Truth Online
How to still sort out what is true on the web — a method collection for skeptical readers.
A TikTok video goes viral — a politician says something outrageous. Millions share it. But it never happened. The clip was generated by AI in under a minute. In the article "Learning with AI", you learned that AI can hallucinate — produce wrong information unintentionally. This article goes one step further: what happens when people deliberately use AI to deceive? You will learn what deepfakes are, how to quickly verify sources, and how to navigate a world where more and more content is AI-generated.
Deepfakes — When AI Lies
Definition
Deepfake
Analogy
Definition
Visible artifacts: wrong finger count, unnatural blinking, blurry facial edges, inconsistent lighting. Detectable by attentive observers.
Professional quality: correct anatomy, natural facial expressions, consistent lighting, lip-synced speech. Virtually indistinguishable from real footage by the human eye.
In 2021, a TikTok account called @deeptomcruise posted videos of "Tom Cruise" playing golf and performing magic tricks. The videos were so convincing that many of the millions of viewers believed they were real. Only after investigation was it revealed that a professional deepfake artist had created them with a Tom Cruise impersonator over months of work. The lesson: if even a globally recognizable face can be faked convincingly, any face can be faked.
Caution
Misconception: "I can spot deepfakes — there are always glitches"
That may have been true in 2019 — weird fingers, strange blinking, blurry edges. But not anymore. Modern deepfakes have largely eliminated these artifacts. Relying on visual detection gives you false confidence. The right approach: verify the source, not the pixels.
Deepfakes can be based on Generative Adversarial Networks (GANs), among other techniques: two neural networks train each other — a generator creates fakes, a discriminator tries to detect them. Through this interplay, the fakes keep improving. In video deepfakes, a face is mapped onto another face frame by frame. In voice cloning, AI analyzes speech patterns, pitch, and rhythm from a short audio sample and then synthesizes any statement in that voice.
SIFT — Your Tool Against Fakes
Definition
SIFT Method
Analogy
Definition
- 1. StopDo not react or share immediately. Disinformation exploits emotional reactions.
- 2. InvestigateCheck the source: Who published this? Reputable outlet or anonymous account?
- 3. FindLook for other coverage: Do independent, reputable sources report the same thing?
- 4. TraceFollow the claim to its primary source: Which study? By whom? Does it exist?
Note
Walkthrough: Fake Health Post on TikTok
A viral post claims: "Scientist confirms: [shocking health claim]." 50,000 shares. SIFT check: **Stop** — Do not share. **Investigate** — The poster has 200 followers, no credentials, account created 3 days ago. **Find** — No reputable health outlet, medical journal, or news agency reports this. **Trace** — The cited "study" does not exist: no DOI (the unique ID number of scientific articles), no author, no journal. Result: fake. Time spent: about 90 seconds.
Caution
Misconception: "Source verification takes too long"
The SIFT method is designed for speed: steps 1 and 2 take about a minute. You do not need to check every post — only those that trigger strong emotions, make extraordinary claims, or ask you to act (share, click, donate). Disinformation thrives on urgency — pausing for just 10 seconds breaks its power.
Interactive: Check a Source Yourself
You have learned the SIFT method. Now apply it: work through the checklist and evaluate any online source. At the end you will receive a credibility score with a traffic-light rating — so you can see at a glance how trustworthy the source is.
Credibility Check
Check an online source for trustworthiness. Answer the following questions with Yes or No.
Synthetic Media — The Big Picture
Synthetic media is the umbrella term for all content — text, images, video, audio — that is wholly or partially generated by AI. This includes deepfakes but also useful applications like AI translations, automatic subtitles, image enhancement, and AI art. You often read that up to 90 percent of all online content could be synthetic by 2026. The figure appeared in Europol's 2022 report "Facing Reality," but Europol removed it in the revised edition of January 2024 because it came from an inaccurate source — a good example of why you trace numbers back to the primary source (Trace). So far, only slices have been measured: in early 2026, about half of newly published English-language online articles were mostly AI-written, according to a sample by the marketing firm Graphite using three AI detectors. The core problem is not synthetic media itself but the lack of labeling.
Voice Cloning
3 -15 sec
That is how little audio material AI needs to convincingly clone a voice
Synthetic Content
~50 %
Share of newly published English-language online articles written mostly by AI, early 2026 (sample by the marketing firm Graphite, measured with AI detectors)
SIFT Quick Check
60 sec
That is how long the first two SIFT steps (Stop + Investigate) take
Note
Quick Check: Is This Real?
- Reverse image search: (Google Lens) — Has this image appeared before in a different context?
- Check metadata: AI-generated images often lack EXIF data (camera model, capture time) from a real camera. Missing metadata proves nothing, though: some platforms strip it on upload, even from real photos.
- Look for C2PA labels: (digital proof of origin, more in the Deep Dive below) — Some platforms already show "AI-generated" badges. You can also have images, videos and audio checked for Google's invisible watermark in Gemini or the SynthID Detector; that way you spot content from Google's AI and from partners such as OpenAI. But no watermark does not mean “real”.
- Apply SIFT: Does a credible source confirm the content?
If all four checks fail to confirm authenticity: treat the content as unverified.
Caution
Misconception: "AI detectors solve the problem"
Only partially true. Detectors like GPTZero (for text) or Hive (for images) are useful but have significant error rates and are in a constant arms race with generators. Generators are usually one step ahead because new generators appear before detectors have been trained on them. Detectors are a helpful addition, but not a definitive answer. Your SIFT skills remain essential.
Since 2 August 2026, Article 50 of the EU AI Act has required that deepfakes be disclosed as artificially generated; providers must also mark AI-generated images, videos, audio and text in a machine-readable way (systems already on the market before then have until 2 December 2026) — a first regulatory step against invisible synthetic media. In parallel, the C2PA consortium (Coalition for Content Provenance and Authenticity) has developed an open technical standard: a digital provenance certificate embedded in media files that shows when, where, and with what tool content was created or modified. Some platforms already support C2PA.
Note
Key Takeaways
- Deepfakes have become too good: Video, audio, and image deepfakes can no longer be reliably detected by human senses. "I can see it's fake" is a dangerous assumption.
- SIFT is your universal tool: Stop → Investigate → Find → Trace. These four steps work for any type of online content — deepfake videos, AI-generated texts, or manipulated images.
- Synthetic media are not inherently bad: AI translations, accessibility features, and art are useful. But missing labeling and malicious use make synthetic media dangerous. Your verification skills are the last line of defense.
This is the final article in this learning path. The SIFT method goes far beyond deepfakes — it is your tool for evaluating any information, in any medium, for the rest of your life.
Check
What This Is About
What is a deepfake?
Select one answer
1. What is a deepfake?
2. A friend sends you a shocking video of a celebrity making a controversial statement. Applying the SIFT method, what should you do FIRST?
3. You see a news post: "New study proves: [shocking health claim]." 50,000 shares. You investigate: anonymous account, no major outlet reports it, the "study" has no author and no DOI. What is your conclusion?
4. A classmate argues: "We don't need to learn source verification — AI detectors like GPTZero will identify all fakes for us." Why is this argument flawed?
Check
Learning Goals
Can you explain this in your own words? Tick what you've got.
0 of 3 done