Maintaining control over artificial intelligence responses remains difficult. An American company claims to have found a simple way to circumvent censorship in certain models. This solution could change the way AI is used in sensitive sectors.
Artificial intelligence is increasingly used in all fields since the arrival of ChatGPT. But behind its prowess, a problem persists: some models refuse to answer certain questions deemed sensitive. This form of automatic censorship sparks debate, particularly when it limits professional or educational uses.
To address this problem, the American startup CTGT has developed a new method. Unlike traditional techniques, it does not require reprogramming the artificial intelligence or retraining it for hours. Their approach consists of: precisely pinpoint what is blocking responses in the model, and then adjust these elements to make the AI freer, without compromising its reliability or accuracy. The company insists that their system allows for the removal of unnecessary censorship without opening the door to dangerous responses, a sensitive subject since AIs like Elon Musk's Grok 3 went haywire at launch or DeepSeek, which was capable of providing guides for making chemical weapons.
CTGT promises to lift AI censorship without affecting its proper functioning
In tests conducted with 100 sensitive questions, the basic DeepSeek model answered barely 32% of them. After modification with CTGT's method, it answered 96% of the questions, without any slip-ups or dangerous content. This technique acts directly when using the AI, making it possible to change its behavior without touching its original programming. The system is also reversible: it is possible to deactivate or adjust the settings at any time.
This advancement could be of interest to many companies that want to use AI while respecting their own internal rules. CTGT explains that its solution avoids heavy investments in reprogramming, while making the models more adaptable to different uses. While models like DeepSeek are seen as potential risks by American authorities, better controlling what this technology can and cannot say is becoming a crucial issue in sectors such as cybersecurity, finance, and healthcare.
Source: CTGT
0 Comments