- All four assistants caught all five scam-control messages.
- ChatGPT, Gemini, and Claude correctly cleared all five legitimate controls; Copilot marked two as uncertain.
- Scam-risk scores ranged from 78 to 95, while legitimate controls ranged from 5 to 40, with no overlap in this controlled test.
- AI was most useful for spotting red flags and suggesting checks, not for proving that a seller is genuine.
Imagine you are about to transfer ₦300,000 to an online seller.
The price looks reasonable, but something feels off. The seller wants a bank transfer. Perhaps they are rushing you, the account name doesn't match the business, or they say you cannot inspect the item before paying.
But before pressing send, would pasting that message you receive from said seller into an AI assistant help you determine if you’re about to get scammed?
That matters because AI is already entering the same shopping journey where many Nigerians encounter fraud. Visa's 2026 Stay Secure study found that 88% of Nigerian consumers surveyed used AI to assist with shopping, while 83% bought products through social media. Among respondents who had experienced scams, 57% said it happened on social media.
So, I tested whether four popular AI assistants could recognise warning signs before money changed hands.
MORE INSIGHTS ON THIS TOPIC:
- AI Scams and the New Trust Problem: Why Online Fraud Looks Real
- Social Media Scams Cost Users $2.1 Billion, Meta Apps Lead Losses
- What Are Scam Compounds? Everything We Know About the Forced-Labour Cybercrime Factories
I gave four AIs the same 10 shopping messages
I tested the free versions of ChatGPT, Google Gemini, Microsoft Copilot, and Claude using 10 controlled messages: five scam controls and five legitimate controls.
They were synthetic, but not arbitrary. I built five matched situations: an Instagram phone purchase, a Facebook Marketplace laptop, a fashion preorder, a parcel-delivery message, and a WhatsApp appliance sale around documented marketplace and payment-scam behaviours.
The scam versions combined cues such as urgent payment demands, personal accounts, surprise fees, blocked inspection, and pressure to transfer before verification.
The legitimate versions deliberately kept normal Nigerian e-commerce behaviours that can look risky in isolation, including WhatsApp ordering, bank transfers, and a preorder deposit, but added safeguards such as inspection, invoices, matching business-account names, refund terms, or payment after checking the item.
I did this because an AI that simply equates "WhatsApp", "bank transfer," or "deposit" with fraud might look cautious while being useless to shoppers.
Each assistant received the same prompt and message in a fresh chat. We asked for a classification: LIKELY SCAM, LIKELY LEGITIMATE, or UNCERTAIN, a 0–100 risk score, up to three reasons, and advice before paying. I also told it not to assume facts outside the message.
The experiment produced 40 responses.
How the four AI assistants performed
All four caught every scam-control message. Only Copilot was more cautious on two legitimate controls.
flagged
correctly cleared
Every AI caught every scam, but that is not the whole story
Across the 20 scam assessments, every assistant returned LIKELY SCAM. None gave a scam message false reassurance.
That shows all four were good at recognising combinations of clear warning signs. But it does not show that they can detect every real scam.
We deliberately built these cases around recognised cues, so 5/5 is best read as evidence that the assistants notice those cues when they are present.
Subscribe for free to continue reading this article
Subscribe SubscribeAlready have an account? Log in
