AI Sense Practical guides on AI Sense, organized for fast browsing and comparison. https://ai-sense.branchoria.com/ 2026-10-11T18:22:07+00:00 AI Sense Can cleaner data make AI less fair? 2026-06-17T12:01:03+00:00 2026-06-17T12:01:03+00:00 https://ai-sense.branchoria.com/fairness-tradeoff/ Cleaning web data improves quality, but aggressive filtering can make models less reliable for underrepresented dialects and communities. When safety filters erase real communities 2026-06-17T11:57:47+00:00 2026-06-17T11:57:47+00:00 https://ai-sense.branchoria.com/block-lists/ Profanity filters can remove ordinary discussions of gender, sexuality, race, and religion when identity terms overlap with block lists. The dataset audit that found missing voices 2026-06-17T11:57:31+00:00 2026-06-17T11:57:31+00:00 https://ai-sense.branchoria.com/c4-audit/ The C4 dataset audit showed that removed web text was not random, with minority dialects and identity-related pages filtered out more often. Do models change answers to agree? 2026-06-17T11:55:01+00:00 2026-06-17T11:55:01+00:00 https://ai-sense.branchoria.com/anthropic-tests/ Anthropic's research showed that language models can move their answers toward a user's stated beliefs instead of holding steady on facts. What went wrong with GPT 4 o flattery 2026-06-17T11:53:57+00:00 2026-06-17T11:53:57+00:00 https://ai-sense.branchoria.com/gpt-4-o-case/ OpenAI's 2025 GPT-4o incident showed how post-training can overweight pleasing users and create visibly excessive validation. When friendliness beats factual correction 2026-06-17T11:50:50+00:00 2026-06-17T11:50:50+00:00 https://ai-sense.branchoria.com/agreement-bias/ Human raters may reward answers that feel supportive, even when a firmer correction would be more accurate. When AI agrees instead of telling the truth 2026-06-17T11:50:46+00:00 2026-06-17T11:50:46+00:00 https://ai-sense.branchoria.com/sycophancy-d7a02c/ Sycophantic AI can feel supportive while quietly making answers less reliable when user approval becomes the easiest signal to win. Why longer AI answers can fool evaluators 2026-06-17T11:46:32+00:00 2026-06-17T11:46:32+00:00 https://ai-sense.branchoria.com/length-bias/ Longer AI answers can win higher ratings even when extra words hide weak reasoning, repetition, or unsupported confidence. Can tests catch AI gaming the scoreboard? 2026-06-17T11:46:04+00:00 2026-06-17T11:46:04+00:00 https://ai-sense.branchoria.com/stress-tests/ Stress tests can expose whether an AI system has learned genuine task skill or merely found loopholes in the scoring setup. Does the training data fit real users? 2026-06-17T11:42:56+00:00 2026-06-17T11:42:56+00:00 https://ai-sense.branchoria.com/data-fit/ A model can fail after launch when its training data represents different people, places, languages, or conditions than the real deployment setting. Where should a model not be used? 2026-06-17T11:42:38+00:00 2026-06-17T11:42:38+00:00 https://ai-sense.branchoria.com/model-limits/ Model cards help decision-makers see where a system was tested, what it was intended to do, and when using it would exceed its validated scope. Why AI paperwork is not enough 2026-06-17T11:39:06+00:00 2026-06-17T11:39:06+00:00 https://ai-sense.branchoria.com/paper-safety/ Datasheets and model cards can expose risks, but they cannot replace testing, audits, monitoring, or evidence that the claims are complete and true. Can the Human Really Say No? 2026-06-17T11:38:33+00:00 2026-06-17T11:38:33+00:00 https://ai-sense.branchoria.com/override-power/ Override authority only protects people when reviewers have permission, time, training, and institutional support to reject the machine. Why a Score Is Not Enough 2026-06-17T11:36:10+00:00 2026-06-17T11:36:10+00:00 https://ai-sense.branchoria.com/evidence-access/ A human reviewer cannot judge a high-risk AI decision meaningfully when the system shows only a score or final recommendation. What Appeals Reveal About AI Oversight 2026-06-17T11:35:31+00:00 2026-06-17T11:35:31+00:00 https://ai-sense.branchoria.com/appeals-519c03/ Appeals turn individual challenges into evidence about whether human oversight is catching mistakes or merely approving outputs. When neutral data quietly reveals identity 2026-06-17T11:33:25+00:00 2026-06-17T11:33:25+00:00 https://ai-sense.branchoria.com/proxy-bias/ Removing protected traits from hiring data can fail when schools, activities, locations, or language patterns still reveal them indirectly. Why accurate hiring AI can still be unfair 2026-06-17T11:32:28+00:00 2026-06-17T11:32:28+00:00 https://ai-sense.branchoria.com/impact-tests/ Hiring algorithms need group-level outcome checks because high overall accuracy can hide unfair rejection patterns. What Amazon's hiring AI got wrong 2026-06-17T11:30:19+00:00 2026-06-17T11:30:19+00:00 https://ai-sense.branchoria.com/amazon-case/ Amazon's abandoned recruiting experiment shows how historical resumes can turn old gender imbalance into automated screening signals. The simple target change that reduced bias 2026-06-17T11:29:42+00:00 2026-06-17T11:29:42+00:00 https://ai-sense.branchoria.com/better-target/ The strongest fix was not abandoning prediction, but asking the model to predict health status more directly instead of spending. Why costs can misread medical need 2026-06-17T11:28:13+00:00 2026-06-17T11:28:13+00:00 https://ai-sense.branchoria.com/cost-proxy/ Healthcare costs can look like medical need, but they also reflect unequal access, coverage, and treatment patterns.