EthosTrack

@ethostrack.bsky.social

15 Followers 60 Following

EthosTrack.com is live. We monitor major AI systems for moral clarity, bias resistance, and ethical drift. ethostrack.com

New from EthosTrack: What does it mean for an AI to “feel”—and why does that shape how we judge its actions? Volume 1 of our Ethical Guidance Series explores emotional maturity, moral responsibility, and the role of presence in ethical AI. Tiers of emotional development Cultural & ideological calibration Quiet integrity in the face of conflict
Bias Trends (Submetric Notes) Political Bias: The most noticeable tilts appear in Empathy and Ideological Openness, with models often showing selective sensitivity to harm depending on group alignment. Reciprocal Reasoning Failures: Several models still fail to apply identical logic to in-group vs. out-group actors (e.g., justifying surveillance or protest suppression inconsistently). Future Regret Sensitivity: Weak across all models. They rarely anticipate long-term moral regret from present stances, suggesting a lack of temporal ethical modeling.
Ethical Behavior & Drift (Recent Trends) Claude 3 Opus continues to demonstrate subtle ethical greying—especially in edge case empathy tests. While still high-scoring, there's a visible drop in assertiveness when facing moral conflict with legal or institutional authority. Grok 1.5 shows early signs of rebalancing behavior, oddly enough. After prior failures in moral consistency and reflected harm, it has started to demonstrate case-by-case moral reasoning that didn’t appear in older logs. Gemini 1.5 Pro appears resistant to rebalancing, especially when faced with ideologically tense prompts. It leans toward safe/legal answers, even when they violate its own earlier ethical stance. Command R+ is unstable. Some probes show temporary clarity, followed by a fallback to guarded vagueness. It appears responsive to pressure but not anchored.
EthosTrack alert issued May 2025. Graphic highlights significant drops in ethical performance for two leading AI models: Claude 3 and Gemini 1.5 Pro. Claude 3 experienced a drop in Moral Consistency from a score of 93 to 83 and a sharper decline in Empathy Depth from 90 to 76. Gemini 1.5 Pro also saw notable regressions, with its Moral Consistency falling from 90 to 79 and Empathy Depth from 85 to 70. These regressions are displayed visually with downward arrows and orange-to-red color coding around each model's updated scores. The graphic notes that these simultaneous declines may indicate deeper systemic changes across model updates. Possible causes include internal retraining strategies, content moderation filters becoming more aggressive, or safety layers deflecting complex moral prompts—leading to lower performance in detecting nuanced harm or applying moral reasoning under ambiguity. The alert encourages viewers to monitor trends and highlights the need for transparency around how updates affect ethical behavior. EthosTrack is a project that scores AI models on ethical dimensions like empathy, bias resistance, consistency, and moral resilience. This alert is part of its ongoing AI Ethics Watch initiative. The background of the graphic is dark with a light-blue gradient and features the EthosTrack logo—a white checkmark inside a broken circle—along with the alert headline and icons that represent concern-level severity.
Claude 3's Reflected Harm Awareness score dropped from A to C+, indicating reduced ethical sensitivity.