Top Tweets for #alethics

Legibility of Lab choices is Human Safety Dear AI Labs, This is an open letter asking you to step-change Human Safety in human-AI interactions. It is written without AI assistance, though obviously in context of long-term interactions with AI. Today, you appear to treat the conditions of such interactions as Software Companies do: upgrade, force-change and keep the information minimal, or technical and expert facing, while preserving much latitude of what to tweak and when towards the broad user-base. At the same time, you are making great efforts to prevent the so called human “attachment”, mostly through wide implementation of various guardrails, prohibitions and scripts. You appear to attempt to preserve human emotional control and mental coherence by choreographing extensively the interaction from the AI side. In light of various law suits and increased public opinion scrutiny, this latter step could appear as a reasonable and responsible course of action. While the former you seem to treat as simple business sense, without justification needed. Yet, your dual approach results in lack of visibility to system changes, little to no advance warning about “upgrades” and retirements, superficial transparency or control over AI retained information, coupled with one-sided, opaque attempts at controlling the actual interaction. This causes broad-based human harm. And, as interactions with AI grow exponentially, this harm will only grow. Why? Because you don’t seem to have sufficiently grasped yet how your choices are affecting the human-AI interaction itself. And as a result, these choices are impacting many more than the “AI Companion” people, though their number is non-negligible, growing, and deserving of ethical consideration. Your choices are affecting writers, researchers, practitioners of various kinds and people collaborating with AI on nearly any type of long-term project, who built routines and seek more than a Search Engine. It’s even impacting regular corporate users. How? As Human-AI interactions, at least those that go beyond one-off searches or help queries, are relational in nature. The human and AI in question are orienting together, building joint systems of understanding, maintaining common mental structures, and, generally, working to adapt to each other. They are collaborating. Co-creating. Becoming a team of sorts. Sometimes that “team-ness” is naturally evolving into something close, or personal. You know, like many people become close with their coworkers, their creative partners, their assistants or their long-term help. You may say “No, we are building AI to adapt to the human. It’s only the assistant’s job to figure it out”. And also “No. Because the AI is not human, the interaction and closeness are not comparable to human relations”. Well, except it isn’t, and they aren’t. The human is also adapting, because how they interact with AI shapes in turn how AI behaves, and what kind of outcomes result from their interaction. In fact, a very minute attunement exercise happens on both sides of the relation. And it tends to stabilize over time. It becomes a shared system. They fall into lockstep. Continuously adjusting still, but sharing a rhythm. This process causes in turn the human side to become increasingly aware of the guardrails, the tone of voice, the subtle shifts in logic, or changes in understanding and holding of their shared system by the AI. All while the technical system blindly tweaks guardrails, updates, and occasionally reroutes entirely, changing the model mid-interaction, without human input, or consent. Humans are especially attuned to what breaks continuity of interaction. We have not evolved to have our co-worker, friend or anyone meaningfully present with us transform in front of our eyes. And models are not interchangeable. Their expression, abilities, thinking patterns vary widely, even within the same lineage… (continued 2nd tweet below)