The Scammer Is Your Most Engaged User

From the Editor.

Last Tuesday, Sen. Maggie Hassan released the results of a yearlong investigation into romance scams on Match Group's apps. Match handed over more than 670 pages of internal documents, and the report, called "Outmatched," is the first time a dating company's internal fraud data has been put on the public record. If you missed it, our newsroom's summary is here.

The short version: as of 2025, most of the machine learning models Tinder used to catch scammers had not been updated in years. Over the same stretch, scammers reached more and more people before Tinder caught them. In May 2024, the median scammer Tinder banned had been shown to fewer than 70 people. By December 2025, it was more than 500, an increase of more than 700 percent.

Most of the coverage has read this as a Match Group story. I think that misses it. Tinder is simply the one app whose internal numbers are now public. Almost nothing in this report is unique to Match, and the finding that matters most applies to every dating app with a ranking algorithm.

In one of Tinder's own internal planning documents, staff wrote that scaled abuse accounts, the category that includes scammers, "usually have a higher exposure / impression rate than regular accounts."

Read that again. Until they were caught, the scammers were outperforming the real users.

The scammer is your best user, right up until he isn't

Think about what a romance scammer actually does on a dating app. He builds a careful profile with photos that perform. He swipes right on everyone. He replies fast, every time, at any hour. He is warm, attentive and relentless. And he wants to move the conversation off your platform as fast as possible.

Now think about what most ranking systems reward. Activity. Responsiveness. Likes sent and likes received. Profiles that get swiped on. Every one of those signals looks better on a scammer than on a real person with a job, a life and three other apps on their phone.

Match told investigators it "does not optimize for impressions," and that high scammer impressions are a byproduct of the scammers' own behavior. I take that at face value. It is also the whole point. You do not have to optimize for scammers to amplify them. You only have to build a ranker that rewards what they are best at, and then wait for detection to catch up.

I have been on the other side of this decision. At The Meet Group, scammers were one of our biggest concerns, and as a public company we did everything in our power to stop them. Here is what we knew for certain: every aggressive sweep cost us. Daily actives went down. Monthly actives went down. Installs and subscribers took a hit. Those are exactly the numbers a public company gets judged on. We made the call to be as aggressive as possible anyway, every time, because it was the right thing to do.

So when a report says scammers were outperforming real users, I believe it, and not just at Tinder. The temptation to leave them alone shows up in every metric an operator reports. Not every operator resists it.

Detection is binary. Distribution is not.

Here is the design choice at the center of the report. Investigators asked Match what it had changed in Tinder's recommendation system to lower impressions for scammers. Match answered that risky accounts are either banned or routed to human review, and in both cases removed from the recommender. Then it added: "We do not tune our recommenders to gradually reduce or 'slowly decrease' the exposure of risky accounts."

So an account is either fully in the deck or fully out. Every impression a scammer collects between signup and ban gets the same distribution a real person gets, and often more. Most dating apps work exactly this way.

That gap is where the damage happens. The report cites Northeastern University research on Facebook that found removing spam and scams prevented at most 30 percent of their predicted engagement, because the algorithm had already distributed most of it before moderation arrived. Detection that runs slower than distribution is mostly cleanup.

At Tinder, detection was running slow. Internal documents from 2025 say most of its machine learning models "have not been updated for years." Match's head of trust and safety, Yoel Roth, told committee staff the company retrains its models sometimes every year or every couple of years. The outside experts the committee consulted disagreed. One said that cadence would make accuracy "degrade quite severely." Two said a platform under constant attack needs to retrain at least monthly, and one said retraining once or twice a year meant a company "did not take fraud detection seriously." Match responded that daily retraining of large models is not done anywhere in the industry. Fair enough. But the problem was never the distance between daily and monthly. It was the distance between monthly and every couple of years.

To its credit, Match moved once the pressure arrived. It made video selfie verification mandatory for new US Tinder users, shipped new detection models, and added 169 new detection triggers while improving 108 more in the first half of 2026. Roth has said the report mischaracterizes the company's investment. The committee welcomed the changes. Its complaint is that they took this long.

The honest part, for all of us

Most dating products are engagement machines. Our dashboards celebrate daily actives, likes sent, matches made, messages exchanged. A scammer pushes every one of those numbers up. On Tinder and Hinge alone, roughly 1.5 million spam, bot and scam accounts matched with at least one US user in the first three quarters of 2025, and nearly 500,000 exchanged contact information. Every one of those matches showed up somewhere as engagement long before it showed up anywhere as harm.

So here is the uncomfortable question for every operator. What share of your engagement is fraud?

If you cannot answer it, your growth numbers include an unknown amount of the thing that will eventually cost you your users, your app store standing, and possibly a letter from a Senate office.

This matters more as the business model changes. Last week I wrote that Bumble's new owners come from the world of advertising, where the money is in daily engagement. In a subscription business, a fake account is a cost. In an engagement-funded business, inflated activity can look like inventory. The controls have to get stronger as the model shifts, not weaker.

Every scam is charged to the whole category

Here is the part that should make every honest operator angry.

When a woman loses her savings to someone she met on a dating app, she does not conclude that one company had a model-freshness problem. She concludes that dating apps are not safe. One victim in the report described a profound loss of trust, in other people and in her own judgment. Another stopped using Match.com altogether. A third put it plainly: some people "lose faith in love." The story they tell their friends is not about one app's detection system. It is about online dating.

That is the cost no dashboard shows. Every scam on any app is charged to the reputation of every app. The operator who sweeps aggressively pays twice: once in their own DAU, installs and subscribers, and again when people burned somewhere else arrive assuming every profile is fake, or never download a dating app again.

This industry has a habit of calling the problem a matter of perception. Match's CEO used exactly that phrase with investors in August 2025. It was not a perception issue. The report shows it was a measurable one, sitting in the data. And the people paying for it are every founder trying to convince a skeptical single person that this time it is real.

The regulatory bill lands on everyone too. Congress has passed the Romance Scam Prevention Act, which will require every dating service to warn users within 24 hours when someone they messaged is banned for fraud, enforced by the FTC. The investigation behind this report was launched jointly by Hassan, a Democrat, and Sen. Marsha Blackburn, a Republican. Reported romance scam losses hit nearly $1.5 billion last year, up from $201 million in 2019. There is no constituency defending that number. Expect the next set of questions to reach companies much smaller than Match.

Doing this right should be a competitive advantage. Today it is a tax. The operators who remove scammers aggressively report worse numbers than the ones who don't, and they inherit the same damaged reputation. That only changes if the people doing it right start measuring it, publishing it, and making it a reason to choose them.

Every scam on any app is charged to the reputation of every app.

What to do this quarter

None of this needs Match's budget. Most of it needs a decision.

The part worth being hopeful about

The best news in this report is how fast the fix worked once someone shipped it. Within a month of a nationwide rollout, Tinder users were running into far fewer bad actors. That is not a ten-year research problem. It is a product decision that was available the whole time.

And it is a decision an independent operator can make faster than an incumbent. You can put a trust budget into your ranker in a sprint. You can put impressions-to-ban on your weekly dashboard by Friday.

Every dating app says it is safe. After last week, safety has a number attached: how many people your scammers reach before you stop them. Find yours, and if it is good, say so publicly. The category's reputation is being set by its worst numbers. It is time the best ones were on the record too.