AI is Software
Over recent weeks, there has been a flurry of AI announcements that (in my opinion) set a concerning precedent. The release of Meta Muse, OpenAI Dots, and even messaging from Anthropic regarding Claude all take steps to “humanise” AI.
Meta Muse is the most obvious example of this trend, with the introduction of a cute avatar that users are expected to treat as a “friend”, building a personal relationship. We have also seen Anthropic announce that they will ban users from “needless abusive or cruel behaviour” towards Claude.
A primary reason behind these decisions is likely driven by the need to shift public perception regarding AI, especially with the younger generation, acknowledging that today there is widespread skepticism regarding its value.
The concern I have is that by positioning an AI as a living entity (like a human or animal), we create a confusing social and legal precedent. For example, should an AI have rights? Should there be laws and rules that protect these rights, such as the legal right to use reasonable force to defend itself? In addition, in the event an AI executes a task that is considered malicious, who is liable?
In a world that could quickly have more AI agents than humans, working in large swarms to plan and execute tasks, the need to ensure alignment and containment has never been more important.
Therefore, we must position, accept, and treat AI as software. Like other software, AI must be positioned as a technology tool (not a living entity) that facilitates specific processes defined and governed by humans (AI must work for humans to solve the problems we define).
This clear positioning supports the enforcement of important controls, including alignment, containment, observability, and the ability to terminate any AI task without judgment.
It also simplifies liability. For example, if you own the software that executes a malicious task, you (the individual or business) are liable.
It should be noted that liability challenges still exist. For example, when a foundational AI model (e.g., Google Gemini) is used to execute a malicious task, is it the AI model provider or the user triggering the task that is liable?
In short, we must treat AI as software, avoiding the temptation to humanise the software.
As I look across the market today, I am most impressed with Microsoft, who have published the Humanist AI Code of Conduct, alongside specific documents regarding responsible AI.