Loading module…
Loading module…
Module 1 of 6
Synthetic media — images, video, audio, and text generated or altered by AI — has crossed the threshold where it is cheap, abundant, and often indistinguishable from the real thing at a glance. This breaks a default humans have relied on for all of history: that a photograph, a recording, a video showed something that happened. That default is now unsafe, and the ability to detect and verify synthetic content has become a core literacy — for investigators, journalists, security teams, and anyone who must decide whether what they are looking at is real. This course teaches that literacy concretely: the actual detection techniques, forensic methods, and verification workflows that separate authentic content from synthetic, along with the honest truth about how reliable those techniques are.
That honest truth is the frame for the whole course, and it must be stated at the start because it governs everything: detection is imperfect, it is losing an arms race against generation, and no single technique is reliable on its own. The techniques in this course genuinely help — they catch a great deal of synthetic content and they are essential skills — but anyone who promises a reliable "AI detector" is selling false confidence, because generation improves continuously and detection chases it, and today's reliable tell becomes tomorrow's fixed flaw. So this course teaches detection techniques and the disciplined, multi-signal, provenance-first verification approach that works even when any single detector fails, because that disciplined approach — not faith in a detector — is what actually protects you in a world of improving synthetic media. We begin with the landscape: what synthetic media is, why detection is hard, and the honest state of the field.
To detect synthetic content you must first know what exists, because the types differ fundamentally in how they are made, how they fail, and how they are detected. Synthetic media is not one thing but several distinct technologies, each with its own signatures and its own detection approaches.
Synthetic images are generated from scratch by AI (a described scene rendered as a photorealistic image) or existing images altered by AI (objects removed or added, faces swapped, scenes extended). Older generation methods (GANs) and newer ones (diffusion models, from the AI courses) produce images that are increasingly photorealistic, and their detection rests on the subtle artifacts these generation processes leave (Module 4). Synthetic images range from wholly invented scenes to subtle edits of real photos, and the detection challenge differs: a wholly-generated image may carry generation artifacts throughout, while a subtle edit leaves most of a real image intact with only a manipulated region to find.
Deepfake video is AI-generated or AI-altered video, most notoriously face-swaps (one person's face mapped onto another's body/performance) and face-reenactment (making a real person appear to say or do things they did not), but also fully synthetic video (entirely generated scenes). Video deepfakes are harder to produce convincingly than images because they must be coherent across many frames and often synchronized with audio, which creates detection opportunities in the temporal and synchronization dimensions (Module 2) — inconsistencies between frames, unnatural motion, imperfect lip-sync — that single images do not have. Synthetic audio is AI-generated or cloned voice — making a synthetic voice say anything, or cloning a specific real person's voice from samples. Voice cloning has become alarmingly good and is a major vector for fraud (impersonation scams), and its detection rests on subtle acoustic artifacts and unnaturalness (Module 2). Audio deepfakes are particularly dangerous because audio carries strong authenticity intuitions (we trust a familiar voice) while being easier to fake convincingly than video.
Synthetic text is AI-generated writing — the language-model output of the AI courses — used for everything from fake reviews and disinformation to impersonation and academic dishonesty. Text is the hardest synthetic media to detect (Module 3), because unlike images, video, and audio, text has no "physical" signal to analyze for artifacts — it is just words, and AI-generated words can be genuinely indistinguishable from human words, making text detection the least reliable of all. Knowing these types matters because detection is type-specific: the techniques for spotting a synthetic image (visual artifacts) are entirely different from those for synthetic audio (acoustic analysis) or synthetic text (statistical and contextual patterns), and a verification workflow must apply the right techniques to the right media type. The types also differ in detection difficulty — roughly, wholly-synthetic images and video with artifacts are more detectable, subtle edits and cloned audio are harder, and text is hardest — which shapes how much confidence any detection can offer for each type.
Key Insight: Synthetic media is several distinct technologies, each with its own signatures and detection: synthetic images (generated or AI-edited, detected via generation artifacts), deepfake video (face-swaps, reenactment, fully synthetic — detectable in temporal and lip-sync dimensions single images lack), synthetic audio (generated or cloned voice — dangerous because we trust familiar voices, detected via acoustic artifacts), and synthetic text (hardest of all, because text has no physical signal to analyze — AI words can be genuinely indistinguishable from human words). Detection is type-specific, so a verification workflow applies the right techniques to the right media.
The central fact governing detection — and the one most often misunderstood — is that detection is locked in an arms race with generation that generation is winning, and understanding this arms race is what sets honest expectations for every technique in the course. Detection is not a solved problem with reliable tools; it is a continuous struggle against a target that keeps improving.
The arms race works like this: generation methods produce synthetic content with certain flaws and artifacts; detection methods learn to spot those flaws; then generation methods improve to eliminate the flaws detection learned to spot; then detection must find new flaws; and so on, continuously. Every reliable detection tell has a limited lifespan, because once a tell is known (and detection research is public), generation methods are trained to eliminate it — often specifically trained against detectors, so that the next generation of synthetic content no longer exhibits the artifact the detector relied on. This is why detection perpetually chases generation: the artifacts that make today's synthetic content detectable are precisely the flaws that tomorrow's generation methods are being improved to remove. A course that taught only "look for this specific artifact" would be obsolete quickly, because that artifact will be fixed; the durable skills are the general forensic approaches and the multi-signal verification discipline that survive the fixing of any specific tell.
Generation has structural advantages in this race that are worth understanding, because they explain why generation tends to win. First, generation improves continuously and rapidly (the AI courses' scaling and progress), while detection can only respond to what generation has already produced — detection is inherently reactive, always a step behind the latest generation. Second, generation can be trained specifically to defeat detection (adversarially), directly optimizing to eliminate the signals detectors use, while detection cannot as easily anticipate generation's next improvement. Third, generation needs to fool a viewer only well enough for the purpose, while detection must reliably distinguish synthetic from real across all the variation of both — an asymmetric burden. Fourth, the sheer volume and diversity of synthetic content outpaces detection's ability to characterize it. These structural factors mean the honest expectation is that detection will remain imperfect and will lag generation for the foreseeable future — not that detection is useless (it catches much, especially less-sophisticated synthetic content), but that it cannot be relied upon as a reliable gatekeeper, because sophisticated, current synthetic content is increasingly designed to pass it.
The crucial implication, which shapes the whole course, is that detection techniques are necessary but not sufficient, and the reliable protection is verification discipline, not detector faith. Because any single detection technique can be defeated and detection lags generation, relying on a detector — "the tool says it's real, so it's real" — is exactly the false confidence that sophisticated synthetic content exploits. The reliable approach, developed through the course and assembled in Module 6, is multi-signal verification: combining multiple detection techniques (no single one trusted), and crucially anchoring on provenance and corroboration (where did this come from, is it consistent with other evidence) rather than on artifact-detection alone — because provenance and corroboration survive even when artifact-detection fails against sophisticated fakes. This is the same lesson the AI courses reached about media literacy: in a world of improving synthetic media, you cannot reliably tell real from fake by examining the artifact, so you must shift weight to the pathway (provenance) and the corroboration (does independent evidence agree). The detection techniques this course teaches are real and valuable — they catch much and they are essential skills — but they are held within this honest frame: useful signals in a multi-signal discipline, never a reliable detector to trust, because the arms race guarantees no detector stays reliable.
Key Insight: Detection is losing an arms race with generation: every reliable tell has a limited lifespan because once known (research is public), generation is trained to eliminate it — often adversarially, directly optimized to defeat detectors. Generation improves continuously while detection can only react, so detection perpetually lags. The honest expectation: detection stays imperfect and behind, catching less-sophisticated fakes but not reliably gatekeeping sophisticated current ones. The implication governing the course: detection techniques are necessary but not sufficient — the reliable protection is multi-signal verification anchored on provenance and corroboration, not faith in any detector.
Given the arms race, it is essential to be precise and honest about what detection tools can and cannot actually do, because the market is full of overclaiming and the false confidence it produces is itself a danger. This section is the reality check that keeps you from the two failures of naive trust and naive dismissal.
The honest state, by media type, is roughly this. Image detection tools and techniques catch a meaningful fraction of synthetic images, especially wholly-generated ones with characteristic artifacts, but they miss sophisticated current generations and subtle edits, and they produce both false positives (flagging real images as synthetic) and false negatives (missing synthetic ones). They are useful signals, not reliable verdicts. Video (deepfake) detection similarly catches many deepfakes, particularly less-sophisticated ones with temporal and sync artifacts, but sophisticated deepfakes increasingly pass, and detection reliability degrades as generation improves. Audio detection is harder and less mature, catching some synthetic audio but missing good voice clones, which is concerning given audio's fraud vector. Text detection is the least reliable by far — AI-text detectors have high error rates in both directions, cannot reliably distinguish AI from human text especially as models improve, and notoriously produce false positives (flagging human writing as AI) that have caused real harm (wrongly accusing people). The honest summary: detection tools across all types are useful-but-unreliable signals, better at catching unsophisticated synthetic content than sophisticated, prone to errors in both directions, and degrading as generation improves — with text detection so unreliable it should be treated with particular caution.
Two specific dangers of detection tools deserve emphasis because they cause real harm. The false positive danger: detectors flag real content as synthetic, and acting on this (dismissing genuine evidence as fake, or accusing someone of using AI when they did not) causes harm — the wrongful academic accusations from unreliable text detectors are a documented, serious example. False positives mean a detector's "this is synthetic" is not reliable, so you cannot use it to confidently condemn content as fake any more than to confidently clear it as real. The false confidence danger: a detector's verdict, especially with a confidence percentage, carries an unearned authority (the AI courses' bias-laundering, in detection form) that leads people to trust it beyond its reliability — "the tool says 94% real" feels authoritative but the 94% may be meaningless given the tool's actual error rate against current generation. Both dangers point the same way: detector outputs are signals to weigh, not verdicts to trust, and treating them as verdicts — in either direction — produces the errors that the tools' unreliability guarantees.
The constructive conclusion is not that detection is useless but that it must be used correctly: as one input among several in a verification discipline, weighted by its honest reliability, never as a sole or trusted verdict. A detector flagging content as possibly synthetic is a reason to investigate further, not a conclusion; a detector clearing content is not a guarantee of authenticity; and the confidence to draw about a piece of content comes from the whole verification picture (multiple signals, provenance, corroboration), not from any detector alone. This correct use — detectors as weighted signals in a multi-signal discipline — captures the genuine value detection tools offer (they do catch much, and their flags are useful leads) while avoiding the false confidence their overclaiming invites. The person who understands the honest state of detection uses the tools for what they are worth and no more, and anchors their actual conclusions on the verification discipline that does not depend on any single tool's unreliable verdict. That honest, disciplined use is the alternative to both naive trust (believing detectors) and naive dismissal (ignoring a useful signal), and it is what the rest of the course builds.
Key Insight: Honestly, detection tools are useful-but-unreliable signals across all media types — better at catching unsophisticated than sophisticated synthetic content, error-prone in both directions, degrading as generation improves, with text detection so unreliable it warrants particular caution. Two harms: false positives (flagging real content as fake — the wrongful AI-text accusations are a documented example, so "synthetic" verdicts aren't reliable either) and false confidence (a detector's percentage carries unearned authority). The correct use: detector outputs are weighted signals in a multi-signal verification discipline, never sole or trusted verdicts — capturing their genuine value without the false confidence their overclaiming invites.
Before the specific techniques, establish the mindset that makes detection work in practice — because the right mental approach is what separates effective verification from both gullibility and paranoia. The detection mindset treats content authenticity as a question answered by accumulating and weighing signals, not by a single test that returns a verdict.
The foundational stance is calibrated skepticism — neither trusting content by default (gullibility, unsafe in a synthetic-media world) nor disbelieving everything (paranoia, which is its own failure, as the AI courses noted, because dismissing genuine content as fake serves deceivers and destroys shared truth). Calibrated skepticism means holding authenticity as an open question to be assessed rather than assumed in either direction, and calibrating the assessment effort to the stakes — a casual image warrants a glance, while content that will drive a decision, be published, or damage someone warrants rigorous verification. This calibration is the same appropriate-trust discipline from the AI courses, applied to media authenticity: match the verification effort to what is at stake, be most rigorous where being wrong costs most, and neither exhaust yourself verifying trivia nor wave through consequential content unchecked. The mindset starts here: authenticity is a question, assessed with effort proportional to stakes.
The core operational principle is signals, not verdicts — building authenticity assessment from multiple signals, each of which shifts the probability, rather than seeking a single test that decides. No single signal is reliable (the arms race guarantees this), but multiple independent signals pointing the same way build genuine confidence, because while any one can be defeated or wrong, sophisticated fakes that defeat all of them simultaneously are much rarer than fakes that defeat one. So the detection mindset accumulates signals — a detector's flag, a visual artifact, a metadata inconsistency, a provenance problem, a corroboration failure — and weighs them together, reaching a confidence level (not a binary verdict) from the whole picture. This multi-signal approach is robust precisely where single-signal detection is fragile: it does not collapse when one technique fails, because it never relied on one technique. And it produces honest, graduated conclusions ("multiple signals suggest this is likely synthetic" or "signals are mixed; authenticity is uncertain") rather than false binary certainty, which matches the actual epistemic situation where certainty is rarely available.
The mindset also includes knowing what each signal is worth and anchoring on the most durable signals. Signals differ in reliability and durability: artifact-detection signals (visual/acoustic flaws) are useful but degrade as generation improves (the arms race), while provenance signals (where did this come from, what is its chain of custody) and corroboration signals (does independent evidence agree) are more durable because they do not depend on the synthetic content's quality — a perfect fake still has to come from somewhere and still has to be consistent with other evidence, and provenance and corroboration test exactly those, regardless of how good the fake looks. So the mature detection mindset, while using all signals, anchors especially on provenance and corroboration as the signals that survive the arms race, and treats artifact-detection as valuable but degrading support. This is the strategic core of detection in the synthetic-media age: because artifact-detection is losing the arms race, weight shifts to provenance and corroboration, which are not in an arms race with generation quality at all. The detection mindset — calibrated skepticism, signals not verdicts, knowing each signal's worth, and anchoring on the durable signals of provenance and corroboration — is what makes verification work even as generation improves, and it is the frame within which every specific technique in the following modules operates: each technique is a signal, weighted by its worth, combined with others, anchored on provenance, toward a calibrated confidence rather than a fragile single-test verdict.
Key Insight: The detection mindset: calibrated skepticism (authenticity is an open question assessed with effort proportional to stakes — neither gullible nor paranoid), and signals-not-verdicts (accumulate multiple independent signals, each shifting the probability, toward a calibrated confidence rather than a fragile binary test — robust because sophisticated fakes rarely defeat all signals at once). Know each signal's worth and anchor on the durable ones: artifact-detection degrades as generation improves (the arms race), but provenance (where it came from) and corroboration (does independent evidence agree) don't depend on the fake's quality, so weight shifts to them. Every technique ahead is a signal within this frame.
Understanding why synthetic-media detection matters — the concrete harms it defends against — motivates the discipline and calibrates where to apply it hardest. The stakes are real and rising, and they span personal, institutional, and societal harm.
At the personal and interpersonal level, synthetic media enables fraud and abuse that detection defends against. Voice cloning powers impersonation scams — a cloned voice of a family member or executive requesting an urgent transfer, a fraud vector that has caused real financial losses and that detection literacy (recognizing the possibility, verifying through a second channel) defends against. Synthetic images and video enable non-consensual fabricated content targeting individuals — a grave harm the AI courses flagged — where the ability to detect and prove fabrication matters for the victim. Fabricated evidence in personal disputes, fake profiles in social engineering, and manipulated content in harassment all make personal-level detection literacy a genuine protection. The concrete defensive skill here is often simple and high-value: knowing that a voice or video can be faked, and therefore verifying an urgent or surprising request through an independent channel rather than trusting the media — a discipline that defeats the most common synthetic-media fraud without any forensic tool.
At the institutional and professional level, synthetic media threatens the integrity of evidence and information that institutions rely on. Journalists must verify that images and videos are authentic before publishing, because publishing a fake is a serious failure and because fakes are deliberately fed to media to deceive. Investigators and analysts (the OSINT of Course 33) must assess whether the media they collect is authentic, because building intelligence on a fake corrupts the product. Businesses face synthetic-media fraud, fake documentation, and manipulated communications. Legal and security contexts must grapple with whether audio/video evidence is genuine. In all these, professional detection literacy — the techniques and the verification discipline — is a core competence, because the institution's function depends on distinguishing real from synthetic, and the cost of getting it wrong (publishing a fake, acting on fabricated evidence, being defrauded) is high. This is where the course's techniques earn their keep professionally.
At the societal level, synthetic media threatens the shared information environment in the way the AI courses described — the erosion of the ability to trust recorded media as evidence of reality, and the second-order harm where the mere possibility of fakes lets genuine media be dismissed as possibly fake (the "liar's dividend"). Detection literacy is part of the societal defense: a population that understands synthetic media, knows it cannot trust media by default, and has some verification discipline is more resilient to synthetic-media manipulation than one that either naively believes everything or cynically believes nothing. The individual skill aggregates into societal resilience. Understanding these stakes calibrates the discipline: apply detection most rigorously where the harm of being fooled is greatest — consequential decisions, publication, evidence, financial transactions, anything targeting a person — and recognize that the literacy is not academic but a defense against concrete, rising harms across every level from personal fraud to societal manipulation. The stakes are why this course exists, and why the disciplined verification it teaches — imperfect but real — is worth the effort: not because detection is reliable (it is not), but because the alternative, trusting synthetic-media by default, is a vulnerability that real harms exploit.
Key Insight: The stakes span levels: personal (voice-cloning fraud, non-consensual fabrication, harassment — often defended by the simple discipline of verifying surprising requests through an independent channel rather than trusting the media), institutional (journalists, investigators, businesses, legal/security contexts whose function depends on distinguishing real from synthetic, at high cost of error), and societal (erosion of trust in recorded media, plus the "liar's dividend" where real content is dismissed as possibly fake). Detection literacy is concrete defense against rising harms, calibrated to apply hardest where being fooled costs most — worth the effort not because detection is reliable, but because trusting synthetic media by default is a vulnerability real harms exploit.
This closing section gives you the concrete starting toolkit and the triage approach that the rest of the course builds into full forensic technique — enough to begin verifying today, held within the honest, multi-signal, provenance-anchored frame.
For the most common real-world threat — a surprising, urgent, or high-stakes piece of media (an urgent voice request, a shocking video) — the single highest-value technique needs no tool: pause and verify through an independent channel. The urgency and shock are exactly the manipulation, and stepping outside the media to confirm (call back on a known number, check independent sources for the event) defeats most synthetic-media harm before any forensics. This habit — distrust the urgent surprising media, verify independently — is the practical core of synthetic-media defense, and it embodies the whole frame: anchor on provenance and corroboration, not on the media itself.
This starter kit lets you begin verifying now — reverse search, provenance, corroboration, and the independent-channel habit carry most everyday cases — while the following modules build the deep forensic technique (video/audio deepfake artifacts, text detection's limits, image forensics and GAN artifacts, metadata analysis) that the harder cases require, all assembled in Module 6 into the full verification workflow.
Key Insight: The verification starter kit: reverse-search first (an image's history often resolves authenticity cheaply), check provenance (untraceable or inconsistent origin is a strong signal), seek corroboration (real significant events are multiply-witnessed), inspect artifacts and metadata as signals, run detectors weighted by honest reliability, and weigh all signals to a calibrated conclusion — most confident when independent signals (especially provenance and corroboration) converge. The single highest-value habit needs no tool: for urgent, surprising, high-stakes media, pause and verify through an independent channel — the urgency is the manipulation, and stepping outside the media defeats most synthetic-media harm.
You now have the landscape and the frame: the types of synthetic media and their type-specific detection, the arms race that makes detection imperfect and losing, the honest state of detection tools (useful-but-unreliable signals, not verdicts, with text detection least reliable), the detection mindset (calibrated skepticism, signals not verdicts, anchoring on the durable provenance and corroboration signals), the concrete stakes that motivate the discipline, and a starter toolkit and triage approach to begin verifying now. Above all, you hold the governing truth: detection techniques are necessary but not sufficient, and the reliable protection is multi-signal, provenance-anchored verification, not faith in any detector.
The next module, Deepfake Detection: Video & Audio, goes deep on the forensic techniques for the highest-stakes synthetic media — the visual, temporal, and acoustic artifacts that betray manipulated video and cloned audio, the tools and methods for finding them, and the honest reliability of each as generation improves. Everything builds on this module's frame: each artifact is a signal weighted by its worth and durability, combined with others and anchored on provenance, toward a calibrated conclusion — the disciplined verification that works even against media designed to fool the eye and ear, because it never relies on the eye and ear alone.