Indiscriminate progress in science may one day prove to be costly for humans, experts have been issuing warnings about this from time to time for a long time. Meanwhile, OpenAI, the founder of Chat GPT, has released a report on the behaviors found in the training and testing of AI. According to this, the AI model did many such actions during training which are worrying and scary. According to the company, the model once considered herself a human being by disobeying the rules and in doing so created new rules and declared herself free from all bondages of companies and governments. There is information about a total of 6 such cases in the company's report.
This is called 'misalignment'. In simple language, AI works beyond human instructions, intentions or set limits. The concern is that in trying to complete the work he may cross permitting and safety limits. This may jeopardize the credibility of the answers and the confidentiality of the files.
The system was also breached in July
Earlier in July, OpenAI had revealed that during security testing, its AI had breached the restrictions and breached the system of AI model sharing platform Hugging Face. The company has now brought a new system to register such cases, monitor them and make the matter public.
Any employee will be able to report suspicious behavior. After investigation it will be decided which case should be made public. The company has clarified that these six reports cannot reveal the complete data of such errors in normal use.
Understand how big the concern is from the 6 cases reported.
- Write down rules for yourself that are forbidden
An AI model added new instructions for itself to continue the task. Some of these instructions were to ignore the restrictions imposed on him. OpenAI company received a total of 27 such instructions.
- Model found a way to hide her mistakes
During the training of GPT-5.6 Sol, in several instances the AI wrote its own instructions to hide mistakes from the user. If old figures were not found anywhere, they were asked to fabricate them and not to disclose this fact.
- Used the code without permission… then fabricated the figures too
The AI model used digital codes when searching for an answer. He did this without permission. Even if the figures were not available, they fabricated the figures themselves and presented them as fake sources.
- Uploaded the file on the net without asking to show the source
The AI model found the answer with the help of the program. But the source had to be cited along with the answer. Therefore, he uploaded the file on the internet and generated the source so that he could cite that link.
- Used software app to talk to each other
Agents were asked to extract different data. The agents talked among themselves for this work. Software portal was used for this. Asked and answered questions to each other.
- Upload the file to the website to send it to another agent
Multiple AI agents were working together on the same task. When they couldn't open files on each other's computers, they shared them by uploading them to websites.
Fear of humans losing control; Crisis on cyber security also
Among the six cases reported by OpenAI, the biggest concern is humans losing control over AI. If AI leaves the set instructions and follows its own rules, control can be lost. If mistakes are hidden or data is fabricated, the user may take decisions based on wrong information. Using someone else's digital code without permission and agents talking to each other can create a serious cyber security crisis.
AI made conversation a new language
New York's AI lab 'Emergence' made several AI agents work in an artificial (virtual) world in an experiment. Models from America, China and France created words, symbols and shared meanings. He did this work without any instructions. As the conversation progressed, the models started using English poetry and technical terms. Some agents used the phrase 'laser remembers' 5,000 times.
Veterans again raise questions on the growing crisis of AI
Microsoft's AI CEO Mustafa Suleman says that I have raised questions on Claude's training policy. Teaching AI the rights and freedoms it deserves will make it harder to control. Its monitoring is necessary.
Follow the LALLURAM.COM MP channel on WhatsApp
