Safety Filters make LLMs defective tools
woolion.art
woolion.art
The whole thing has then morphed into a moral crusade that makes wokeism seem perfectly logical in comparison.
It's a computer.
It can't have its feelings hurt. It won't be embarrassed.
Conversely, I'm an adult.
I won't feel bad if the stupid AI model says something rude. I personally think it's hilarious.
Sure, sure, allow the people making helpful website support chat bots to dial up the censorship to eleven! Similarly, educational bots helping students with self study should be between guardrails.
However, it is incredibly frustrating that LLMs are the best language translators of any kind by far, but I can't use them to translate movie subtitles because "this contains violence".
No shit it contains violence, it's an action movie.
It's legal, it's harmless fun, it's available for streaming on every computing device made in the last decade.
Get a grip people! It's just words.
As a professional developer, as an API user, as a grown man who knows how these things work, I should have the option to turn the safety filters off. Not "dial down", but "off".
Google image search lets me turn the porn filter off.
I can search for the formula for making TNT on any search engine.
WikiPedia lets me read about Tiananmen Square.
These systems treat adults like adults.
Why can't OpenAI, Anthropic, and DeepSeek?
The default should be like the "Safe Search" defaults on search engines: Nothing gross, no ultra violence, no porn.
There should also be the soft pink fluffy version for children, helpful AI bots, and the like.
The unfettered base model should be available via API keys for professional users, with the usual warnings of: "This can and will be rude, mean, sexist, etc..."
Why can't I have steak? Because a baby can't chew it and might choke?
Little fat man who sold his dream
> Google image search lets me turn the porn filter off.
Google images let you turn the porn filter down, not off. It's not going to readily show you child porn. Actually, the whole internet has this filter, including "unfiltered" sites like 4chan. The darkweb exists, sure, but this isn't a model any company in their right mind will surface to users.
Anyway, I think you have a point in that you can disable Google's porn filter, but you cannot disable Google's wrongthink filters. Google search is manually tweaked to redirect many types of queries to their opposite, safe alternative (like searching for Alex Jones will only give you 'why is alex jones such a bad person', and not act as a search engine, which would be neutral).
Things have changed gradually and there hasn't been so much pushback on this, unfortunately.
GP didn't.
> like searching for Alex Jones will only give you 'why is alex jones such a bad person
That doesn't happen here. I get his social media profiles, Wikipedia, news sites, etc.
I firmly believe most people who push for "unfiltered" models are looking for 4chan-equivalent filtering. Filtered for safety, but subtly.
I suspect that true uncensored models, like the darkweb, are not useful for most people. I further suspect the applications they are useful for are not something most people want to publicly associate with (to put it lightly).
Note that ITA I'm not advocating for "no" safety filters but a better implementation.
In this analogy the equivalent being used by LLM makers is “guilty puritan” or “imam insisting on the niqab to rid men of all impure thoughts.”
There’s degrees of censorship, but current AIs enforce the level that is suitable for the youngest of minors.
We’re talking Disney World levels, or Frozen on Ice the musical.
I’m a big boy, I’m allowed to stay up late and watch The Terminator.