When parents ask whether a game is safe, the answer is often: "they have AI moderation." A new report shows what that actually means.

Australia's eSafety Commissioner used its legal powers to require four of the most popular gaming services among children, Roblox, Minecraft, Fortnite and Steam, to report how they protect young players from child sexual exploitation, violent material, cyberbullying and online hate. It published the findings on October 1.

The regulator is careful to say that issuing a notice "does not necessarily mean that a company has deficient safety practices." But the differences between services are striking.

Roblox: an AI that caught about half. Roblox's "Image Auto-mod" tool, used to detect violence and child exploitation material across every image on the service, "had a recall rate of only 51%." In plain terms, it detected roughly half of what it should have. The regulator notes low recall "can indicate under-enforcement."

Steam: chats largely unscanned. Valve "did not use tools to proactively detect harms in text or voice chat on Steam Chat," apart from a link-detection tool that "mainly blocked scams." Valve also didn't use hash-matching tools to detect known abuse material, and relied on users to self-declare their age.

Minecraft: say you're 18, get in. Players "self-declared as 18 or older" could access Minecraft servers hosted by third parties that were "not proactively moderated by Mojang Studios."

Fortnite: slow on one key category. Epic Games took the longest of the four, 18 hours and 53 minutes, to assess automated reports of possible child exploitation material (other than grooming and sexual extortion).

Stale training data. Roblox and Valve trained their language-analysis models only on internal data, and some tools weren't updated regularly.

Fairness matters here:

eSafety's own research found that 89% of Australian children aged 8 to 17 played online games in the past year, and 40% have played online with people they didn't know offline.

Our view: automated moderation is doing much of the watching in kids' games, and it's genuinely useful. But this report shows that "we use AI moderation" can mean very different things. One tool might catch most of the harm. Another might catch half. Some spaces may not be scanned at all.

Three practical takeaways:

Ask which spaces are moderated. Official servers and community-run servers can have very different protections. Minecraft's third-party servers are one example.

Age settings matter. Several services relied on kids to state their own age. Setting up a child account properly unlocks the protections built for children.

Talk about chat. Text and voice chat were the least consistently protected areas. Make sure your child knows they can tell you about anything that feels wrong, without losing their game.

A conversation starter: ask your child who they play with online, and whether they know them in real life. Listen first.

AI moderation isn't a force field. It's a tool, and this report shows how well each company's tool actually works.

Source: eSafety Commissioner (Australia), "Responses to transparency notices: Online game and gaming-adjacent services," October 2026 report. https://www.esafety.gov.au/industry/basic-online-safety-expectations/online-game-and-gaming-adjacent-services

Quotations verbatim from eSafety's key findings. The interpretation and family guidance are ours.

Disclosure: drafted with Claude, made by Anthropic.

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter