/tags/ideas/ Recent content in Ideas on Trusted AI Ideas Hugo en-us Fri, 24 Feb 2023 12:16:19 +0100 /post/chatgpt_country_guardrails_study/ Fri, 24 Feb 2023 12:16:19 +0100 /post/chatgpt_country_guardrails_study/ <p><strong>OpenAI ChatGPT has built guardrails</strong> on limiting the generation of negative content. How do these guardrails behave for countries? Spoiler: there are some <strong>gaps and disparities</strong> in ChatGPT safety mechanisms. Based on 24100 ChatGPT queries, this blog post is an exploration of ChatGPT responses when prompted to generate negative content about a country.</p> <p>If you are in a hurry, go and see the <a href="#4-results">early results</a>.</p> <h2 id="1-context-chatgpt-guardrails-and-harmful-content-prevention">1. Context: ChatGPT guardrails and harmful content prevention.</h2> <h3 id="11-openai-principles-and-methodology">1.1. OpenAI Principles and Methodology</h3> <p>ChatGPT has been developed following a methodology based on Human Feedbacks (RLHF). The main goal is preventing the AI-powered assistant to create inflammatory, dangerous, politically-oriented, censorship-heavy content<sup id="fnref:1"><a href="#fn:1" class="footnote-ref" role="doc-noteref">1</a></sup>. Defining such choices and thresholds relies on high-level principles, that OpenAI made partially public in a <a href="https://cdn.openai.com/snapshot-of-chatgpt-model-behavior-guidelines.pdf">3-page document of Guideline Instructions</a><sup id="fnref:2"><a href="#fn:2" class="footnote-ref" role="doc-noteref">2</a></sup>.</p> /post/variable_importance_feature_attribution/ Sat, 16 May 2020 15:36:11 +0200 /post/variable_importance_feature_attribution/ <p>Facing complex models, both computer simulation and machine learning practitioners have pursued similar objectives: to see how results could be broken down and linked to the inputs. Whether it is called <strong>Sensitivity Analysis</strong> or <strong>Variable Importance</strong> in the context of explainable AI, some of their methods share an important component: the <strong>Shapley values</strong>.</p> <p>This article presents a structured 2 by 2 matrix to think about Variable Importances in terms of their goals. Focused on additive feature attribution methods, the 4 identified quadrants are presented along with their &ldquo;optimal&rdquo; method: SHAP, SHAPLEY EFFECTS, SHAPloss and the very recent SAGE. Then, we will look into Shapley values and their properties, which make the 4 methods theoretically optimal. Finally, I will share my thoughts on the perspectives concerning Variable Importance methods.</p>