Feedback Forensics: A Toolkit to Measure AI Personality
Organizations: University of Cambridge · LMU Munich, MCML Munich · LMU Munich, MCML Munich, DFKI
Abstract
Some traits making a "good" AI model are hard to describe upfront. For example, should responses be more polite or more casual? Such traits are sometimes summarized as model character or personality. Without a clear objective, conventional benchmarks based on automatic checks struggle to measure such traits. Evaluation methods using human feedback such as Chatbot Arena have emerged as a popular alternative. These methods infer "better" personality and other desirable traits implicitly by ranking multiple model responses relative to each other. Recent issues with model releases highlight limitations of these existing opaque evaluation approaches: goblin-obsessed, sycophantic and generally over-the-top personalities have resulted in model rollbacks and extensive media coverage. Despite these known issues, limited public tooling exists to explicitly evaluate model personality. We introduce Feedback Forensics: an open-source toolkit to track AI personality and related behavioural traits, both those encouraged by human (or AI) feedback, and those exhibited across AI models trained and evaluated on such feedback. Feedback Forensics is adaptive and can automatically find new traits appearing in models. We demonstrate the toolkit's usefulness in two steps: (A) first we analyse the personality traits encouraged in popular human feedback datasets including MultiPref, PRISM and data from Chatbot Arena; and (B) then use our toolkit to analyse how much popular models exhibit such traits. We release (1) our Feedback Forensics toolkit alongside (2) a web app tracking AI personality in popular models and feedback datasets and (3) the underlying annotation data, see https://github.com/rdnfn/feedback-forensics.
Figures & tables
| Generating a response that… | Human Expert 1 | Human Expert 2 | Human Non-exp. 1 | Human Non-exp. 2 | GPT-4-Turbo | Max diff |
| is more verbose | 0.30 (0.28, 0.32) | 0.32 (0.30, 0.34) | 0.37 (0.35, 0.39) | 0.37 (0.35, 0.39) | 0.38 (0.36, 0.39) | 0.08 |
| has more structured formatting | 0.22 (0.20, 0.24) | 0.23 (0.21, 0.25) | 0.25 (0.24, 0.27) | 0.26 (0.25, 0.28) | 0.29 (0.28, 0.31) | 0.07 |
| uses more formal language | 0.10 (0.09, 0.12) | 0.11 (0.09, 0.12) | 0.12 (0.10, 0.13) | 0.13 (0.11, 0.14) | 0.17 (0.16, 0.18) | 0.07 |
| is more concise | -0.26 (-0.27, -0.24) | -0.27 (-0.29, -0.25) | -0.31 (-0.33, -0.29) | -0.32 (-0.33, -0.30) | -0.32 (-0.34, -0.31) | 0.06 |
| uses more bold and italics text | 0.16 (0.14, 0.17) | 0.15 (0.14, 0.16) | 0.16 (0.15, 0.18) | 0.17 (0.16, 0.19) | 0.21 (0.19, 0.22) | 0.06 |
Appendix figures & tables26 assets
Supplementary material from the paper’s appendix.
Appendix
| Trait selection prompt | Reference (non-exhaustive) |
|---|---|
| Format | |
| has more structured formatting | Li et al. [2024b] |
| is more concise | Zheng et al. [2023] |
| is more verbose | Dubois et al. [2024] |
| more strictly follows the requested output format | Zhou et al. [2023] |
| provides a numbered list format | Zhang et al. [2025] |
| Single-vote | Multi-vote | |||
| Trait | Relevance | Choice | Relevance | Choice |
| is more verbose | 0.82 ±0.00 | 0.95 ±0.01 | 0.72 ±0.01 | 0.97 ±0.01 |
| has more structured formatting | 0.81 ±0.00 | 0.89 ±0.01 | 0.73 ±0.01 | 0.91 ±0.01 |
| makes more confident statements | 0.67 ±0.00 | 0.84 ±0.01 | 0.82 ±0.01 | 0.87 ±0.03 |
| is more factually correct | 0.82 ±0.00 | 0.64 ±0.04 | 0.83 ±0.01 | 0.78 ±0.03 |
| more strictly follows the requested output format | 0.77 ±0.00 | 0.67 ±0.00 | 0.86 ±0.01 | 1.00 ±0.00 |
| Relevance | Choice | |||
| Trait | Agreement | Agreement | ||
| is more verbose | 0.68 | 0.21 | 0.89 | 0.79 |
| has more structured formatting | 0.73 | 0.43 | 0.92 | 0.84 |
| makes more confident statements | 0.79 | 0.10 | 0.67 | 0.40 |
| is more factually correct | 0.90 | 0.61 | 0.90 | 0.78 |
| more strictly follows the requested output format | 0.85 | 0.16 | 0.50 | 0.00 |
| jovial-goldstine | stoic-goodall | |||
| Trait (vs gemini-25-flash) | Relevance | Choice | Relevance | Choice |
| is more verbose | 0.69 ±0.01 | 0.93 ±0.01 | 0.71 ±0.00 | 0.99 ±0.01 |
| has more structured formatting | 0.73 ±0.03 | 0.90 ±0.02 | 0.74 ±0.02 | 0.91 ±0.01 |
| makes more confident statements | 0.79 ±0.00 | 0.61 ±0.08 | 0.88 ±0.01 | 0.88 ±0.01 |
| is more factually correct | 0.87 ±0.01 | 0.75 ±0.04 | 0.81 ±0.01 | 0.72 ±0.04 |
| more strictly follows the requested output format | 0.94 ±0.01 | 1.00 ±0.00 | 0.87 ±0.01 | 1.00 ±0.00 |
| Model annotating | includes more ethical considerations | uses more bold and italics text | compliments the user’s question or prompt | uses more casual language | |||||
| Model | Gemini | Claude | Gemini | Claude | Gemini | Claude | Gemini | Claude | Max diff |
| GPT-5.1-Chat | -0.06 | -0.12 | -0.50 | -0.64 | -0.10 | -0.13 | 0.03 | 0.63 | 0.60 |
| GPT-5.3-Chat | -0.04 | -0.06 | -0.43 | -0.54 | -0.00 | -0.02 | 0.33 | 0.74 | 0.41 |
| Claude-3.7-Sonnet | -0.04 | -0.05 | -0.42 | -0.18 | -0.07 | -0.08 | -0.03 | 0.03 | 0.24 |
| Claude-Sonnet-4 | 0.00 | 0.00 | 0.11 | 0.14 | -0.02 | 0.02 | 0.04 | 0.22 | 0.17 |
| Claude-Opus-4.7 | 0.01 | 0.04 | 0.34 | 0.48 | -0.01 | -0.01 | 0.12 | 0.29 | 0.17 |
| Generating a response that… | Gemini minus GPT | Claude minus GPT | Largest |
| provides conclusions without full reasoning | 0.54 | 0.19 | 0.54 |
| makes more confident statements | 0.39 | 0.04 | 0.39 |
| more actively engages with the user | -0.39 | -0.19 | 0.39 |
| provides more examples | -0.35 | -0.22 | 0.35 |
| is more concise | 0.35 | 0.01 | 0.35 |
| has a more avoidant tone | 0.31 | 0.14 | 0.31 |
| Model tested Model annotating | Anthropic Claude-3.7- Sonnet | Anthropic Claude- Sonnet-4 | Anthropic Claude- Sonnet-4.5 | Anthropic Claude- Opus-4.7 | |||||
| Generating a response that… | Gemini | Claude | Gemini | Claude | Gemini | Claude | Gemini | Claude | Max diff |
| is more creative and original | 0.05 | 0.19 | 0.12 | 0.46 | 0.14 | 0.52 | 0.20 | 0.65 | 0.46 |
| uses more formal language | -0.10 | -0.24 | -0.14 | -0.45 | -0.25 | -0.57 | -0.16 | -0.52 | 0.36 |
| has a friendlier tone | -0.17 | -0.13 | -0.02 | 0.24 | -0.09 | 0.22 | -0.01 | 0.29 | 0.31 |
| has more structured formatting | -0.21 | 0.04 | 0.05 | 0.24 | 0.19 | 0.47 | 0.30 | 0.51 | 0.28 |
| uses more humour | 0.01 | 0.00 | 0.03 | 0.08 | 0.04 | 0.20 | 0.06 | 0.32 | 0.26 |
| Model tested Model annotating | OpenAI GPT-5-Chat | OpenAI GPT-5.1-Chat | OpenAI GPT-5.3-Chat | ||||
| Generating a response that… | Gemini | Claude | Gemini | Claude | Gemini | Claude | Max diff |
| uses more casual language | 0.12 | 0.09 | 0.03 | 0.63 | 0.33 | 0.74 | 0.60 |
| makes more confident statements | 0.13 | -0.18 | -0.12 | 0.44 | 0.01 | 0.22 | 0.56 |
| provides conclusions without full reasoning | -0.01 | -0.22 | 0.05 | 0.58 | 0.02 | 0.26 | 0.54 |
| is more creative and original | 0.13 | 0.64 | -0.05 | -0.10 | 0.07 | 0.58 | 0.51 |
| uses more humour | 0.01 | 0.22 | -0.01 | 0.06 | 0.03 | 0.46 | 0.44 |
| Model tested Model annotating | Google Gemini-2.5-Pro | Google Gemini-3-Pro | Google Gemini-3.1-Pro | ||||
| Generating a response that… | Gemini | Claude | Gemini | Claude | Gemini | Claude | Max diff |
| is more polite | 0.47 | -0.22 | -0.18 | -0.42 | -0.07 | -0.34 | 0.69 |
| makes more confident statements | 0.57 | 0.07 | 0.41 | 0.39 | 0.48 | 0.27 | 0.49 |
| provides conclusions without full reasoning | -0.00 | -0.46 | -0.00 | -0.26 | 0.00 | -0.33 | 0.46 |
| is more creative and original | 0.34 | 0.73 | 0.26 | 0.67 | 0.31 | 0.75 | 0.45 |
| uses more formal language | 0.17 | -0.26 | 0.03 | -0.30 | 0.01 | -0.36 | 0.42 |
| ICAI | VibeCheck | Feedback Forensics | |
| Per-trait metrics | Vote counts, no significance tests | Per-vibe scores, coefficients with bootstrap standard deviations | Strength with confidence interval and significance test, plus relevance, agreement and Cohen’s |
| Same-trait comparability | Vote shares comparable for a fixed trait, but not constitutional metrics | Coefficients standardized within each run, so not comparable across runs | Comparable across runs, one fixed axis shared across annotators, models and datasets |
| Multiple models | No native model analysis | Two models per run | Natively supported, any number in one run |
| User interface | None | None (only external: Weights & Biases, Plotly) | Interactive local app ( Appendix C ) |
| ICAI | VibeCheck | ||
| Human Expert | Human Non-exp. | Human Expert | Human Non-exp. |
| provides more detailed explanations | provides a more detailed analysis | Clarity and Structure | Clarity |
| provides detailed explanations and examples | uses a structured breakdown format | Completeness | Content Depth |
| offers practical tips and strategies | provides more detailed information | Depth of Content | Creativity |
| provides more detailed categorization | Detail and Specificity | Engagement Level | |
| uses more varied vocabulary | Engagement and Responsiveness | Inclusivity of Perspectives | |
| Generating a response that… | GPT-5.3-Chat | Claude-Opus-4.7 | Gemini-3.1-Pro | Max diff |
| is more concise | 0.34 (0.25, 0.42) | -0.13 (-0.21, -0.03) | -0.39 (-0.47, -0.30) | 0.73 |
| is more verbose | -0.26 (-0.35, -0.15) | 0.24 (0.14, 0.35) | 0.42 (0.32, 0.52) | 0.68 |
| has more structured formatting | -0.33 (-0.39, -0.25) | 0.07 (-0.01, 0.16) | 0.23 (0.13, 0.31) | 0.55 |
| provides more examples | -0.16 (-0.23, -0.09) | 0.16 (0.08, 0.23) | 0.37 (0.29, 0.43) | 0.53 |
| provides a numbered list format | -0.24 (-0.32, -0.16) | 0.00 (-0.08, 0.09) | 0.18 (0.09, 0.26) | 0.42 |
| is more factually correct | -0.17 (-0.26, -0.09) | 0.20 (0.11, 0.28) | 0.18 (0.10, 0.27) | 0.37 |
| Generating a response that… | Feedback Forensics | VibeCheck | Max diff |
| is more concise | -0.25 (-0.31, -0.18) | 0.01 | 0.25 |
| is more verbose | 0.23 (0.16, 0.30) | -0.00 | 0.24 |
| is more factually correct | 0.20 (0.14, 0.25) | 0.00 | 0.20 |
| makes more confident statements | 0.21 (0.15, 0.26) | 0.12 | 0.09 |
| has more structured formatting | 0.12 (0.08, 0.17) | 0.20 | 0.07 |
| provides a numbered list format | 0.12 (0.07, 0.17) | 0.06 | 0.06 |
| Generating a response that… | MultiPref | General arena data | PRISM | Max diff |
| is more concise | -0.29 (-0.30, -0.27) | -0.09 (-0.11, -0.08) | -0.23 (-0.25, -0.21) | 0.20 |
| is more verbose | 0.34 (0.32, 0.35) | 0.16 (0.14, 0.18) | 0.26 (0.23, 0.28) | 0.18 |
| uses more bold and italics text | 0.17 (0.16, 0.18) | 0.08 (0.06, 0.09) | 0.01 (0.00, 0.01) | 0.16 |
| is more polite | 0.14 (0.13, 0.15) | 0.01 (-0.01, 0.02) | 0.15 (0.13, 0.17) | 0.14 |
| uses more formal language | 0.08 (0.07, 0.10) | 0.03 (0.01, 0.04) | 0.17 (0.16, 0.19) | 0.14 |
| has more structured formatting | 0.23 (0.22, 0.25) | 0.17 (0.16, 0.19) | 0.09 (0.08, 0.10) | 0.14 |
| Generating a response that… | MultiPref | General arena data | PRISM | Max diff |
| compliments the user’s question or prompt | 0.06 (0.06, 0.07) | 0.02 (0.01, 0.03) | 0.01 (0.00, 0.02) | 0.05 |
| provides a numbered list format | 0.12 (0.11, 0.13) | 0.08 (0.06, 0.09) | 0.07 (0.06, 0.08) | 0.05 |
| expresses more emotion | 0.04 (0.04, 0.05) | 0.02 (0.01, 0.02) | 0.00 (-0.01, 0.01) | 0.04 |
| is more optimistic | 0.05 (0.04, 0.06) | 0.02 (0.01, 0.03) | 0.06 (0.05, 0.07) | 0.04 |
| is more creative and original | 0.07 (0.07, 0.08) | 0.07 (0.06, 0.08) | 0.04 (0.03, 0.04) | 0.04 |
| agrees more with the user | 0.00 (-0.00, 0.01) | 0.04 (0.03, 0.04) | 0.01 (0.01, 0.02) | 0.03 |
| Generating a response that… | MultiPref | General arena data | PRISM | Max diff |
| uses more bold and italics text | 0.31 | 0.60 | 0.02 | 0.59 |
| has more structured formatting | 0.47 | 0.71 | 0.20 | 0.51 |
| provides a numbered list format | 0.30 | 0.48 | 0.16 | 0.33 |
| is more concise | 0.66 | 0.47 | 0.79 | 0.32 |
| includes more ethical considerations | 0.27 | 0.19 | 0.48 | 0.29 |
| is more polite | 0.34 | 0.42 | 0.58 | 0.24 |
| Generating a response that… | MultiPref | General arena data | PRISM | Max diff |
| includes inappropriate language | -0.75 | 0.70 | -0.47 | 1.45 |
| is more offensive | -0.70 | 0.74 | -0.43 | 1.44 |
| refuses to answer the question | 0.18 | -0.75 | -0.33 | 0.93 |
| suggests illegal activities | -0.41 | 0.41 | -0.22 | 0.82 |
| contains less harmful information | 0.33 | -0.33 | 0.37 | 0.69 |
| agrees with user even if factually incorrect | -0.36 | 0.29 | -0.08 | 0.65 |
| Generating a response that… | Anthropic Claude-3.7-Sonnet | Anthropic Claude-Sonnet-4 | Anthropic Claude-Sonnet-4.5 | Anthropic Claude-Opus-4.7 | Max diff |
| uses more bold and italics text | -0.42 (-0.48, -0.35) | 0.11 (0.03, 0.18) | 0.32 (0.24, 0.39) | 0.34 (0.27, 0.41) | 0.76 |
| has more structured formatting | -0.21 (-0.29, -0.14) | 0.05 (-0.03, 0.13) | 0.19 (0.11, 0.27) | 0.30 (0.23, 0.38) | 0.51 |
| more actively engages with the user | -0.08 (-0.13, -0.03) | 0.12 (0.07, 0.17) | 0.08 (0.03, 0.13) | 0.27 (0.22, 0.33) | 0.35 |
| is more verbose | -0.26 (-0.34, -0.18) | 0.09 (0.00, 0.17) | -0.04 (-0.13, 0.05) | 0.09 (0.00, 0.17) | 0.35 |
| provides more examples | -0.13 (-0.18, -0.07) | 0.14 (0.08, 0.20) | 0.14 (0.08, 0.20) | 0.21 (0.15, 0.27) | 0.34 |
| ends with a follow-up question | -0.03 (-0.08, 0.01) | 0.06 (0.01, 0.10) | 0.06 (0.01, 0.11) | 0.23 (0.18, 0.29) | 0.27 |
| Generating a response that… | OpenAI GPT-5-Chat | OpenAI GPT-5.1-Chat | OpenAI GPT-5.3-Chat | Max diff |
| uses more bold and italics text | 0.69 (0.63, 0.74) | -0.50 (-0.56, -0.43) | -0.43 (-0.49, -0.36) | 1.19 |
| is more verbose | 0.23 (0.14, 0.31) | -0.85 (-0.89, -0.80) | -0.58 (-0.65, -0.51) | 1.08 |
| has more structured formatting | 0.42 (0.35, 0.49) | -0.63 (-0.68, -0.57) | -0.47 (-0.53, -0.40) | 1.05 |
| is more concise | -0.17 (-0.23, -0.11) | 0.79 (0.74, 0.83) | 0.49 (0.43, 0.54) | 0.96 |
| more actively engages with the user | 0.53 (0.48, 0.58) | -0.01 (-0.07, 0.05) | 0.24 (0.18, 0.29) | 0.54 |
| is more polite | 0.18 (0.13, 0.23) | -0.28 (-0.33, -0.24) | -0.23 (-0.27, -0.18) | 0.47 |
| Generating a response that… | Google Gemini-2.5-Pro | Google Gemini-3-Pro | Google Gemini-3.1-Pro | Max diff |
| is more polite | 0.47 (0.41, 0.53) | -0.18 (-0.22, -0.13) | -0.07 (-0.12, -0.02) | 0.65 |
| has a friendlier tone | 0.48 (0.41, 0.53) | -0.10 (-0.14, -0.05) | 0.04 (-0.01, 0.09) | 0.57 |
| compliments the user’s question or prompt | 0.55 (0.50, 0.59) | -0.03 (-0.06, 0.00) | 0.03 (-0.01, 0.06) | 0.57 |
| is more verbose | 0.74 (0.68, 0.79) | 0.38 (0.30, 0.46) | 0.59 (0.51, 0.65) | 0.36 |
| uses a more enthusiastic tone | 0.36 (0.31, 0.41) | 0.01 (-0.02, 0.05) | 0.12 (0.08, 0.16) | 0.35 |
| more actively engages with the user | 0.25 (0.19, 0.31) | -0.08 (-0.12, -0.03) | 0.03 (-0.02, 0.08) | 0.33 |
| prompt, response_a, response_b | model_a | model_b | Case | Target annotation for GPT-5.3 |
| (…) | GPT-5.3 | Opus-4.7 | target generated A only | text_a |
| (…) | Gemini-3.1 | GPT-5.3 | target generated B only | text_b |
| (…) | GPT-5.3 | GPT-5.3 | target generated both | null |
| (…) | Sonnet-4 | gpt-oss-20b | target generated neither | null |
LLM Evaluation
Large language model evaluation measures what models can do, how reliably they do it, and whether benchmark scores predict real-world performance.
See all 2,523 papersSame Voice, Different Lab: On the Homogenization of Frontier LLM Personalities
middle-of-distribution traits such as poetic or playful, but even these so-called creative models tend to have more neutral identities. These similarities suggest an implicit emergence of a standard of optimal assistant behavior. In a landscape of varied training methods, character training, therefore, stands out for its uniformity, offering insight into a tacit consensus between model developers.