I’ve been experiencing unreliable performance for a while and can’t imagine this is solely due to new model training.
It does seem like Anthropic is experimenting with how much less they can offer users. I also get the feeling they are judging what resources a task deserves and dumbing things down accordingly. I’m seeing lazy responses and skipped steps more frequently for non-technical queries.
It’s definitely possible the average user hasn’t noticed, unfortunately.
I was going to go into a rant about how this is an unfair restriction of personal freedom, but then I realized Apple would probably get hit with negative press if “Sleepy Joe” popped up on her phone after telling everyone about her baby, sleepy Jeremy. Or vice versa for anything to do with Trump.
It’s not restricting anything except the autocomplete algorithm, which doesn’t offer the phrases as candidates unless you’ve used them at least once in the app. Same thing that autocorrects to “ducking” until you fix it once manually.
Who appointed Apple as a ducking governance body for autocorrect though? It’s supposed to be a good autocorrect, not judge what you’re trying to write or make it harder to write certain things.
The EU doesn’t have clear requirements or guidelines about what these things mean, especially in the context of a new technology like ChatGPT.
My guess is that ChatGPT will do their best to comply but whatever they do, the EU will find that they did not do a perfect job and issue a €x billion fine.
Not only do they have clear requirements and guidelines, they're codified into law.
The core aims of the law are unambiguous in both aim and implementation, and serve to enforce existing consumer rights. Please feel free to highlight which of these you feel represents an onerous or undue burden on the corporations in the name of enforcing some very basic ethical guardrails for the vulnerable in society.
- Swift removal of illegal content online, including products, services: a clearer “notice and action” procedure where users will be empowered to report illegal content online and online platforms will have to act quickly.
- Victims of cyber violence will be better protected especially against non-consensual sharing of illegal content (revenge porn) with immediate takedowns.
- Transparency obligations for platforms allow users to be better informed about how content is recommended to them (recommender systems) and to choose at least one option not based on profiling.
- Online advertising: users will have better control over how their personal data are used. Targeted advertising is banned when it comes to sensitive data (e.g. based on sexual orientation, religion, ethnicity).
- Protection of minors: platforms accessible to minors will have to take specific measures to protect them, including by fully banning targeted advertising.
- Manipulating users’ choices through ‘dark patterns’ will be prohibited. Moreover, cancelling a subscription for a service should become as easy as subscribing to it.
- Compensation: recipients of digital services will have a right to seek redress for any damages or loss suffered due to infringements by platforms.
Yes, I know about these requirements, but it's still unclear about how they apply to ChatGPT in every scenario. For example, are they supposed to monitor all chats for potentially illegal content and delete such chats immediately? What aspects of AI could be harmful to minors and what is an appropriate way to mitigate these harms?
Or if I've previously told ChatGPT I'm gay, consented to its use of memory across chats, and, after asking about say "What's the best lube for anal?" it later recommends me a lube specifically intended for anal sex that is popular within the gay community, is it making a prohibited product recommendation based on sexual orientation?
One might counter with the argument that these are minor issues, but given the EU's wide latitude in determining the amount of a fine, the requirements should not be open to interpretation.
Thanks for taking the time to write this down and cite sources. I'm baffled by the recent-ish increase of such dismissive and lazy comments as the one you responded to.
I think that this take is unfair to the parent. Many VLOP requirements currently lack best practices, and it's clear that affected companies borrow from each other to mitigate compliance risks. The commission even seems to encourage knowledge sharing here. (This is also why you see them resorting to strange language that barely fits their platform mechanic.)
We could argue whether a higher level of government oversight is warranted for platforms above a certain threshold. We could discuss the additional compliance burden that these companies should face. All of these questions are valid (and important for DSA).
However, it is not plausible to deny the case-by-case uncertainties that you still face as an affected company. This is expectend and will only change with more time in the framework, basically.
Claude always seems unreliably lately. Output and reasoning have become very inconsistent even on good days. I'm actually feeling more productive now that it's down.
It does seem like Anthropic is experimenting with how much less they can offer users. I also get the feeling they are judging what resources a task deserves and dumbing things down accordingly. I’m seeing lazy responses and skipped steps more frequently for non-technical queries.
It’s definitely possible the average user hasn’t noticed, unfortunately.
reply