Good Prompts Are Not Enough:
How to Verify and Trace AI Answers
Many people use AI as a search engine, tutor, or advisor. But AI output often mixes facts, plausible guesses, and fabricated details. What matters is not only how to ask, but how to verify what comes back. This chapter connects source tracing, lateral reading, reverse search, and citation habits into one AI-era workflow.
Context Manipulation: The Most Overlooked Technique
First Draft (a Harvard research project), after analyzing thousands of disinformation cases, categorized visual content misinformation into seven types. The most common isn't "completely fabricated content" but "False Context" — using real images or videos but pairing them with false time, location, or event descriptions.
Why is context manipulation so effective? Because:
- Real images/videos pass the first "looks real" visual check
- Most people's verification habit is "check if the image was modified," not "check the image's original source"
- Reverse image search remains a skill mastered by few; most ordinary users have never used it
| Type | Description | Frequency |
|---|---|---|
| False Context | Real content + false context | ★★★★★ |
| Imposter Content | Impersonating credible sources | ★★★★☆ |
| Manipulated Content | Real images modified | ★★★☆☆ |
| Fabricated Content | Entirely fabricated content | ★★☆☆☆ |
Reverse Image Search: Three Free Tools
Reverse image search is the most effective image tracing tool. The principle: use the image itself as a search key to find all locations where the image has appeared online, thereby confirming when and where it first appeared.
| Tool | Strength | Weakness | Best For |
|---|---|---|---|
| Google Images | Largest database, global coverage | Lower accuracy for Asian/Middle Eastern images | Western media news images |
| TinEye | Precisely tracks first appearance date | Relatively smaller database | Confirming an image's "age" |
| Yandex Images | Particularly high accuracy for Asian/Eastern European images | Interface in Russian | Taiwan, China, Japan/Korea, Middle East images |
Step-by-Step (Using Google)
- In Chrome browser, right-click any image → select "Search image with Google," or go to images.google.com and click the camera icon
- Upload the image file or paste the image URL
- Check the "Visually similar images" and "Other sizes of this image" sections in the results
- Compare the earliest reporting date in the results to confirm if the image first appeared before the claimed event date
Video Tracing: How to Use InVID
Videos can't be directly reverse-searched, but the InVID/WeVerify browser extension greatly simplifies this process:
- Install Chrome/Firefox extension (search "InVID WeVerify")
- Use the "Video analysis" function on video links from YouTube, Facebook, Twitter, etc. — automatically captures 16 key frames and runs reverse image search
- Check search results for each key frame to confirm if the video appeared in other contexts before
Geolocation Verification: Using SunCalc and Google Earth
When a video claims to be filmed in a certain location but you suspect the location is wrong, use:
- Google Street View: Screenshot buildings, street signs, storefronts from the video and search for matches in Street View. Bellingcat investigators typically locate video filming sites within 2-3 minutes.
- SunCalc.org: Enter the claimed location and date on SunCalc to calculate the sun's azimuth and elevation at any time of day. Compare with shadow directions in the video to verify whether the "date + location" combination is consistent.
- Google Earth Timelapse: View historical satellite imagery of a location to confirm timelines of construction and landscape changes, ruling out the possibility that "this was filmed years ago."
EXIF Metadata: The Invisible Diary of Photos
Every digital photo contains hidden EXIF (Exchangeable Image File Format) metadata: capture time (accurate to the second), camera model, lens information, and in cases where it hasn't been stripped, GPS coordinates.
Important caveats:
- Social media platforms like Facebook and Instagram usually strip GPS data on upload, but don't necessarily strip capture time
- EXIF data can be modified, making it "supporting evidence" rather than "final proof"
- Use Jeffrey's EXIF Viewer (jeffreyfriendship.com) or exiftool to read EXIF data
Lateral Reading: Stanford University's Breakthrough Research
Stanford University's History Education Group (SHEG) conducted an important experiment: they had three groups — high school students, college students, and professional fact-checkers — evaluate the same suspicious websites. The results were surprising: professional fact-checkers had 3× higher accuracy than the other two groups.
What exactly did they do differently? The researchers recorded the professional fact-checkers' screen actions and found one critical difference:
- Ordinary people (vertical reading): Carefully read the suspicious website's articles, trying to judge credibility from the site's own "About Us" page
- Professional fact-checkers (lateral reading): After seeing the suspicious site's headline, immediately opened new tabs in the search engine to search "what is this website's background," finding others' evaluations of this source
When you encounter an unfamiliar source, don't read the article first. Instead:
- Open a new tab, search "[site name] + credibility / bias / fake news"
- Search "[site name] + site:en.wikipedia.org" to see if Wikipedia describes this source
- Check Media Bias / Fact Check (mediabiasfactcheck.com) ratings
- Spend just 60-90 seconds for a basic credibility assessment of the source
Taiwan Fact-Checking Resource Map
For Taiwan users, these are the most practical Chinese-language verification resources:
台灣最大事實查核機構，LINE 機器人可即時查詢。
tfcctw.org
在 LINE 中加入好友，直接傳送可疑訊息即可查詢。
mygopen.com
社群協作查核 LINE 群組流傳的假訊息。
cofacts.tw
專注台灣政治相關假訊息的獨立查核媒體。
rumtoast.com
Slide Deck
Case Studies
Within hours of the April 3, 2024 Hualien magnitude 7.2 earthquake, multiple dramatic photos of collapsed buildings and road damage spread massively on LINE groups and Facebook, labeled as "Hualien live disaster footage." During the tense period when people urgently sought information about family and friends' safety, these images spread extremely rapidly.
After emergency verification by Taiwan FactCheck Center (TFC) researchers, they found: one widely circulated "collapsed building" photo was actually taken during the 2018 Hualien earthquake; another "broken road" photo came from the 2021 Taroko Express train accident; and one "injured rescue scene" photo had first appeared in coverage of the 2016 Kumamoto earthquake in Japan.
TFC researchers used Google Reverse Image Search + TinEye "earliest appearance" function, spending no more than 3 minutes per image to confirm each one's true original source. TinEye's "Sort by Oldest" function directly showed the earliest indexed date for each image online, all predating April 3, 2024.
Emergency disaster periods are peak times for disinformation. Reverse image search is the most effective first-line verification tool — no more than 3 minutes per image to determine if it's a genuinely "live" photo.
The first 48 hours after Russia's February 24, 2022 invasion of Ukraine was one of the highest-density periods of disinformation on social media. Bellingcat later documented over 300 incorrectly labeled visual items. These included: air combat footage claimed to show "Ukrainian Air Force shooting down six Russian aircraft" — actually from a 2013 flight simulator game — plus multiple "Russian armored forces invading" clips that came from the 2014-2016 Donbas conflict and Syrian civil war.
Bellingcat investigators used a systematic approach: ① Screenshot key frames for reverse image search ② Geolocate buildings (Google Street View, Yandex Maps) ③ Use shadow direction and vegetation state to estimate season and time ④ Compare vehicle identification codes to confirm equipment era. The key "Ghost of Kyiv" flight simulator video was identified through key frame search finding the original YouTube upload, confirming its game nature.
When wars and major conflicts break out, disinformation creators know news consumption spikes — making it the optimal moment to release fake footage. Reverse search + geolocation matching + timeline analysis is the golden combination for distinguishing real battlefield documentation from disinformation.
- For "disaster," "battlefield," "accident scene" images, do reverse search before sharing
- Use TinEye "Sort by Oldest" to confirm an image's "age"
- When unsure, wait for Taiwan FactCheck Center or Reuters Fact Check reports