State of the Art: Ideogram 1.0

When ChatGPT started "breathing" during voice interactions this year, this design choice seemed innocuous. However, it highlights a concern around increasingly human-like AI features being implemented without proper evaluation of their necessity or impact. The tragic case of Character.AI -- where anthropomorphic design choices allegedly contributed to a teenager's death -- demonstrates the urgent need for systematic analysis of these features. This article proposes a two-step framework for evaluating anthropomorphic AI design: first assessing both social and technical necessity, then implementing appropriate guidelines and safeguards (some of which are suggested herein). While some applications (like accessibility tools) may well justify human-like features, others (like companionship AI) tread on ethically difficult territory. As AI becomes more sophisticated, we need clear criteria for determining when anthropomorphic design serves legitimate social purposes. An LLM doesn't really need to breathe, and perhaps we ought to think much harder about how we would like human-to-human relationships to look like in the future before delving too far into enabling human-to-AI facsimiles.

Did you know we're all mind readers? Humans develop what’s called “theory of mind” (ToM) around the age of four, enabling us to hold views and beliefs about our own mental states and the mental states of others. It’s one of those things that, well, makes us human. However, a recent paper by Strachan et al. explores this capability in the context of AI -- a finding which has some interesting implications. The authors compared the performance of a series of LLMs (GPT-3, GPT-4, and LLaMA2) against around 1,900 humans across a battery of tests designed to explore ToM. They determined that, in every task bar one (designed to test faux pas), LLMs can perform at or above human level. My own cursory testing suggests that GPT-4o may even be able to demonstrate ToM across the board. While we can’t conclude that these models actually have ToM (in a cognitive sense) they certainly seem to behave as if they do. So what does this mean in practice? These LLMs open the door to exciting potential applications across areas such as mental health and decision support making. Ultimately, this paper’s conclusions give us a moment to stop and think about the state of our technology: to be impressed by what we’ve accomplished, and acknowledge the risks that should be considered as we move forward.

On 4 June 2024, employees at OpenAI and Google DeepMind released an open letter imploring their companies to facilitate what they call a right to warn about advanced AI. In the absence of proper government oversight, they call upon their companies to: (1) not enter or enforce any non-disparagement clauses when it comes to risk-related concerns; (2) develop internal, verifiably anonymous, whistleblower processes; (3) support a culture of open criticism which enables disclosure of risk-related concerns; and (4) not retaliate against employees (past or present) who choose to publicly voice risk-related concerns where other processes have failed. In the uncomfortable light of commercial reality, most of these requests appear unrealistic. A demand for robust internal whistleblowing processes, however, is an actionable and arguably best next step -- particularly given the poor state of protection under the current whistleblowing regime in England and Wales. There should be a right to warn about advanced AI. This demand may well be the starting point.

The past few weeks have brought the issue of AI Safety to the fore in two contrasting ways: first, OpenAI launched GPT-4o, a promising new model that will be free for public use. Around the same time, Ilya Sutskever (OpenAI's renowned co-founder and chief scientist) left the company amidst reports that OpenAI had dissolved its long-term AI Safety team. Although the company reiterated its commitment to safe development, questions linger. Second, the UK's Department of Science, Innovation and Technology released its interim International Scientific Report on the Safety of Advanced AI. The report focuses on general-purpose AI and explores various risks as well as some technical risk mitigation techniques. It also leaves plenty unanswered. With so much uncertainty, it's imperative we seek to resolve the most practically-relevant question as soon as possible: who's responsible for AI Safety?

With this week's condensed version of the Sunday Sweep, we focus on a couple of key developments. First, Neuralink's released an incredible video showing its first human patient, Noland Arbaugh, manipulating a cursor on a monitor entirely with his mind. This sort of technology demands serious consideration both for the tremendous benefits it can bring as well as the potential ethical issues that come hand-in-hand with it. On the regulatory side, representatives of the US Congress proposed the Protecting Consumers from Deceptive AI Act -- federal legislation designed to protect the public from deepfakes. The White House Office of Management and Budget issued its first major AI policy, and a pair of additional government departments released their own AI reports (including a Treasury report on managing AI cybersecurity risk in the financial services sector).