Hello,

The more I look at your safety age protection and message filtering systems, the less I can agree to them, despite I support efficient protective and preventive systems.

A good system shall ask three basic questions:
- Does the message AND past behavior indicate the user is (not) underage?
- Is the person's message (obvious) satire/humor?
- Is the (counter)-action appropriate? (required by law such as EU-DSA)

Example one: people are banned for messages like "I am xyz years old" (usually below 13). Their account gets immediately locked until age verification. Despite them obviously joking or having an clear account history, indicating they are adults without doubt

Example two: recently Twitch introduced and removed multiple times message filters banning the beginning of the word "***"(an). Weird, especially in non-English contexts

Example three: the public Twitch start page shows FSK 18+ streams being freely accessible without login for underaged people

How is the current system upholding the requirements outlined in the question and examples? While fulfilling legal compliance and cost minimization?

As outlined in (example three), everyone can see, regardless of age adult content on Twitch's main page without login, while people sending jokes in public chat getting banned (example one) or as in (example two) messages beginning with blocked words are deleted or banned, even outside the lingual context as "***" is not discriminatory in non-English lingual contexts. Twitch's system obviously has no way to differentiate and contextualize user messages. Despite all ads and recommendations on Twitch are indeed contextualized based on the user data available.

From a business perspective, we need to look at income-cost-risk-benefit analysis. The streaming market environment has ever more rivalry between competitors (high risk and cost), so the incentive for user's to switch platform (medium risk and cost), which are the requirement for getting advertisers who may follow them (high risk, very high cost) who pay for ads, is greater than ever before. On the other hand, the legal risk (high cost & risk) of not blocking underaged costs a lot, and simple-filter systems are operationally cheap (benefit) with additional medium cost of post-moderation (systems) for correcting the high error-rate (medium cost), which increases user's frustration as in bad user experience (medium risk), due to lacking contextualization. The platform's continuing decline due to destruction of the user base may outweight all of these considerations in terms of risk and cost, and the basic right infringement (high risk) may even increase this, especially in times of global political and economic pressure (e.g. shown by the introduction of Twitch Turbo; high risk and cost). While contextual moderation systems may look expensive at first glance, the are much cheaper (benefit, low cost), efficient (benefit, lower cost), and posing less legal risk (benefit, less risk and cost), if we look at the fact Twitch has already age-profiled data prepared (benefit, low cost) waiting to be reused for this purpose, as they run ads with it (benefit, no cost).

Legally, Twitch is a public platform under EU-DSA, every moderation can only withstand legal questioning when being transparent, balanced and weighted (Proportional & Compliant). So Twitch is bascially a public townsquare obligated to uphold and protect basic rights (e.g. freedom of opinion/expression), anti-censorship and children and youth protection regulations (restricting adult or harmful content access). Regardless of Terms of Service and contract law, with the limitation of Criminal Code. Every message, even by a minor is a opinion, thus touches forementioned basic rights, technically above all other rights, aslong unconflicting. If two legal requirements or legal principles are conflicting, we have to logically assess what is legally higher, the established precendence, what and how we act on it, then we have to transparently communicate our decisions and check whether it can legally uphold. The approach to generally ban people for messages or delete them using simple-filters infringes their freedom of opinion/expression more and by the post-moderation violates the cost/effort and harm reduction legal principle more. Contextualization increases compliance, bolsters the legal argument and reduces harm & cost by reusing the existent age-profiling data of the advertising & recommendation systems.

Thank you very much!