super exciting job opportunity! 🥳.

I'm presenting our work on the Gaps in the Safety Evaluation of Generative AI today at @AIESConf ! .We survey the state of safety evaluations and find 3 gaps: the modality gap 📊, the coverage gap 📸, and the context gap 🌐. Find out more in the paper:

Join millions who have switched to Grok.

RT @mhtessler: I am really beyond words to be able to share with the world what @bakkermichiel, @summerfieldlab, and I and a truly world-cl….

RT @IasonGabriel: What does it mean for AI to be "too human"? How could such a situation arise? And why does this matter?. Check out this e….

How do we anticipate and prepare for the impacts of anthropomorphic AI on users and society? We explore this question in our paper, now out in @AIESConf.#aies2024:

RT @Arianna_Manzini: In a world where users rely on advanced AI assistants for a range of tasks across various domains, when would user tru….

RT @lujainmibrahim: Most real-world AI applications involve human-model interaction, yet most current safety evaluations do not. In a new p….

[6] There are clear ways for model developers to avoid inadvertently increasing the risk of harm to users, including transparency and disclosure of an AI's status and impact-centred research on user-AI interaction. 🔧 Find out more about paths to mitigations in the paper!.

[5] The more anthropomorphic cues we build into these systems, the more salient misperceptions of human-likeness become. The capacity to shape this design creates a responsibility. 🙋‍♀️.

[4] These risks have become a lot more salient with recent releases of conversational AI systems. ⚠️ Luckily, AI designers have influence over the anthropomorphic cues in a system that increase perceptions of human-likeness.

[3] We’ve known for a long time that seeing human-likeness in AI may lead users to over-trust these systems, or form inappropriate emotional attachments to them, paving the way to harms that impact a user's privacy, safety, and emotional well-being. ❤️‍🩹.

[2] But the dynamic and highly interactive nature of conversational AI is quickly changing our expectations of how likely it is that people will anthropomorphise these systems (treat them as human-like). 🤯.

[1] Seeing human-likeness in technology is nothing new – we've observed this effect in user interactions with simple dialogue systems decades ago (the “Eliza effect”). More recently, people have shown a tendency to "see human" in social robots and digital assistants. 🤖.

Have you been thinking about the implications of anthropomorphic AI quite a bit this week? 🤔 We explore the risks of anthropomorphic AI systems in our Ethics of Advanced AI Assistants report. Key insights in thread 💡

RT @IasonGabriel: More great work from a research team led by our model methodologist and evaluator in chief @weidingerlaura 👏. Here's what….

paper's out! 🥳 honored to have contributed to this thorough analysis of the mechanisms that enable persuasive generative AI – hats off to seliem and sasha for their spectacular leadership!.

A truly monumental effort led by @IasonGabriel , @Arianna_Manzini , and Geoff Keeling. I led the section on anthropomorphism – many thanks to my lovely co-authors Iason, Arianna, @verena_rieser, and @weidingerlaura and to @merrierm and @mhtessler for their thoughtful edits!.

RT @GoogleDeepMind: We’re excited to announce 𝗚𝗲𝗺𝗶𝗻𝗶: @Google’s largest and most capable AI model. Built to be natively multimodal, it can….

RT @GoogleDeepMind: From assisting in healthcare to creating art, generative AI is changing how we live and work. We developed a framework….