李怡志：如何生成有效的圖騙

1. In one sentence

Whether a generated image fools you has never turned on how real it looks, but on whether it lands on your position, your emotions, the knowledge gap you are unaware of, and your trust in the other people who liked and shared it — and 李怡志 took those four apart and demonstrated each.

2. The session

Teaching generated imagery since 2003

He introduces himself as an assistant professor at NCCU who has been teaching generated imagery there since 2003, accumulating about eight credits' worth of courses. He opened with a student assignment: one student made "a Big Mac in 1942", a fake historical photograph of Big Macs existing during the Second World War, and used it to fool his own mother, who believed it. In the same batch, other students made a fake photograph of a Birmingham strike, and fabricated a story that "Taiwan has grown coffee since antiquity, so it should have been called Aroma rather than Formosa" to fool a Korean language exchange partner. He said these images have plenty of flaws, and yet used on mothers, older relatives and foreign friends they were believed — which is what genuinely interests him and what he wanted to talk about: why do people still believe images with obvious flaws.

He pointed out the pun in "圖騙" in his title: the second character can be the 片 of 圖片 (image) or the 片 of 騙人 (deception).

What a fake image is actually trying to do to you

He splits the purposes of a fake image into four:

• Implanting a false memory: machine learning needs data fed to it, and a fake image feeds bad data into an audience's brain. His examples include: a composite photograph placing a man accused of sexual assault during Me Too together with another woman (no independent reporting was found to corroborate the identity of the person in that composite, so it is marked unverified); the photograph of Martin Luther King Jr. apparently giving the finger, where the original frame is in fact him making a peace sign (two fingers) on learning the Senate had passed the Civil Rights Act, later altered to leave only one finger (subsequently verified, see section seven); the 2023 fake photograph of an explosion at the Pentagon that briefly moved US markets (subsequently verified, see section seven); and two AI-generated images published respectively by the Chinese embassy in France and by 陳衛華, head of *China Daily*'s EU bureau, denouncing Israeli attacks on Palestine (subsequently verified, see section seven).
• Manipulating your attitude: he cited the fake image of Madonna in an intimate relationship with the Pope, aimed at changing the audience's view of the Pope; and a commentary co-written by NTU's 蔡蕙如 and NTNU's 林玉鵬 discussing depicting the Israeli military in Ghibli style, drawing violence as adorable — essentially decontextualising violence into a warm image (subsequently verified, see section seven).
• Amplifying your emotions: positive or negative both count. He cited an image a friend shared on social media, obviously AI-generated yet moving enough to make you want to cry, and how when he pointed out to them that it was AI-generated, they asked back "why would these people lie to me?"
• Lowering your judgement: which he considers the most frightening, because eventually you simply give up verifying. He cited the AI-generated "Trump as king" image Trump published himself, and the AI composite of himself dressed as the Pope that Trump posted after the Pope's death (both subsequently verified, see section seven) — images this hard to tell apart and this frequent make people simply abandon the step of asking "is this real?"

Why you get fooled: from heuristic to systematic processing, and the source monitoring framework

He borrowed the dual-process framework from communication theory: scrolling on a phone, most people use heuristic processing, skimming past and sharing on a feeling, since nobody can seriously study every image; the alternative is systematic processing, which takes time to assess truth carefully. He said social media is by design an encouragement to heuristic processing.

He then introduced a framework called the source monitoring framework, tracing its lineage to a 1987 study he referred to as Jackson, and on to a 1993 development (what external material points to is Johnson, Hashtroudi and Lindsay's 1993 *Source Monitoring*; no author or exact title could be found for a 1987 paper, and "Jackson" appears to be a slip for "Johnson", with the year unverified — see section seven). The framework holds that in judging whether a piece of information is true, we are influenced not only by the visual (does it look real) but simultaneously by information, emotion and trust.

He broke those four into operational detail:

• Visual characteristics: including copy-move, splicing, retouching, cropping and blurring. His examples were the photograph of Trump's first inauguration cropped to remove the angle showing an empty area in front, and the US National Archives blurring out placards protesting Trump.
• Semantic characteristics: the content has to be consistent with common sense, with period, place and cultural detail all matching. He reminds students that if they are generating a scene set between 2020 and 2023, the people must be wearing masks — without them it looks fake.
• Emotional characteristics: can the image or video make someone angry, happy or sad? The closer to a situation the audience knows, the easier it is to stir emotion and the easier it is to be believed.
• Trust strategies: mixing real and generated frames (a real background with an AI-generated person in the foreground), faking posting times and watermarks, mimicking the traces of a social platform repost (a verified account, a news screenshot), and the credibility manufactured by like and share counts.

On top of these he added two crucial psychological mechanisms: the knowledge gap effect — we receive far too much information daily to verify all of it, and the part we do not know and cannot be bothered to check is the easiest breach; and confirmation bias — as long as a piece of information fits our existing position we tend to believe it is true, regardless of whether it actually is.

He used the widely circulated letter of a reply to the education minister, "corrected in red pen until it bled", as an example: the corrections underneath were made by an internet user themselves and were not an official Harvard response, but because the plot fits the public's fixed impression that "the education minister is unread" so well, many people never even thought to check before sharing — he nearly reshared it himself at first.

He also gave several Taiwanese social media cases showing how the whole mechanism compounds: one being the circulated claim that "a Chinese athlete's competition kit was printed Chinese Taipei" (one of the claims circulating during the World Masters Games; he mentioned that people who reshared it were subsequently fined under the Social Order Maintenance Act); another taking the genuine 剴剴 childcare abuse case and layering on the claim that "the accused carer is a mainland Chinese spouse", paired with the rhetorical device of "do you know her? no? so it's been covered up", tying knowledge gap, confirmation bias and emotion together at once, so that people would believe an entire narrative without so much as looking at an image; and a third being the circulated claim that a mainland spouse, "雅雅", returned to China to remarry and brought mainland relatives to Taiwan as dependants. He also mentioned a friend commenting under his own Facebook post citing "周偉航 says it's true", adding "he said it on a political talk show, so it's more credible" — that kind of compounding misattribution being another common technique (he also mentioned a reshared RTHK Facebook post confirmed as disinformation, with the link on the slide; it is listed in section seven).

Demonstrating how to make a fake image more real

He gave a live "reverse tutorial": first the quality prompts (realistic, Canon EOS 5D, natural, plus a little imperfection), then the semantic setting (an apartment block in Ukraine, a person wearing a red cross armband), then layering on emotional elements (fear and sadness on a child's face). Having produced an image of "a Ukrainian child receiving battlefield medical aid", he then layered on the look of a CNN-style Facebook post — but told everyone to look closely: there is an extra dot beside that "CNN", making it an impersonation account rather than the real CNN. He also demonstrated another layer of trust strategy: if the image shows likes from 李慕約 and 李怡志, ordinary people believe it more readily, because we unconsciously trust the opinion leaders we follow, even though those opinion leaders are themselves easily fooled.

How not to get fooled

His approach is practical, and does not ask everyone to become a fact-checker:

1. Forget whether it looks real — anyone can produce a very realistic image now, and that is no longer the point.
2. Check whether you have tripped any of four switches: does the image fit your existing values and position, do you have an emotional reaction, do you actually know the truth, and are a lot of people liking and sharing it (he cautions that most successful disinformation has very high engagement, and engagement itself is not evidence of truth).
3. The moment you want to share, do one simple check first — right-click the image and run a Google image search, and very often a fact-checking organisation's explanation pops up straight away. His live example was an image of a meteor streaking past treetops billed as "a National Geographic award-winning photograph of the year", which one search revealed had already been flagged as fake.

3. Figures and cases

• He has taught generated imagery at NCCU since 2003, accumulating over twenty years of teaching (speaker's own account).
• The 2023 fake Pentagon explosion image briefly moved the S&P 500 down by about 0.26% (subsequently verified, see section seven).
• The Martin Luther King Jr. "giving the finger" photograph is in fact a two-fingered peace sign, and the doctored version was dug up and widely reshared in 2017 (subsequently verified, see section seven).
• The Harvard "red pen correction of the education minister's letter" incident occurred in May 2025, and the letter was corrected by an individual internet user rather than being an official Harvard response (subsequently verified, see section seven).
• Trump published an AI composite in "King Trump" style in 2025, and after the Pope's death an AI composite of himself dressed as the Pope; both drew controversy (subsequently verified, see section seven).

4. Lines worth keeping

• "That character can be the 片 of image, or the 片 of deception."
• "Everyone's going to stop verifying altogether — that's exactly what Trump is up to."
• "The more you want to share something, the more likely there's a problem with it, and the more you need to check."
• "You're all experts in generative AI. The least we can hope for is that we all try not to get fooled."

5. Tools and terms mentioned

• Source monitoring framework: the psychological framework for judging where a memory came from, separating the visual, informational, emotional and trust dimensions that influence judgement; the usual reference is Johnson, Hashtroudi and Lindsay's 1993 *Source Monitoring*.
• Heuristic / systematic processing: the dual-process model of information handling common in communication and psychology, used to explain why people mostly judge quickly with the former while scrolling social media.
• Knowledge gap effect: the more information there is, the easier the gap between what an individual knows and does not know is to exploit.
• Confirmation bias: anything fitting an existing position gets believed, true or not.
• Copy-move / splicing / retouching / cropping / blurring: the five traditional image manipulation techniques he groups together.
• Google reverse image search: what he demonstrates as a first check on whether a suspicious image has already been flagged as fake.

6. Wider observations

The session is not trying to teach anyone to spot the flaws of AI generation, because he knows perfectly well those flaws can no longer be exhaustively caught. What matters is changing the question from "is this image AI-generated?" to "why do I want to believe this image?" The four switches he identifies — position, emotion, knowledge gap, herd behaviour — apply to almost every moment anyone has wanted to hit share while scrolling on a phone, and laying out the psychological mechanism rather than handing over a pile of detection tools is both more memorable and more durably useful than teaching technical identification alone.

7. Sources