Get your free copy of Editor's Digest
FT editor Roula Khalaf picks her favourite stories in this weekly newsletter.
Supporters and opponents of AI tend to agree that the technology will change the world. People like OpenAI's Sam Altman envision a future in which humanity thrives; critics, meanwhile, predict social chaos and excessive corporate power. Which predictions come true will depend in part on the foundations laid today. But recent controversy at OpenAI, including the departure of its co-founder and chief scientist, suggests that major AI players have become too opaque for society to chart the right course.
According to an index developed by Stanford University, AI leaders Google, Amazon, Meta, and OpenAI are not as transparent as they should be. AI was born from the collaboration of researchers and experts across platforms, but the companies have been silent since OpenAI's ChatGPT led the commercial AI boom. Given the potential dangers of AI, these companies need to return to a more open past.
AI transparency falls into two main areas: inputs and models. The large-scale language models that underpin generative AI, such as OpenAI's ChatGPT and Google's Gemini, are trained by trawling the internet to learn from analyzing “datasets” that vary from Reddit forums to Picasso paintings. In the early days of AI, researchers often published their training data in scientific journals, allowing others to assess the quality of the inputs and diagnose flaws.
Currently, major companies tend not to release details of their data to protect themselves from copyright infringement lawsuits and secure a competitive advantage, making it difficult to evaluate the veracity of AI-generated answers, and leaving writers, actors, and other creators unable to know whether their privacy or intellectual property is being knowingly violated.
The models themselves also lack transparency. How a model interprets inputs and generates language depends on its design. AI companies tend to consider a model's architecture as its “secret sauce.” The originality of OpenAI's GPT-4 and Meta's Llama rests on the quality of their computations. AI researchers once published papers on their designs, but the race for market share has put an end to such disclosures. However, without understanding how a model works, it is difficult to evaluate the AI's outputs, limitations, and biases.
This lack of transparency makes it difficult for the public and regulators to assess AI safety and prevent potential harm. This situation is even more concerning after Jan Reike, who led OpenAI's push for super-powerful AI tools, left the company this month, alleging that the company's leaders prioritized a “flashy product” over safety. The company claims it can regulate its own products, but the new security committee will report to those very same leaders.
Governments are beginning to lay the groundwork for AI regulation, through the Bletchley Park conference last year, President Joe Biden's executive order on AI, and the EU's AI law. While these steps are welcome, they focus on guardrails and “safety testing” rather than full transparency. The reality is that most AI experts work for companies, and the technology is developing too quickly for regular safety testing to be sufficient. Regulators should demand transparency of models and inputs, and experts from these companies should work with regulators.
AI has the potential to change the world for the better, perhaps more powerfully and faster than the Internet revolution. Companies may argue that transparency requirements will slow innovation and blunt their competitiveness, but the recent history of AI shows that this is not the case. These technologies have advanced on the back of collaboration and joint research. A return to these standards will increase societal trust and enable faster, more secure innovation.
