocalls discussed in Lesson 2, the 2024 election cycle saw documented AI-generated audio of multiple candidates circulated on social media in at least six countries, including Slovakia (where audio falsely attributed to candidate Michal Ε imečka discussing election fraud circulated days before the September 2023 election, too close to the vote for fact-checkers to respond), Bangladesh, and Pakistan. The European Digital Media Observatory documented at least 65 incidents of AI-generated election misinformation in EU member states during 2023–24 election periods.

Non-consensual intimate imagery (NCII). The Internet Watch Foundation's 2023 annual report identified a significant increase in AI-generated child sexual abuse material (CSAM), produced without any real victim using image generation tools. Separately, a 2023 study by the nonprofit Sensity AI estimated that 96% of deepfake videos online were non-consensual pornographic depictions of real women β€” celebrities and, increasingly, private individuals targeted by acquaintances using consumer-grade tools.

Financial fraud. The Hong Kong deepfake CFO fraud of January 2024 is the highest-documented single-incident financial loss attributed to AI false content: a finance employee was summoned to a video call in which all other participants β€” including a person presented as the company's CFO β€” were real-time deepfakes. The employee authorized a transfer of HK$200 million (approximately US$25.6 million). The attack used publicly available video footage of company executives to build the deepfake models.

Reputational harm. In 2023, false AI-generated articles claiming Baltimore County Public Schools Athletic Director Dazhon Darien had made racist statements to students were investigated by police. The articles were sophisticated enough to include false quotes attributed to real school board members. The case was later tied to a colleague, and Darien was briefly arrested β€” but only after the false content had circulated widely and the arrest itself became news.

The Velocity Problem

A consistent pattern across documented AI misinformation incidents is that the false content reaches its target audience before fact-checking infrastructure can respond. The Slovak election audio circulated 48 hours before the vote β€” within the blackout period that prohibited campaign materials. The Pentagon image moved markets in under 90 minutes. Fact-checking organizations, which typically require multi-step verification processes, are structurally slower than social media amplification. Institutional responses that depend on post-publication debunking are working against this asymmetry.

Platform Policy Responses

In 2024, Meta, Google, TikTok, and X (formerly Twitter) all announced policies requiring disclosure labels on AI-generated political content. Meta's policy, announced in November 2023 and expanded in 2024, requires advertisers to disclose when political ads contain "digitally altered or created" content. The implementation faced immediate challenges: Meta's own enforcement mechanisms rely partially on self-disclosure and partially on automated detection β€” both of which have the limitations documented in Lesson 3.

YouTube's policy, updated in November 2023, requires creators to disclose when realistic AI-generated content depicts real people, events, or places. YouTube can remove AI labels applied by creators who don't disclose, and can apply labels itself on content that could "confuse or mislead" viewers. Critics noted that "realistic" is undefined in the policy language and that enforcement at YouTube's scale requires significant automation.

The EU's AI Act, which entered force in August 2024, requires that AI-generated content be labeled as such β€” with specific provisions for deepfakes, requiring that viewers be clearly informed when they are watching or listening to synthetic media. The Act includes fines of up to 6% of global turnover for systemic violations. Enforcement begins in stages through 2026.

What Individuals Can Actually Do

Given the limitations of automated detection and the speed of AI content spread, individual verification practices focus on source and context rather than content analysis. The lateral reading technique β€” developed by researchers at the University of Washington's SIFT project and adopted by professional fact-checkers at AP and Reuters β€” involves opening new tabs to search for what independent sources say about a claim or content source, rather than analyzing the content itself. Studies of professional fact-checkers show they spend roughly 70% of their evaluation time on lateral reading and 30% on content analysis; novice readers reverse this ratio.

For suspicious images, reverse image search remains a useful first step for recycled content β€” and tools like Google Lens's "about this image" feature, launched in 2023, can surface publication history and context. For suspicious audio, the question of whether a clip was offered with full audio context (not excerpted) and whether the claimed speaker has acknowledged or denied it through official channels is often more reliable than any acoustic forensic tool.

The Structural Insight of Module 1

AI-generated false content is not primarily a technology problem. The underlying technologies β€” diffusion models, LLMs, voice cloning β€” are general-purpose tools with legitimate uses. The problem is the intersection of near-zero production cost, near-zero distribution cost, and human cognitive architecture that evolved to treat audiovisual evidence as reliable. The most durable responses will address that cognitive architecture β€” through media literacy, lateral reading, provenance infrastructure, and regulatory frameworks β€” rather than relying on automated detection that is structurally in a losing arms race with generation.

Lesson 4 Quiz

Five questions Β· Select the best answer Β· Instant feedback
1. The Hong Kong deepfake CFO fraud of January 2024 is notable because:
Correct. The attack was remarkable for its scale: an entire video conference populated with real-time deepfakes of known company executives, convincing enough that a finance employee authorized a HK$200 million (roughly US$25.6 million) wire transfer. The deepfake models were built from publicly available video footage of those executives β€” no insider access required.
The Hong Kong fraud involved a complete fake video conference β€” every other participant, including the apparent CFO, was a real-time deepfake constructed from public video footage. The scale of the deception (HK$200 million) and the technical sophistication (multiple simultaneous real-time deepfakes) make it the highest-documented single-incident financial loss attributable to AI false content.
2. What made the AI-generated election audio in Slovakia's September 2023 election particularly harmful?
Correct. The timing was the critical factor. Releasing false content during the blackout period β€” when campaign responses are legally restricted and voters are finalizing decisions β€” is a deliberate exploitation of the velocity asymmetry: false content spreads faster than corrections, and corrections themselves may be prohibited under blackout rules.
The timing was the decisive factor. The audio circulated during the pre-election blackout period, when campaign materials are prohibited β€” which meant the targeted candidate could not legally respond, and fact-checkers had insufficient time to debunk the content before polls opened. This is a deliberate exploitation of the velocity asymmetry between misinformation and correction.
3. According to studies of professional fact-checkers, what proportion of evaluation time do they spend on lateral reading versus content analysis?
Correct. Professional fact-checkers at organizations like AP and Reuters spend roughly 70% of their evaluation time on lateral reading β€” opening new tabs to search for independent source commentary on a claim β€” and only 30% analyzing the content itself. Novice readers invert this, spending most of their time examining the content directly, which is precisely where they are most vulnerable to being deceived by convincing-looking false content.
Studies show professional fact-checkers allocate roughly 70% of evaluation time to lateral reading β€” researching what independent sources say about a claim or its source β€” and only 30% to content analysis. Novices reverse this ratio, which makes them more susceptible to being deceived by plausible-looking content, because they're examining the very thing that has been designed to mislead them.
4. The EU AI Act provision on synthetic media requires:
Correct. The EU AI Act, which entered force in August 2024, requires that AI-generated synthetic media be disclosed to viewers β€” specifically including deepfakes depicting real people in realistic scenarios. Fines can reach 6% of global annual turnover for systemic non-compliance. Enforcement phases in through 2026, and critics note that "realistic" determination at platform scale still relies on automated systems with the limitations documented in Lesson 3.
The EU AI Act requires disclosure labeling for AI-generated synthetic media β€” viewers must be clearly informed when they're watching or listening to AI-generated content depicting real people or events. The fines are substantial (up to 6% of global turnover), and the Act represents the most comprehensive binding regulatory framework applied to AI-generated false content so far.
5. What does the Stanford Internet Observatory's "Pink Slime 2.0" research identify as the key difference from earlier waves of misinformation websites?
Correct. The defining difference is labor cost. The 2016-era Macedonian misinformation factories required rooms full of human writers. The 2023 wave required minimal human labor after initial setup β€” AI generation produced hundreds of articles per day at essentially zero marginal cost. This eliminates the production constraint that previously limited how much false content any single operation could generate.
The key distinction is production cost. Earlier misinformation operations β€” like the Macedonian content farms active during the 2016 US election β€” required significant human labor. The 2023 AI-generated news sites produced content at volumes impossible for human teams, with minimal ongoing labor. The economic model of misinformation has fundamentally changed when production costs approach zero.

Lab 4: Policy and Individual Response

Interactive AI discussion Β· Minimum 3 exchanges to complete

Your Task

Discuss the institutional and individual response landscape with the lab assistant. Evaluate the effectiveness of platform policies, regulatory approaches like the EU AI Act, and individual verification techniques. The assistant will ask you to assess tradeoffs and apply the module's core insights to real scenarios.

Try: "Evaluate Meta's disclosure policy for AI political ads." Or: "Why is lateral reading more effective than content analysis for most people?" Or: "Apply the module's structural insight to explain why detection tools will always lag generation."
AI Lab Assistant
Lesson 4 Β· Response & Policy
Welcome to Lab 4. We're going to evaluate institutional and individual responses to AI-generated false content. I'll present policy proposals and real-world scenarios β€” you'll assess their logic, their limitations, and how the module's core concepts apply. What aspect of the response landscape would you like to examine first?

Module 1 Test

15 questions Β· 80% required to pass Β· Covers all four lessons
1. Which term correctly describes a large language model producing a confident, detailed, entirely false court case citation?
Correct. A hallucination is a confident, plausible-sounding factual error generated by an LLM without malicious intent β€” it arises from next-token prediction without a truth-verification mechanism.
This is a hallucination β€” the model generates plausible-sounding false information because it predicts statistically likely text without verifying factual accuracy. Deepfakes are audiovisual synthetic media; computational propaganda is intentional persuasion at scale.
2. The March 2023 fake Pentagon explosion image was significant for AI misinformation research primarily because it was:
Correct. The image had no photographic source β€” it was natively generated, likely by Midjourney. It briefly moved the S&P 500 and was debunked within 90 minutes. Its significance lies in demonstrating that natively generated imagery can cause real financial harm and that traditional photo-forensic tools are inapplicable to it.
The image's significance was twofold: it was natively synthetic (not manipulated from a real photo) and it moved financial markets before being debunked. This demonstrated both the real-world harm potential of AI-generated imagery and the inadequacy of traditional forensic approaches against natively generated content.
3. In a diffusion model, photorealism in the output indicates:
Correct. Photorealism in diffusion outputs is a measure of distribution-fit β€” the image looks real because it resembles the statistical patterns the model learned from real images. It says nothing about whether the depicted scene ever occurred.
Photorealism is a property of distribution-fit, not documentary authenticity. A diffusion model produces realistic-looking outputs when they land in high-probability regions of its learned visual distribution β€” the content could be entirely fabricated and still look completely convincing.
4. The term "burstiness" in AI text detection refers to:
Correct. Burstiness describes how much perplexity varies within a text. Human writing alternates between predictable and surprising sentences; LLM output tends toward consistently low perplexity throughout. AI text detectors use this pattern as a heuristic β€” which paraphrasing can defeat.
Burstiness is the variation in perplexity across a text passage. It's a heuristic used by AI text detectors: human writing tends to have variable perplexity (some sentences predictable, others surprising), while LLM output tends to be uniformly low-perplexity. This difference is exploitable by paraphrasing.
5. Microsoft's VALL-E voice cloning paper demonstrated speaker modeling from:
Correct. VALL-E demonstrated convincing speaker modeling from three-second audio clips β€” a threshold that makes any short publicly available recording of a person sufficient source material for voice cloning.
VALL-E required only three seconds. This is the threshold that makes voice cloning a practical threat for virtually any public-facing individual β€” the source audio requirement is now shorter than most voicemail greetings.
6. Retrieval-Augmented Generation (RAG) reduces hallucination by:
Correct. RAG is a significant mitigation: by pulling real documents into the generation context, it reduces the frequency of fabrication. But the model still generates using next-token prediction and can misquote, misapply, or blend retrieved content β€” RAG is not a solution, it's a meaningful improvement.
RAG grounds generation in retrieved documents, reducing hallucination frequency substantially. But it doesn't eliminate it β€” the model still generates text using next-token prediction and can misrepresent retrieved content or blend it with confabulated material. It's a mitigation, not a cure.
7. GAN-specific forensic detection tools are largely ineffective against current AI image generators because:
Correct. GAN detectors target spectral artifacts produced by convolutional upsampling in GAN architecture. Diffusion models generate images through an entirely different process (iterative denoising) and don't produce those artifacts. The detectors are looking for the wrong evidence type.
The architecture changed. GAN models produced characteristic frequency-domain artifacts; diffusion models don't. When the industry shifted to diffusion-based generation (Midjourney, Stable Diffusion, DALL-E 3), GAN detectors became inapplicable β€” they're looking for evidence that the new models simply don't leave.
8. OpenAI discontinued its AI text classifier in July 2023 because:
Correct. OpenAI acknowledged at launch that its classifier achieved only 26% true positive detection on AI-written text with a 9% false positive rate on human text. A tool that misses three-quarters of AI content and wrongly accuses nearly one in ten human writers is not fit for any high-stakes purpose.
OpenAI's own documentation acknowledged the tool's poor performance: 26% true positive, 9% false positive. Detecting AI text is inherently difficult because the statistical properties that distinguish it from human text are subtle and can be removed by simple paraphrasing.
9. C2PA content credentials provide reliable provenance information only when:
Correct. C2PA is a chain-of-custody system. The credential is only meaningful when it has been preserved from capture or generation to the moment of viewing. A simple re-save as JPEG strips the metadata, breaking the chain. Where the chain is intact, C2PA provides strong provenance evidence; where it's broken, it provides nothing.
C2PA depends on an unbroken chain of custody. Credentials attach at creation and are verified at display β€” but the chain is broken the moment the image is re-saved, cropped, or processed through a platform that strips metadata. C2PA is a powerful provenance tool where adoption is complete; it provides no protection where the chain is broken.
10. The New Hampshire AI robocall voter suppression incident of January 2024 was attributed to:
Correct. Kramer publicly admitted commissioning the calls, claiming he did so to "start a conversation" about AI in elections. The calls reached over 20,000 Democratic primary voters with an AI-generated voice mimicking President Biden, telling them not to vote. The FCC clarified that the TCPA covers AI-generated voice calls.
Political consultant Steve Kramer admitted commissioning the calls. This case illustrates that AI election interference is not exclusively a foreign threat β€” domestic political operatives with modest resources can deploy voice cloning for voter suppression at scale.
11. The "lateral reading" technique, as practiced by professional fact-checkers, involves:
Correct. Lateral reading means opening new tabs to find independent source commentary on a claim or its publisher, rather than evaluating the content on its own terms. Professional fact-checkers use this approach for roughly 70% of their evaluation time, because it avoids the trap of being misled by convincing-looking false content.
Lateral reading means navigating away from the content to search for what independent sources say about it β€” its source, its claim history, its publisher's reputation. This is the opposite of content analysis: instead of examining the artifact, you research its context and reception in external sources.
12. Statistical watermarking in LLM outputs works by:
Correct. Statistical watermarks work by biasing the probability distribution over token choices during generation in ways that a detector with the key can identify, without noticeably affecting output quality. The limitation is that paraphrasing, translation, or fine-tuning can destroy the signal.
Statistical watermarking biases the token selection process during generation β€” certain token choices are systematically favored in ways that produce a detectable pattern for authorized detectors. The watermark is statistical, not textual, and doesn't affect perceived text quality. Paraphrasing can remove it.
13. The Stanford Internet Observatory's "Pink Slime 2.0" research (2023) identified how many AI-generated news websites across how many countries?
Correct. The scale β€” over 1,000 sites in 60 countries, 52 languages β€” illustrates the production cost collapse. These operations were impossible at this scale before AI generation tools became available; the human labor costs would have been prohibitive.
The research identified over 1,000 such sites in 60 countries producing content in 52 languages. The scale is the point: this volume of content production was simply impossible before AI generation tools became accessible, because the human labor costs would have been prohibitive.
14. The 2023 Nature Communications study on video deepfake detection found that social media compression:
Correct. Detectors rely on subtle visual artifacts that compression removes. The practical implication is that deepfake detectors trained on clean, high-resolution video perform dramatically worse on the compressed video formats used by WhatsApp, Telegram, and most social platforms β€” precisely where deepfakes are most likely to spread.
Compression degraded detector accuracy toward coin-flip levels in some conditions. This is a critical practical problem: the environments where deepfakes spread most β€” mobile messaging platforms with aggressive compression β€” are exactly where current detectors are least reliable. Lab benchmarks significantly overstate real-world performance.
15. According to the module's structural insight, the most durable responses to AI-generated false content will address:
Correct. The structural insight is that detection is in an asymmetric arms race with generation β€” any detection method that becomes known can be trained against. Durable responses build cognitive infrastructure (media literacy, lateral reading) and provenance infrastructure (C2PA, regulatory disclosure requirements) rather than depending on automated detection winning an arms race it is structurally positioned to lose.
Detection tools are in a structural arms race with generation tools β€” one they are positioned to lose, because generation can adapt to evade any known detection method. The most durable responses build human cognitive capacity (media literacy, lateral reading) and provenance infrastructure (C2PA, regulatory labeling requirements) that remain useful even as generation improves.