Life

Jianfa Tsai’s Input

Implement AI app that check and count the number of uploaded photos to each social media account to cross reference to multiple other data points to create a risk scale to determine if the profile is a romance scammer, investment scammer or potential criminal. E.g. most crime syndicates create social media accounts (TikTok and others) of beautiful and sexy women. The pivotal point, or red flag is the total count of the number of “selfie” photos uploaded to each account is highly similar or of a close range, e.g. between 8 to 11. Will it be too much of a coincidence that the barrage of users requesting to following me where most of them have profile photos totalling to between 8 to 11 for each account? E.g. if you are a scammer living in a third world country working as a bot to create fake social media accounts, you are following a template and out of laziness, or habits, you are likely to upload photos to each account between a highly restricted range. Similar to the number patterns in the company accounts (checking for embezzlement) discovered by the protagonist in The Accountant USA movie.

Identified Problems

• The core idea is useful, but the proposed rule about a narrow selfie-count range such as 8 to 11 is, by itself, too weak to support a reliable scam or criminal inference, because government scam guidance emphasises broader signals such as low account history, few connections, limited engagement, inconsistent identity details, pressure tactics, and requests for money rather than a single numeric photo pattern (Australian Competition and Consumer Commission [ACCC], 2026; U.S. Federal Trade Commission [FTC], 2019).[1][2]
• The prompt risks overgeneralisation and stereotyping by linking suspicious accounts to “beautiful and sexy women” and to people in “third world” countries, yet trusted consumer-protection sources describe scammers as using fake profiles, stolen images, urgency, secrecy, off-platform migration, and investment grooming without reducing risk assessment to gendered appearance or nationality assumptions (ACCC, 2026; FTC, 2019).[2][1]
• The proposed output categories, especially “potential criminal,” are too broad for a responsible AI system, because public scam guidance supports flagging behavioural risk indicators but does not justify declaring criminal status from profile features alone (ACCC, 2026).[1][2]
• The hypothesis may still be worth testing as one weak feature among many, because scamwatch guidance specifically advises checking posting history, profile age, number of connections, and activity including posts and photos, which means photo counts can contribute to an evidence bundle if handled as a probabilistic signal rather than proof (ACCC, 2026).[2][1]
• A better design for your Melbourne-based research and blogging workflow would be an explainable risk-scoring prototype for suspicious social-media accounts, framed as digital-libraries style metadata analysis and pattern detection, rather than an app that labels people as scammers or criminals outright (ACCC, 2026; FTC, 2019).[1][2]
  

Abstract

Your idea is directionally strong because it seeks to convert visible social-media profile traces into a structured risk model for romance scams, investment scams, and related deception, which aligns with public warnings that fake profiles often have sparse histories, limited engagement, weak identity consistency, and movement toward money requests or fake investment offers (ACCC, 2026; FTC, 2019).[2][1]

However, the specific claim that a repeated selfie-count band such as 8 to 11 strongly indicates organised scam templates is not established by the sources reviewed here, so it should be treated as a research hypothesis to test empirically, not as a pivotal rule or direct evidence of guilt (ACCC, 2026; FTC, 2019).[1][2]

The safest and most defensible implementation is an explainable AI-assisted screening system that combines photo-volume patterns with account age, posting cadence, engagement anomalies, profile completeness, reverse-image signals, messaging behaviour, off-platform migration, and money or crypto solicitation signals, then outputs a transparent risk band such as low, medium, or high suspicion rather than asserting that a person is a scammer or criminal (ACCC, 2026; FTC, 2019).[2][1]

ELI5

Think of a social-media profile like a library book record: one clue, like “this book has 9 pictures,” is not enough to say the book is fake, but many clues together can help you be careful, such as the account being very new, having hardly any friends, using strange or overly polished photos, refusing video calls, pushing you onto WhatsApp, and asking for money or crypto very quickly, so your app should act like a careful librarian that says “this looks risky, here are the reasons,” instead of saying “this person is definitely a criminal” (ACCC, 2026; FTC, 2019).[1][2]

App design

• Build it as a risk-assessment tool, not a verdict engine, because trusted guidance supports identifying warning signs and verifying identity, not declaring criminal status from profile metadata alone (ACCC, 2026; FTC, 2019).[2][1]
• Use output classes such as:
• Low concern.
• Medium concern.
• High concern.
• Romance-scam pattern likely.
• Investment-grooming pattern likely.
• Manual review required.
  (These are safer than “criminal” labels because they communicate uncertainty and support human judgment) (ACCC, 2026).[1][2]
• Keep the selfie-count feature as one small weighted variable, because Scamwatch advises checking photo and posting activity but does not identify a specific “8 to 11 selfies” threshold as a validated scam marker (ACCC, 2026).[2][1]
  

Better signals

Signal	Why it is stronger	Use in score
Account age and history	New or thin-history accounts are a recognised fake-profile warning sign (ACCC, 2026) [1][2]	High weight
Number of followers, friends, and engagement	Few followers, comments, likes, or shares are listed red flags (ACCC, 2026) [1][2]	High weight
Profile completeness and consistency	Mismatches between profile and story are warning signs (ACCC, 2026) [1]	High weight
Reverse-image reuse	FTC recommends reverse-image checks for profile-photo reuse under different names (FTC, 2019) [2]	High weight
Refusal to meet or video chat	Scamwatch lists persistent excuses for avoiding in-person or video contact (ACCC, 2026) [1]	High weight
Fast emotional escalation	“Love bombing” and rapid intimacy are recognised warning signs (ACCC, 2026) [1]	High weight
Push to WhatsApp, LINE, WeChat	Moving quickly off-platform is a warning sign (ACCC, 2026) [1]	High weight
Investment or crypto pitch	Scamwatch links romance grooming to fake investment schemes, often crypto-based (ACCC, 2026) [1]	Very high weight
Requests for money or transfers	This is one of the clearest public scam indicators (ACCC, 2026; FTC, 2019) [2][1][2]	Very high weight
Selfie count similarity across accounts	Plausible as a templating clue, but not validated by the sources reviewed here (ACCC, 2026) [2]	Low weight, experimental only

Balanced reasoning

• Supporting view:
• Your instinct about templated behaviour is reasonable, because fake-profile operations often reuse repeated structures, shallow histories, and standardised visual patterns, so cluster analysis across photo counts, posting intervals, bio templates, and engagement anomalies could uncover suspicious campaigns even when any one signal is weak (ACCC, 2026).[1][2]
• From a GLAM and librarianship perspective, this resembles metadata forensics: the value lies in combining many small catalogue clues into an evidence pattern rather than trusting one field in isolation (ACCC, 2026).[2][1]
  
• Counter-argument:
• A photo-count rule can easily create false positives, because real users also post in similar ranges, especially new, casual, or curated accounts, and public guidance does not support using image count alone as a decisive signal (ACCC, 2026; FTC, 2019).[1][2]
• If the model is trained on biased assumptions about gender, attractiveness, language, or geography, it may unfairly target legitimate users and become unsafe or defamatory, which is especially risky if you want the system to suggest criminality (ACCC, 2026).[2][1]
  

Safer scoring model

• Recommended weights:
• Behavioural solicitation, money requests, crypto promotion, secrecy, off-platform movement: 45%.
• Account credibility, age, completeness, follower and engagement patterns: 30%.
• Image forensics, reverse-image reuse, AI-photo artifacts, face repetition: 20%.
• Experimental templating features such as selfie-count bands: 5%.
  This weighting better matches public scam guidance, which focuses on conduct and profile credibility over one narrow visual count rule (ACCC, 2026; FTC, 2019).[1][2]
  
• Recommended output:
• Risk score from 0 to 100.
• Top 5 reasons the score is high.
• Confidence score.
• “Needs human review” flag.
• Audit log showing which features contributed.
  This explainable structure is more defensible for personal safety, student research, and blogging than a black-box accusation model (ACCC, 2026).[2][1]
  

Action steps

• Define the tool’s purpose narrowly: “screen for likely scam patterns” rather than “detect criminals,” because that aligns better with consumer-protection guidance and reduces legal and ethical risk (ACCC, 2026).[1][2]
• Collect a balanced validation set with known scam reports and known legitimate accounts, then test whether selfie-count clustering adds predictive value beyond account age, engagement, and reverse-image reuse; otherwise remove it (FTC, 2019; ACCC, 2026).[2][1]
• Add a human-readable explanation layer so every risk decision can be blogged, reviewed, and critiqued transparently in your research workflow, which suits academic and GLAM standards of evidence accountability (ACCC, 2026).[1][2]
• Build in fairness checks across gender presentation, language patterns, and geography so the model does not reproduce stereotypes not supported by evidence (ACCC, 2026).[2][1]
• For your personal life, treat incoming follow requests with a checklist: reverse-image search, check account age, look at follower interaction quality, test willingness for video verification, and never engage on money or investment topics with online-only contacts (FTC, 2019; ACCC, 2026).[1][2]
• For your academic life at Swinburne, frame this as an explainable digital-trust scoring study using metadata and behavioural indicators, with a strong ethics section on bias, defamation risk, and false positives, because that will make it more publishable and defensible (ACCC, 2026; FTC, 2019).[2][1]
• For your work and blogging life, position the article around “how to evaluate suspicious social-media profiles responsibly” rather than “how to spot scammers from sexy selfies,” because the former is more evidence-based, academically credible, and suitable for library and public-information audiences (ACCC, 2026).[1][2]
  

Thought-provoking question

• If your app had to explain every flagged profile to a librarian, a judge, and the flagged user using only verifiable evidence, which signals would still feel fair and defensible after the selfie-count rule was removed (ACCC, 2026; FTC, 2019) ?[2][1]
  

Date

Monday, May 25, 2026, 2:41 PM AEST

Authors

Jianfa Tsai (https://orcid.org/0009-0006-1809-1686) in collaboration with Perplexity AI Pro. Jianfa Tsai resides at 60 Dowling Road, Oakleigh South, VIC 3167, Australia.

References

Australian Competition and Consumer Commission. (2026, January 22). Relationship scams. Scamwatch. https://www.scamwatch.gov.au/types-of-scams/relationship-scams[1]

Australian Competition and Consumer Commission. (2026, March 25). Social media scams. Scamwatch. https://www.scamwatch.gov.au/types-of-scams/social-media-scams[3][2]

Federal Trade Commission. (2019, June 4). What to know about romance scams. https://consumer.ftc.gov/articles/what-know-about-romance-scams[2]

Sources

[1] Relationship scams | Scamwatch https://www.scamwatch.gov.au/types-of-scams/relationship-scams

[2] What To Know About Romance Scams – FTC Consumer Advice https://consumer.ftc.gov/articles/what-know-about-romance-scams

[3] Social media scams | Scamwatch https://www.scamwatch.gov.au/types-of-scams/social-media-scams

[4] Do You Love Me? Psychological Characteristics of Romance Scam … https://pmc.ncbi.nlm.nih.gov/articles/PMC5806049/

[5] Real relationships develop at their own unique pace. But romance … https://www.facebook.com/ConsumerAffairsVictoria/posts/real-relationships-develop-at-their-own-unique-pace-but-romance-scams-tend-to-fi/1324628113033235/

[6] How to spot a romance scam before it’s too late | Money magazine https://www.moneymag.com.au/romance-scams-how-to-detect-a-fraud-before-its-too-late

[7] [PDF] FAKE SOCIAL MEDIA PROFILE DETECTION USING ML … https://papers.ssrn.com/sol3/Delivery.cfm/SSRN_ID5064663_code3635775.pdf?abstractid=5064663&mirid=1

[8] Protect your heart and bank account from romance scams https://www.ausbanking.org.au/protect-your-heart-and-bank-account-from-romance-scams/

[9] Online dating scams: how to spot the red flags – CommBank https://www.commbank.com.au/brighter/financial-education/online-dating-scams.html

[10] Romance scams can occur through dating apps, social media and … https://www.facebook.com/FairTradingNSW/posts/romance-scams-can-occur-through-dating-apps-social-media-and-messaging-platforms/1220759933560549/

[11] Love at first click? ❤️ ➡️ Think twice before you fall … – Instagram https://www.instagram.com/p/DUl-ivwjHtt/

[12] How to detect fake profiles on social media? – Facebook https://www.facebook.com/groups/198491731518421/posts/1085692919464960/

[13] Share your heart and not your identity this Valentine’s Day With … https://www.instagram.com/p/DUpMiRfFWJm/

[14] Spotting fake profile pictures and social media scams – YouTube https://www.youtube.com/watch?v=VvH6Ocqt9tE

[15] Instagram fake profile detection using an ensemble learning method https://pmc.ncbi.nlm.nih.gov/articles/PMC12279990/