Few problems can harm a platform faster than illegal content. That’s why the Safety Team at Wingtalks focuses on prevention. In the past year, we rebuilt the way we do reviews. The process now runs on machine learning, so we can spot harmful material early, respond fast, and keep the platform safe. Below, we take a close and practical look at how the system works.
Illegal Content Is Now the Biggest Risk for Platforms
In 2026, illegal content became a legal duty. The EU’s Digital Services Act and the UK’s Online Safety Act both ask platforms to remove these types of material and report it clearly. Under the Digital Services Act, a platform that fails to act can be fined up to 6% of the money it makes worldwide each year, according to the European Commission.
But fines aren’t the only cost. Scams and other dishonest acts wear down trust. And once trust is gone, people leave.
Part of what makes illegal content so persistent is the perception of impunity. When people act through avatars, normal social accountability breaks down. That gap is what bad actors exploit.
The U.S. Federal Trade Commission reports that scams that start on social media cost people $2.1 billion in 2025. Today, every platform has to keep its users safe.
Why Did Wingtalks Rebuild Its Content Review System?
The old setup worked at small volumes, but couldn’t keep up with a growing product and trends. These problems made us start over.
The first was the law. The Digital Services Act and the Online Safety Act established new expectations for how everything should work, and we needed to adapt to them. The second reason was about the platform itself. The social platforms market itself now dictates new expectations for moderation effectiveness and for creating a comfortable space for users, and as a platform, we aim to meet those expectations. Machine learning solved both issues. It lets us review more content faster. Also, it has the potential to spot deceptive accounts by how people act and what they share.
Inside the Platform’s Security Measures
Trust & safety on Wingtalks are priorities. Automated review handles the scale, while behavioral signals catch risky patterns. Also, the Safety Team steps in when a human call is needed. You can see how the main pieces fit together next.
Automated Review of Shared Content
Every piece of shared content gets a machine-learning score for how likely it is to break the rules. This covers public areas, like profiles and the feed, as well as material shared in private. The system removes clearly illegal media on its own.
Behavioral Scores and Preventive Blocks
Removal only deals with content after it’s posted. So, we also look at how accounts behave. The models weigh many signals at once and watch for patterns that match serious violations. If the signs clearly show a serious problem, we block the account for good.
Serious Complaints Reach Us First
Reports from users are always helpful. So, every flag goes straight to the software, which sorts them and puts the most serious ones at the top. That way, we can always deal with urgent flags first and keep the Wingtalks website safe.
The Numbers Behind the Platform’s Safety
When safety is at stake, we tend to look at what the numbers are actually showing us. The data below comes from research we have carried out internally: how much harmful content we are managing to catch, how quickly we can act on it, and how the volume and nature of reports have changed over time.
A Measurable Drop in Illegal-Content Flags
After the rebuild, complaints about illegal content dropped 26%, from 1,323 cases to 979. The most serious categories fell too: violence went from 5.96% to 4.70% of all reports, and child-safety cases from 3.38% to 2.39%. The machine-learning model now catches illegal photos with 73% accuracy. Taken together, that’s a real improvement in security measures on Wingtalks.
Faster Flag Review and What That Means for Users
When you flag something, you generally hope it gets dealt with within minutes rather than days. That is why we have built the way we handle reports around speed as a primary consideration: roughly 95% of flags are now typically resolved within a 24-hour window, and the majority of those are taken care of on the same day they come in. The process has gotten faster and more consistent over time, and there is ongoing work to keep pushing it in that direction.
The Approaches We Considered and Dropped
We considered other approaches before this one and ruled two out. Neither could handle the scale.
The Problem With Reports Alone
One option was to rely on flags alone in private spaces and step in only when one member flagged another. We dropped it. With only reports, harm can build for weeks before anyone speaks up. It’s usually too late by then. So instead, the software reviews shared content on its own; no human reads your private conversations.
Manual Review Can’t Scale
The other option was risk detection by hand. In that case, a person would approve every profile and post before it went live. But the math doesn’t work. At our size, a manual check would be slow and costly, and members would wait hours just to see their own posts go up. The software does the bulk of the work, and our team steps in where human judgment counts.
The Hard Part: Teaching Machines to Recognize Harm
The Wingtalks safety system tested us. Good data was in short supply, and many of the calls were in a gray area. Here’s how we handled each problem.
Sourcing Training Data the Responsible Way
Good models for platform security need plenty of clear examples. However, harmful content is rare and hard to collect. So the team built a mixed training set from three sources.
-
Openly licensed public datasets from Google APIs, Hugging Face, Roboflow, and Kaggle, used under Creative Commons licenses.
-
Anonymized data from past reports, so the models learn the real language and context of the platform.
-
Synthetic examples of policy-violating content, generated using a language model.
Together, these three sources cover the rare cases and deliberate tricks that wouldn’t show up on their own. They also let us train responsibly and keep everyone’s private material safe.
The High Cost of a Mistake
A machine automation system for online safety isn’t perfect. On its own, it may remove a safe post or block a member who didn’t break rules. So we plan for errors. We’re strict with harmful content, but fair to respectful members. It takes daily work to get both right.
Many Reports Miss the Mark
In the earlier stages, a fairly large share of incoming complaints were not hitting the mark. That piled extra work onto the system and added friction to how quickly the team could get to the threats that mattered most. The software can screen out a good number of those low-quality reports, but the more useful fix has been to make the rules clearer for users. Better-informed reporting tends to produce sharper feedback and fairer outcomes.
What’s Next for Safety on Wingtalks?
The work isn’t done. The people who abuse the platform change their methods often, so the safety tools have to adapt.
Bad actors constantly update their tactics, and the ML models have to keep pace. The team regularly retrains them on new evasion patterns that weren’t in the original datasets but have since appeared in real usage.
From Media Analysis To Text in Context
The current system mostly checks photos and videos. The next step is to look at the words people share, and to read them in context. The settings matter because the same words can be fine in one chat and clearly against the rules in another.
Think about how much the meaning changes in a public post. For example, even something as small as a caption saying “hey” text lands differently depending on who posts it and where.
Helping the Community Report What Matters
A moderation system tends to work better when the people using the platform are actively involved. We want members to be able to flag harmful content whenever they come across it, so we made the full set of rules on the Wingtalks website even more accessible and visible to users and added clearer prompts in the app that explain what each report type is for. The goal is to end up with fewer reports that go nowhere and to make it easier to get real help in situations where something has genuinely gone wrong.
Long-Term Trust on Wingtalks
Safety is something we treat as an ongoing commitment. The level of trust that members have in Wingtalks at any given point is built on a handful of things that are worth laying out:
-
Early detection that is able to catch dishonest behavior before it has a chance to spread further through the platform
-
A Safety Team that has been trained to handle the more difficult cases and knows when those cases need to be escalated
-
Guidance that is clear enough to actually help members take steps to protect themselves
-
Verification processes that are set up to kick in at the points where the risk level happens to be higher
We also make a point of sharing what we learn on Wingtalks.com, on the grounds that being open about how the system works tends to make the platform safer in the broader sense. The system is in a stronger position than before, but there is always more work to be done in this area. Getting it to a better place remains a focus for us.
FAQ
How does the platform detect illegal material?
The site is built around protection that is intended to be proactive rather than reactive. There are several layers of ML models that can check both shared content and member behavior for signs of illegal material, with scanning running automatically across public posts and private content alike. Violations that are clear enough to act on are removed before anyone submits a report about them.
Can users flag suspicious behavior on the platform?
Yes. Abuse reporting is possible with buttons that appear throughout the platform. A “flag” is one of the most useful ways you can help keep the community safe.
What happens after a complaint has been submitted?
Once you send a report, the software sorts it and ranks it by urgency, so the most serious cases move to the front of the line. The Safety Team then reviews what was flagged and acts wherever the rules have been broken. About 95% of cases are resolved within 24 hours.
Does automated moderation read private messages?
Automated processing checks shared content for illegal material. It runs as software, not a person who reads your conversations. The aim is to find serious harm and protect your privacy at the same time. You can find fuller details on Wingtalks.
