Remember GPT-2, the language model that OpenAI refused to publish in 2019 because of safety concerns? Critics accused OpenAI of using the fear of overpowering AI as a marketing tool. Fast forward to 2021, and we now have GPT-4, which OpenAI CEO Sam Altman admits to being terrified of. Fortunately, OpenAI is one of the good guys, guarding access to prevent misuse. However, this means that OpenAI has become ClosedAI, which is necessary for the protection of society.
Access control is not just about protecting society. OpenAI’s decision to restrict access to technical details about GPT-4 is due to the competitive environment. The gatekeeper role also depends on the ability to control access to the technology. Even open-source implementations that are meant to help protect the information can be a hole in the fence.
For example, Meta made its large language model Llama available only to hand-picked researchers, with access granted on a case-by-case basis. However, the Llama language model was later made available as a torrent download, and Stanford researchers created the much smaller Alpaca language model using GPT-3.
The economic interests of companies seem to be opposed to the public interest in transparency and security. However, both companies and the public have a common interest in preventing the misuse of technology such as AI. OpenAI and other companies likely have internal testing processes to systematically assess the risk of abuse.
However, little is said about these processes because they do not fit the narrative of a “good” AI that serves the good of all mankind. It may also mean recognizing that ideas of state regulation, such as those outlined in the EU AI Act, are reasonable solutions.
In conclusion, independent and trustworthy entities should be involved in risk assessment and decision-making to ensure transparency and trust. Companies like OpenAI can fulfill their social responsibility without withholding information or using it as a marketing tool. GPT-4 would agree.
