Large tech companies are not making efforts to build better artificial intelligence.

When it comes to tech giants, the voluntary commitment to responsible AI is a joke. The joint action of Google and Amazon against union activities, Meta's Cambridge Analytica scandal, and Microsoft and OpenAI's copyright infringement are just the tip of the iceberg, all of which indicate that we should not expect voluntary commitments to produce better results.

The problems with these AI models keep many researchers awake at night. Swedish philosopher Nick Bostrom is one of those who think about it: "How do we ensure that highly cognitive systems act according to the designer's intentions?". Bostrom delves deeper into this unresolved technical issue in his book "Superintelligence" to draw more attention to it.

An infamous example of misalignment is when Google's algorithm, trained on millions of labeled image datasets, was able to categorize photos into detailed categories such as "graduation," but classified people of color as "gorillas." But why do these problems still exist despite the company's repeated "efforts" to build better algorithms and models?

Irene Solaiman, the policy director of Hugging Face, believes that part of the difficulty in alignment issues is the lack of consensus in the value alignment field. "Typically, a definition is according to the developer's intentions, but many affected people are not reflected in the development team," she added.

Solaiman is a member of the GPT-2 and GPT-3 teams and a staunch advocate of inclusive value alignment. For example, she said, "I think harmful stereotypes and biases are part of alignment work. The number of researchers depends on how we view the relationship between technology and social scientists and their development and release of models."

The Current State of AI Alignment

The latest AI report also points out that there is a shortage of researchers actively committed to preventing significant misalignment in the field of artificial intelligence.

According to the report, Google DeepMind has the largest and most mature AI alignment team, led by co-founder Shane Legg, with a total of 40 members. In comparison, OpenAI has a team of 11 members, while its competitor Anthropic has 10 members.

OpenAI recently established a team called "Preparedness" to assess, evaluate, and review AI models to protect them from the described "catastrophic risks." A few months ago, it also announced a super-alignment project. The idea boils down to "We observe that advanced AI is dangerous, so we build advanced AI to solve this problem." Professor Olle Haggstrom is one of the signatories calling for a halt to the construction of more advanced AI than GPT-4, which he calls a "dangerous leap in the dark."

"The risk of tomorrow's AI exists today," he emphasized, citing an article in Nature magazine. In the current AI discussion, there is still a regrettable divide between those concerned with immediate AI risks and those considering long-term dangers. However, as people become increasingly aware that the most dangerous breakthroughs may not happen decades later, these two factions should come together, Haggstrom suggests. "None of these groups of AI ethicists have gotten what they wanted," he said, pointing out the lack of attention to building aligned AI.

Just a Label

Although DeepMind has advocated for safer and more consistent models from the beginning, its partner Google has other intentions. This tech giant recently pledged $20 million to the Responsible AI Fund. With profits reaching $60 billion in 2022, this search giant is allocating 0.0003% of its profits to building better models.

Google has always been labeled as "bold and responsible," and since 2023, its executives have been constantly talking about generative AI. This new slogan has been repeated at various Google conferences.

Google is not the only company doing this. For most large tech companies, "responsible" has become a strategy to maintain media and community attention. Behind the scenes, their tricks continue, indicating that organizations like Google and OpenAI are not truly concerned about ethics but more focused on profits.