A £12 “Apple” charger arcs after ten seconds on a test bench. Inside, someone packed modeling clay just to make it feel heavier in your hand.
That charger sat on a major marketplace for months before anyone caught it.
Not chatbots. Not generative art. Image-recognition models trained to spot a fake charger before it ships to your door — that’s where AI is doing real work right now.
Why Fake Chargers Still Flood Online Marketplaces
Counterfeit electronics stopped being a niche problem a while back. The OECD’s most recent study puts the global trade in fake goods at $467 billion, and digital retail keeps pulling more of that volume online.
Online channels now carry 83% of counterfeit trade, up from 64% in 2015. A fake storefront takes minutes to launch and just as little time to vanish once a regulator starts asking questions.
Consumer group Which? bought 15 USB chargers from Amazon, AliExpress, B&Q Marketplace, Debenhams and eBay, then ran them through electrical safety testing. Nine could deliver an electric shock. Eight carried a fire or explosion risk. Not one of them should have been legal to sell.
Nearly a decade earlier, a test of 400 counterfeit Apple chargers sourced from eight countries found 99% failed basic safety checks. Same category, same failure rate, different decade.
Where This Hits Refurbished-Phone Buyers Hardest
Anyone buying a Refurbished iPhone 15 is usually shopping for a new charger at the same moment — and that’s exactly the purchase window counterfeit listings target. A new owner rarely knows what the genuine accessory should feel or weigh like.
USB-C compatibility raises the stakes further. USB Power Delivery lets a device and charger negotiate the correct power level automatically, but only if the charger’s internal components are actually built to spec. A counterfeit board can print “PD supported” on the box without implementing any of it safely.
Marketplaces listing a refurbished iPhone next to generic third-party accessories sit in a similar blind spot. Detection models trained to catch trademark violations don’t always flag a no-name charger falsely claiming device compatibility, because there’s no logo being stolen — just a safety gap that nobody has coded a rule for yet.
How the Detection Models Actually Work
Manual review can’t keep pace with millions of daily listings. Machine learning earns its keep in exactly that gap.
Modern detection systems stack several signals rather than leaning on one:
- Image similarity models catch packaging mismatches a human reviewer would scroll straight past.
- Price disparity flags trip when a “genuine” charger lists far below the manufacturer’s floor.
- Seller-behavior clustering groups accounts sharing shipping addresses or upload patterns — the digital fingerprint of a counterfeit ring.
- Content anomaly checks compare stated specifications against certification claims buried in the listing text.
Trained on large sets of authentic-versus-fake examples, these classifiers now report accuracy above 99% under controlled testing. The accuracy number matters less than the fact that they run around the clock. A listing pulled within hours never reaches the volume a weekly manual sweep would.
The Gap the Accuracy Number Hides
Detection speed and removal speed aren’t the same thing, and marketplaces rarely advertise the difference.
A model can flag a listing in seconds. Actually taking it down still runs through legal review, seller appeals, and rules that shift by jurisdiction. Reported takedown timelines range from a few hours to several days depending on the platform and category.
That gap is likely where the modeling-clay charger sat undisturbed. Something probably flagged it early. The removal process just didn’t move fast enough to matter before someone bought it.
Which produces a strange kind of trust paradox: the technology can catch nearly every fake before checkout, yet the fakes buyers actually run into are rarely the ones the model missed — they’re the ones stuck in a queue.
Three Checks No Algorithm Runs for You
AI detection has genuinely improved the odds of avoiding a dangerous charger. It hasn’t eliminated the risk, and treating it as a solved problem skips the basic checks that still matter.
- Look for a manufacturer name, model number, and compliance marking in the listing photos themselves, not just the title text.
- Treat a price far below category average as a warning sign, not a bargain — it’s the same signal the detection models are trained to catch.
- Confirm the listing states actual power output and charging standard rather than a vague “fast charging” claim.
None of that takes technical knowledge. It takes the five seconds an algorithm doesn’t get to spend on your behalf.
What AI Still Can’t Fix
Detection models are only as useful as a marketplace’s willingness to act on what they flag. The technology side of this problem looks close to solved. Enforcement isn’t, and until platforms close that gap, buying from a seller who never needed an algorithm to prove they’re legitimate is still the safer bet.
Related: How AI Search Decides What Your Business Is Known For in 2026
