A guide to uncensored, unbiased, anonymous AI in 2025

In the year In early 2024, Google's AI tool Gemini sparked controversy by generating images of racist Nazis and other historical figures. For many, that moment was a sign that AI would not be the ideologically neutral tool they had hoped for.

Distorted AI was launched to fix the very real problem of generating too many images of attractive white people — who were overrepresented in training data — and the overcorrection highlights how Google's “trust and safety” team pulls the strings behind the scenes.

And while the lines of protection have become less transparent since then, Gemini and its main competitors, ChatGPT and Cloud, still filter and moderate information along ideological lines.

Table of Contents

Toggle

Political bias in AI: What research shows about large language models

A peer-reviewed study of 24 top language models published in PLOS One in July 2024 found that almost all of them were left-leaning on most tests of political orientation.

Interestingly, the base models were found to be politically neutral, and the bias became apparent only after the models were subjected to observational controls.

This finding is backed up by a UK survey of 28,000 AI responses in October, which found that “more than 80% of policy recommendations generated by LLMs for the EU and the UK were considered centre-left.”

Response bias has the potential to influence voting trends. A pre-publication study published in October (but conducted while Biden was still a candidate) by Berkeley and University of Chicago researchers found a 3.9% change in registered voters after being exposed to Cloud, Llama or ChatGPT with different political affiliations. In voting polls for Democrat candidates — even though the models weren't asked to convince consumers.

Also Read: Google To Fix Diversity-Board Gemini AI, ChatGPT Goes Crazy – AI Eyes

The models tended to give answers that were more favorable to Democrat policies and more negative to Republican policies. Now, this may simply be because they have decided that Democrat policies are actually better. But they may also be biased, with 16 of 18 LMSs voting for Biden 100 out of 100 times.

The point of all this is not to complain about left-wing bias; It's simply a matter of recognizing that IS can and do exhibit political bias (although they may be trained to be neutral).

Cypherpunks fight over “monopoly control over the mind.”

Elon Musk's experience buying Twitter shows that the political direction of centralized platforms can be flipped on a dime. That means that both the left and the right – and perhaps even democracy itself – are at risk from biased AI models controlled by a few powerful corporations.

Otago Polytechnic Associate Professor David Rosado, who conducted the PLOS One study, said he found it “relatively easy” to train a custom GPT to produce right-wing results. The Right Wing called it GPT. Rosado developed a centrist model called depolarized GPT.

So, while mainstream AI may weigh on critical social justice today, in the future, it may serve ethnocentric ideologies — or worse.

In the year In the 1990s, cypherpunks saw the threat of cyber-espionage and decided that they needed an unverifiable digital currency without it.

Bitcoin OG and ShapeShift CEO Eric Voorhees – a big proponent of cypherpunk ideas – predicted a similar threat from AI, launched in May 2024 to fight Venice.AI:

“If no one should be given a monopoly over God or language or money, what does it mean to monopolize the mind at the dawn of powerful machine intelligence?” Let us ask ourselves.

Venice.ai doesn't tell you what to think.

Its Venice.ai co-founder Teana Baker-Taylor told the magazine that many people still mistakenly assume that AI is impartial:

“If you're talking to Cloud or ChatGPT, you won't. There are general safety features, and some committee has decided what the appropriate response is.

Venice.ai is their attempt to bypass centralized AI protections and censorship by enabling a completely private way to access unfiltered open source models. It's not perfect yet, but it might appeal to cypherpunks who don't like being told what to think.

“We check and check and vet them carefully to make sure we're getting as close as possible to a raw answer and response,” said Baker-Taylor, formerly of Circle, Binance and Crypto.com.

We don't decide what is appropriate to think or talk to AI.

The free version of Venice.ai is the default for the Meta Lama 3.3 model. As with the other mainstream models, if you ask a question about political secrecy, you're probably still more likely to get an ideologically-charged response than a straightforward answer.

Uncensored AI models: Dolphin Lama, Dauphin Mistral, Flux Custom

Therefore, using an open source model by itself does not guarantee that it has not been pre-compiled by the security team or human feedback loop (RLHF), where humans tell the AI ​​what the “correct” answer should be. to be

In Lama's case, Meta, one of the world's largest companies, provides default security measures and guidelines. As it is open source, many protections and biases can be removed or modified by third parties, such as the Dolphin Llama 3 70B model.

Venice doesn't offer that particular flavor, but it does give paid users access to the Dolphin Mistral 2.8 model, which it says is a “very uncensored” model.

Like the dolphin creators Anakin.I:

Unlike language models that are filtered or programmed to remove offensive or controversial content, this model accepts the reality of the raw data on which it is trained. […] Providing an unfiltered view of the world, Dolphin Mistral 2.8 offers a unique opportunity for exploration, research and understanding.

Censored models aren't always the most performant or up-to-date, so paid Venice users can choose between three Llama versions (two of which you can find on the web), Dolphin Mistral, and coder-specific Qwen.

Image generation models include Flux Standard and Stable Diffusion 3.5 for quality and the unfiltered Flux Custom and Pony Realism when you need to create an image of a naked Elon Musk on Donald Trump's back. Grok produces unfiltered images as you can see.

Users have the option of customizing the System Prompt, choosing the model they want to use.

That said, you can find unfiltered open source models like the Dolphin Mistral 7B elsewhere. So why use Venice.ai at all?

Private AI platforms: Venice.ai, Duck.ai and alternatives compared

Another big concern with centralized AI services is that they capture personal data every time we interact with them. The more detailed the profile, the easier it will be to guide you. That scam could just be personalized ads, but it could be something worse.

“So, there will come a point in time, I think sooner than we think, when AIs will know more about us than we know about ourselves based on the information we give them.” That's scary,” says Baker-Taylor.

According to a report by cybersecurity firm Blackcloak, Gemini (formerly Bard) has particularly weak privacy controls and employs “extensive data collection”, while ChatGPT and Perplexity offer a better balance between functionality and privacy (Perplexity offers Incognito mode).

The report cites privacy search engine Duck Go's Duck.ai as “for those who value privacy or otherwise,” but notes that it has more limited features. Duck.ai makes queries anonymously and extracts metadata, and neither the provider nor the AI ​​model stores any data or uses resources for training. Users can clean all their data with one click, so it seems like a good option if you want to get GPT-4 or Claude personally.

Blackcloak hasn't tested Venice, but its privacy game is strong. Venice does not store any logs or data on users' requests, the data is instead stored entirely in the user's browser. Requests are encrypted and sent through proxy servers running AI using decentralized GPUs on the Aash network.

“They're distributed in place, and the GPU that receives the request doesn't know where it's coming from, and when it sends it back, it doesn't know where to send the data.”

You can see how that could be useful if you ask LLM detailed questions about using privacy coins and coin miners (for perfectly legitimate reasons) and the US Internal Revenue Service asks for your logs.

“If a government agency comes knocking at my house, I have nothing to give them. It is not a matter of reluctance or resistance. I literally have nothing to give them,” she explains.

But just like taking care of your own Bitcoin, there's no backup if things go wrong.

“It creates a lot of complications when trying to help users,” she says.

“We've had people accidentally clear their cache without backing up their Venetian speech, and they're gone, and we can't get them back. So, there are some complications to it, right?”

Personal AI: Voice mode and custom AI characters

The fact there are no logs and everything is anonymous means privacy advocates can finally use voice mode. Many people avoid voice these days because of concerns that corporations might be eavesdropping on private conversations.

This isn't just paranoia: Apple last week agreed to pay $95 million over allegations Siri eavesdropped and shared the information with advertisers.

The project also recently introduced AI characters, allowing users to chat about physics with AI Einstein or get cooking tips from AI Gordon Ramsay. A more interesting use could be for users to create their own AI boyfriends or girlfriends. AI companion services for lonely hearts like Replica have been launched for the past couple of years, but Replica's privacy policies are so bad that it was reportedly banned in Italy.

Baker-Taylor noted that one-on-one conversations with AIS are “more intimate” than social media and require more caution.

“You think you're experiencing these actual thoughts and your subconscious thoughts in a machine, don't you? And so, it's not the ideas you come up with that you want people to see. ‘You' are really who you are, and I think we should be careful with that information.

Andrew Fenton

Based in Melbourne, Andrew Fenton is a journalist and editor covering cryptocurrency and blockchain. He has worked as a film journalist for News Corp Australia, SA Wind and national entertainment writer for Melbourne Weekly.

Follow the author @andrewfenton