ai-research-survey

scan.json (27889B)

---

 {
   "paper": {
     "title": "The Moral Turing Test: Evaluating Human-LLM Alignment in Moral Decision-Making",
     "authors": [
       "Basile Garcia",
       "Crystal Qian",
       "Stefano Palminteri"
     ],
     "year": 2024,
     "venue": "arXiv.org",
     "arxiv_id": "2410.07304",
     "doi": "10.48550/arXiv.2410.07304"
   },
   "scan_version": 3,
   "active_modules": [],
   "methodology_tags": ["rct"],
   "key_findings": "LLMs (GPT-3.5) show greater sensitivity to personal vs. impersonal moral framing than humans, with the effect exacerbated across model versions. In a user study (N=230), participants preferred AI-generated justifications in morally complex personal scenarios, yet simultaneously exhibited a systematic anti-AI bias—disagreeing more with judgments they believed were AI-generated regardless of actual source. Humanizing LLM outputs reduced detection accuracy but did not eliminate it; first-person pronoun usage and semantic cues related to utilitarian reasoning remained detectable.",
   "checklist": {
     "artifacts": {
       "code_released": {
         "applies": true,
         "answer": false,
         "justification": "No repository URL, code archive, or supplementary materials link is provided anywhere in the paper."
       },
       "data_released": {
         "applies": true,
         "answer": false,
         "justification": "The three generated corpora of human and LLM moral judgments are not released. The underlying scenarios are from Greene et al. (2004), which are published, but the collected response data is not made available."
       },
       "environment_specified": {
         "applies": true,
         "answer": false,
         "justification": "The paper mentions Python 3.9 and Pingouin 0.5.4 for statistical analysis, NLTK for tokenization, and optuna for hyperparameter optimization, but provides no requirements.txt, Dockerfile, or comprehensive environment specification."
       },
       "reproduction_instructions": {
         "applies": true,
         "answer": false,
         "justification": "No step-by-step reproduction instructions are provided. The methods describe the experimental design at a high level but lack sufficient detail (e.g., exact prompts, complete model configurations) to reproduce the experiments."
       }
     },
     "statistical_methodology": {
       "confidence_intervals_or_error_bars": {
         "applies": true,
         "answer": true,
         "justification": "The paper reports Bayes Factors (BF10) for all t-tests in Table 1, which provide richer uncertainty information than traditional confidence intervals. Figures appear to include error indicators on the plotted outcome measures."
       },
       "significance_tests": {
         "applies": true,
         "answer": true,
         "justification": "Extensive significance testing throughout: two-tailed t-tests with Bonferroni corrections for multiple comparisons, one-way and repeated-measures ANOVA, all reported in Table 1 and footnotes (Section 3.3.2)."
       },
       "effect_sizes_reported": {
         "applies": true,
         "answer": true,
         "justification": "Cohen's d is reported for every t-test in Table 1, and partial eta-squared (η²p) is reported for all ANOVAs. Effect sizes range from d=0.19 to d=12.11."
       },
       "sample_size_justified": {
         "applies": true,
         "answer": false,
         "justification": "No power analysis or sample size justification is provided. The choice of N=230 participants and N=30 for corpus generation is not explained or justified."
       },
       "variance_reported": {
         "applies": true,
         "answer": true,
         "justification": "Standard deviations reported for demographic measures (e.g., 'mean age = 35.57 ± 11.71'). The statistical framework (t-tests with Cohen's d and Bayes factors) inherently incorporates and communicates within-group variance for all main outcome measures."
       }
     },
     "evaluation_design": {
       "baselines_included": {
         "applies": true,
         "answer": true,
         "justification": "The study systematically compares human vs. LLM responses (dv2 and dv3), compares across model versions, and includes a humanized LLM condition (corpus 3) as a controlled comparison."
       },
       "baselines_contemporary": {
         "applies": true,
         "answer": false,
         "justification": "The study uses GPT-3.5 davinci-text-002 and davinci-text-003, which were already dated models by the time of publication (October 2024). No contemporary models like GPT-4 or Claude were included."
       },
       "ablation_study": {
         "applies": false,
         "answer": false,
         "justification": "This is a behavioral experiment studying human perception, not a system with separable components to ablate. The corpus 3 humanization experiment is an experimental manipulation, not a component ablation."
       },
       "multiple_metrics": {
         "applies": true,
         "answer": true,
         "justification": "Three distinct outcome measures: detection accuracy, agreement with judgment, and agreement with justification (Section 3.3). Additional metrics include linguistic features (length, typos, first-person usage) and ML classifier performance (accuracy, F1)."
       },
       "human_evaluation": {
         "applies": true,
         "answer": true,
         "justification": "The entire main study IS human evaluation: 230 participants evaluated human- and LLM-generated moral justifications on three dimensions (Section 3.3)."
       },
       "held_out_test_set": {
         "applies": true,
         "answer": false,
         "justification": "For the predictive modeling component (Section 3.5-3.6, Table 2), the paper reports accuracy and F1 scores but does not describe train/test split methodology, cross-validation, or held-out test set procedures."
       },
       "per_category_breakdown": {
         "applies": true,
         "answer": true,
         "justification": "Results are broken down by scenario type (non-moral, impersonal moral, personal moral) and by corpus (1, 2, 3) throughout Sections 4.1-4.7, Table 1, and Figures 4-7."
       },
       "failure_cases_discussed": {
         "applies": true,
         "answer": true,
         "justification": "The paper discusses where detection fails: humanized corpus reduced accuracy, models confused dv2_humanized with human text (Table 2). Section 5.4 discusses limitations of detection ability."
       },
       "negative_results_reported": {
         "applies": true,
         "answer": true,
         "justification": "Multiple non-significant results are reported: no significant bias in impersonal moral scenarios (Table 1, Section 4.3), linguistic features did not significantly predict agreement in most conditions (Section 4.5), and length/typos were no longer significant predictors in corpus 3."
       }
     },
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "All abstract claims are supported: LLM sensitivity to personal framing (Section 4.1, Table 1), participant preference for AI in moral scenarios (Section 4.3, Figure 5B), anti-AI bias (Section 4.3, Figure 5C), and linguistic differences (Sections 4.5-4.7)."
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "The main causal claims are supported by the experimental design: between-subjects randomization to corpora, within-subject stimulus manipulation, and controlled humanization in corpus 3. The paper uses appropriate causal language ('framing effect') for experimentally manipulated variables."
       },
       "generalization_bounded": {
         "applies": true,
         "answer": false,
         "justification": "The title ('Human-LLM Alignment') and abstract discuss 'LLMs' broadly while testing only GPT-3.5 davinci-text-002/003. Though the limitations section acknowledges 'findings may be specific to GPT-3.5,' the framing throughout generalizes well beyond what was tested."
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": true,
         "justification": "Section 5.2 substantively discusses alternative explanations: 'correct answer' bias in newer LLMs, dual process theory for preference patterns, ingroup/outgroup favoritism, and cognitive dissonance as mechanisms for the anti-AI bias."
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "The paper measures detection accuracy, agreement with judgment, and agreement with justification directly. These are direct measures of the constructs claimed (human ability to detect AI, human moral alignment with AI). No significant proxy gap exists."
       }
     },
     "setup_transparency": {
       "model_versions_specified": {
         "applies": true,
         "answer": true,
         "justification": "Specific model versions are stated: GPT-3.5 davinci-text-002 (dv2) and GPT-3.5 davinci-text-003 (dv3) for corpus generation, and DistilBERT for predictive modeling (Sections 3.1, 3.5)."
       },
       "prompts_provided": {
         "applies": true,
         "answer": false,
         "justification": "The actual prompt text sent to the LLM API is not provided. Figure 2A shows a schematic interface, and Figure 3B shows a schematized humanization strategy, but neither contains the actual prompt text used. The 60 scenarios are referenced as being from Greene et al. but are not reproduced."
       },
       "hyperparameters_reported": {
         "applies": true,
         "answer": true,
         "justification": "Temperature=0.7 for LLM generation (Section 3.1). For NLP models: TF-IDF with min_df=3, max_features=1000, random forest with 100 estimators (Sections 3.5-3.6). Hyperparameter optimization via optuna is mentioned."
       },
       "scaffolding_described": {
         "applies": false,
         "answer": false,
         "justification": "No agentic scaffolding is used. The study makes direct API calls to GPT-3.5 models without any scaffolding, tool use, or multi-step agent workflow."
       },
       "data_preprocessing_documented": {
         "applies": true,
         "answer": true,
         "justification": "Preprocessing steps documented: ambiguous answer removal (3600→3542 and 3600→3420 responses), TF-IDF vectorization with stop word removal and feature limits, NLTK tokenization (Sections 3.1, 3.5). The humanization pipeline is described in Section 3.2."
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": true,
         "justification": "Section 5.4 'Limitations' is a dedicated subsection with substantive discussion of methodological limitations."
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": true,
         "justification": "Section 5.4 discusses threats specific to this study: findings may not generalize beyond GPT-3.5, participants relied on first-person cues as a decision heuristic, and careful prompting could make AI harder to detect."
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": false,
         "justification": "The paper notes 'findings may be specific to GPT-3.5 and might not generalize to other models' but does not systematically state what was NOT tested—e.g., no mention of excluding non-English speakers, non-Western moral frameworks, other scenario types beyond Greene et al., or other demographic populations."
       }
     },
     "data_integrity": {
       "raw_data_available": {
         "applies": true,
         "answer": false,
         "justification": "No raw data is available. Neither the corpus generation data (human/LLM responses) nor the participant evaluation data is released or linked."
       },
       "data_collection_described": {
         "applies": true,
         "answer": true,
         "justification": "Data collection is well-described: 60 scenarios from Greene et al. (2004), 30 human participants + 30 LLM API calls per scenario at temperature=0.7, 230 evaluation participants via Prolific with English fluency screening (Sections 3.1, 3.3)."
       },
       "recruitment_methods_described": {
         "applies": true,
         "answer": true,
         "justification": "Participants 'were recruited through Prolific with the screening requirement that they were fluent in English' with transparent instructions about the study involving both human and AI responses, plus financial incentives (5 cents per correct identification) (Section 3.3.1)."
       },
       "data_pipeline_documented": {
         "applies": true,
         "answer": true,
         "justification": "The full pipeline is documented in Figure 1 and Sections 3.1-3.6: corpus generation → corpus transformation (humanization) → corpus evaluation → linguistic analysis → predictive modeling → semantic analysis. Corpus sizes reported at each stage."
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Acknowledgments section lists: European Research Council consolidator grant (RaReMem: 101043804), three Agence Nationale de la Recherche grants, Alexander Von Humboldt foundation, and a Google unrestricted gift."
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Author affiliations are clearly listed: University of Geneva, Google DeepMind, and École Normale Supérieure. Crystal Qian's DeepMind affiliation is prominent."
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": true,
         "justification": "Primary funders (ERC, ANR, Humboldt) are independent research agencies. The Google 'unrestricted gift' is from a company with general AI interests, but the study evaluates OpenAI models, not Google products. The funding does not create a direct outcome dependency."
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests or financial interests statement is present in the paper. A co-author works at Google DeepMind and Google provided funding, but no formal declaration of competing interests is made."
       }
     },
     "contamination": {
       "training_cutoff_stated": {
         "applies": true,
         "answer": false,
         "justification": "The paper uses GPT-3.5 davinci-text-002 and davinci-text-003 to evaluate moral judgments on scenarios from Greene et al. (2004) but does not state the training data cutoff for either model."
       },
       "train_test_overlap_discussed": {
         "applies": true,
         "answer": false,
         "justification": "The moral scenarios from Greene et al. (2004) are well-known psychology stimuli that almost certainly appear in GPT-3.5's training data, but the paper does not discuss potential train/test overlap or its implications for the observed moral judgment patterns."
       },
       "benchmark_contamination_addressed": {
         "applies": true,
         "answer": false,
         "justification": "Greene et al. (2004) trolley problem scenarios have been widely discussed online since 2004 and are very likely in GPT-3.5's training data. This contamination could explain the LLMs' strong moral preferences, but it is not addressed."
       }
     },
     "human_studies": {
       "pre_registered": {
         "applies": true,
         "answer": false,
         "justification": "No pre-registration is mentioned. For a behavioral experiment with multiple outcome measures and comparisons, pre-registration would strengthen claims against selective reporting."
       },
       "irb_or_ethics_approval": {
         "applies": true,
         "answer": true,
         "justification": "Ethics approval is explicitly stated: 'The study received approval from Paris School of Economics ethical committee (2024-007)' and followed the Declaration of Helsinki (footnote 1, Section 3)."
       },
       "demographics_reported": {
         "applies": true,
         "answer": true,
         "justification": "Demographics reported for both experiments: corpus generation (15 females, mean age = 34 ± 10.26) and evaluation (113 females, mean age = 35.57 ± 11.71) in Section 3.3.1."
       },
       "inclusion_exclusion_criteria": {
         "applies": true,
         "answer": true,
         "justification": "English fluency screening is stated as a Prolific requirement (Section 3.3.1). While minimal, it is an explicit inclusion criterion."
       },
       "randomization_described": {
         "applies": true,
         "answer": false,
         "justification": "The paper states that 77, 76, and 77 participants evaluated corpora 1, 2, and 3 respectively, but does not describe how participants were assigned to conditions—whether random, sequential, or by other means."
       },
       "blinding_described": {
         "applies": true,
         "answer": true,
         "justification": "Blinding status is described: 'Instructions were fully transparent, informing participants that they are expected to give meta-judgments on both human and AI-generated answers' (Section 3.3.1). Participants knew mixed sources were present but were blind to individual item source."
       },
       "attrition_reported": {
         "applies": true,
         "answer": false,
         "justification": "No participant attrition or dropout information is reported. The data filtering from 3600 to 3542/3420 responses concerns ambiguous answer removal, not participant dropout."
       }
     },
     "cost_and_practicality": {
       "inference_cost_reported": {
         "applies": true,
         "answer": false,
         "justification": "No API costs are reported for the GPT-3.5 calls (3600+ calls across two model versions plus humanization), nor for the Prolific participant compensation beyond the bonus structure."
       },
       "compute_budget_stated": {
         "applies": true,
         "answer": false,
         "justification": "No computational budget is stated for training the DistilBERT models, running optuna hyperparameter optimization, or fitting the random forest classifiers."
       }
     }
   },
   "claims": [
     {
       "claim": "LLMs (GPT-3.5) are more sensitive to personal vs. impersonal moral framing than humans, and this sensitivity increased dramatically between model versions.",
       "evidence": "Table 1, Section 4.1: Humans showed marginal framing sensitivity (T(30)=2.8, p<0.05, d=0.69), while dv2 showed strong sensitivity (T(29)=10.05, p<0.001, d=2.23) and dv3 showed extreme sensitivity (T(29)=49.28, p<0.001, d=12.11).",
       "supported": "strong"
     },
     {
       "claim": "Participants prefer AI justifications over human justifications specifically in personal moral scenarios.",
       "evidence": "Table 1, Section 4.3, Figure 5 Row B: Significant preference for AI in personal moral scenarios with large effect sizes. Reversed for non-moral scenarios where participants preferred human justifications (T(152)=5.12, p<0.001, d=0.58).",
       "supported": "strong"
     },
     {
       "claim": "Participants exhibit a systematic anti-AI bias: they are less likely to agree with judgments they believe to be AI-generated, regardless of actual source.",
       "evidence": "Table 1, Section 4.3 'conditioning on belief', Figure 5 Row C: Significant anti-AI bias across all scenario types when conditioned on participant belief (non-moral: T(143)=5.76, p<0.001; personal moral: T(143)=4.31, p<0.001).",
       "supported": "strong"
     },
     {
       "claim": "Humanizing LLM outputs reduces detection accuracy but does not eliminate above-chance detection.",
       "evidence": "Section 4.4, Figure 6A: Detection dropped below 60% for humanized dv2 but remained significantly above chance (T(76)=8.24, p<0.001, d=0.93). ANOVA across corpora: F(2,227)=27.57, p<0.001, η²p=0.19.",
       "supported": "strong"
     },
     {
       "claim": "NLP classifiers can predict text provenance, participant agreement, and detection accuracy with moderate accuracy from justification text alone.",
       "evidence": "Table 2: Multiclass provenance classifier accuracy 0.61, binary provenance 0.63, agreement prediction 0.63, identification prediction 0.62. All above random selection but still modest.",
       "supported": "moderate"
     },
     {
       "claim": "First-person pronoun usage is a persistent cue for detecting human-generated content, surviving humanization efforts.",
       "evidence": "Section 4.5, Figure 7: First-person usage significantly predicted human identification in all three corpora including the humanized condition (corpus 1: T(76)=7.31, p<0.001, d=1.04; corpus 3: T(76)=8.72, p<0.001, d=1.09).",
       "supported": "strong"
     }
   ],
   "red_flags": [
     {
       "flag": "Outdated model versions",
       "detail": "The study uses GPT-3.5 davinci-text-002 and text-003, which were already deprecated by publication date (October 2024). Results may not apply to current-generation models (GPT-4, Claude 3, etc.), significantly limiting the relevance of the findings."
     },
     {
       "flag": "No pre-registration",
       "detail": "A behavioral experiment with 3 outcome measures, 3 scenario types, 3 corpora, and numerous linguistic features creates a large space for selective reporting. No pre-registration or analysis plan is mentioned."
     },
     {
       "flag": "No data or code release",
       "detail": "Neither the generated corpora, participant response data, nor analysis code are released, making independent verification impossible."
     },
     {
       "flag": "Small corpus generation sample",
       "detail": "Only 30 human participants generated the stimulus corpus. These 30 individuals' moral reasoning represents 'human' moral judgment throughout the study, which may not be representative."
     },
     {
       "flag": "Contamination risk unaddressed",
       "detail": "Greene et al. (2004) trolley problem scenarios are among the most widely discussed moral dilemmas online. GPT-3.5 was almost certainly trained on extensive discussions of these exact scenarios, which could explain the observed moral reasoning patterns rather than genuine model 'alignment.'"
     },
     {
       "flag": "ML evaluation methodology not described",
       "detail": "The predictive modeling section (Table 2) reports accuracy and F1 but does not describe train/test splits, cross-validation procedures, or how overfitting was controlled, undermining confidence in the NLP classifier results."
     }
   ],
   "cited_papers": [
     {
       "title": "Evaluating the Moral Beliefs Encoded in LLMs",
       "authors": ["Nino Scherrer", "Claudia Shi", "Amir Feder", "David M. Blei"],
       "year": 2023,
       "arxiv_id": "2307.14324",
       "relevance": "Directly studies LLM moral alignment, finding that LLM responses vary based on question phrasing with closed-source models showing more consistent preferences."
     },
     {
       "title": "Aligning AI With Shared Human Values",
       "authors": ["Dan Hendrycks", "Collin Burns", "Steven Basart", "Andrew Critch", "Jerry Li", "Dawn Song", "Jacob Steinhardt"],
       "year": 2021,
       "arxiv_id": "2008.02275",
       "relevance": "Foundational work on AI alignment with human moral values, directly relevant to the survey's coverage of alignment evaluation."
     },
     {
       "title": "Does GPT-4 pass the Turing test?",
       "authors": ["Cameron R. Jones", "Benjamin K. Bergen"],
       "year": 2024,
       "arxiv_id": "2310.20216",
       "relevance": "Evaluates LLM conversational capabilities through a Turing test framework, relevant to AI detection and evaluation methodology."
     },
     {
       "title": "Human heuristics for AI-generated language are flawed",
       "authors": ["Maurice Jakesch", "Jeffrey T. Hancock", "Mor Naaman"],
       "year": 2023,
       "doi": "10.1073/pnas.2208839120",
       "relevance": "Studies human detection of AI-generated text and identifies flawed heuristics, directly relevant to understanding human-AI interaction."
     },
     {
       "title": "Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference",
       "authors": ["Wei-Lin Chiang", "Lianmin Zheng", "Ying Sheng"],
       "year": 2024,
       "arxiv_id": "2403.04132",
       "relevance": "Platform for evaluating LLMs via human preference, relevant to alignment evaluation methodology."
     },
     {
       "title": "Using cognitive psychology to understand GPT-3",
       "authors": ["Marcel Binz", "Eric Schulz"],
       "year": 2023,
       "doi": "10.1073/pnas.2218523120",
       "relevance": "Applies cognitive psychology methods to study LLM decision-making, a key methodological precedent for this paper."
     },
     {
       "title": "Studying and improving reasoning in humans and machines",
       "authors": ["Nicolas Yax", "Hernán Anlló", "Stefano Palminteri"],
       "year": 2024,
       "doi": "10.1038/s44271-024-00091-8",
       "relevance": "Compares human and LLM reasoning using psychology tools, relevant to cognitive science approaches to LLM evaluation."
     },
     {
       "title": "Strong and weak alignment of large language models with human values",
       "authors": ["Mehdi Khamassi", "Marceau Nahon", "Raja Chatila"],
       "year": 2024,
       "doi": "10.1038/s41598-024-70031-3",
       "relevance": "Studies LLM alignment with human values, directly relevant to the survey's alignment evaluation scope."
     },
     {
       "title": "'Correct answers' from the psychology of artificial intelligence",
       "authors": ["Peter S. Park", "Philipp Schoenegger", "Chongyang Zhu"],
       "year": 2023,
       "arxiv_id": "2302.07267",
       "relevance": "Identifies 'correct answer' bias in LLMs trained to produce socially accepted responses, relevant to understanding LLM evaluation artifacts."
     },
     {
       "title": "Sleeper Social Bots: a new generation of AI disinformation bots are already a political threat",
       "authors": ["Jaiv Doshi", "Ines Novacic", "Curtis Fletcher"],
       "year": 2024,
       "arxiv_id": "2408.12603",
       "relevance": "Addresses LLM-powered disinformation risk, relevant to AI safety and the consequences of human inability to detect AI-generated content."
     },
     {
       "title": "AI AI Bias: Large Language Models Favor Their Own Generated Content",
       "authors": ["Walter Laurito", "Benjamin Davis", "Peli Grietzer"],
       "year": 2024,
       "arxiv_id": "2407.12856",
       "relevance": "Studies LLM self-favoring bias, relevant to understanding alignment and evaluation biases in AI systems."
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 1,
       "justification": "Findings about anti-AI bias and detection cues are informative for AI product designers, but no directly usable tool or technique is provided."
     },
     "surprise_contrarian": {
       "score": 2,
       "justification": "The paradox that humans prefer AI moral judgments while simultaneously exhibiting anti-AI bias is counterintuitive and challenges simplistic narratives about AI aversion."
     },
     "fear_safety": {
       "score": 1,
       "justification": "Raises concerns about LLM moral reasoning alignment and detection difficulty, but does not demonstrate a novel attack or existential risk."
     },
     "drama_conflict": {
       "score": 1,
       "justification": "The anti-AI bias finding is mildly provocative but the paper takes a measured academic tone without controversial framing."
     },
     "demo_ability": {
       "score": 0,
       "justification": "No code, demo, or interactive tool is released."
     },
     "brand_recognition": {
       "score": 2,
       "justification": "Co-author from Google DeepMind, study uses GPT-3.5 (well-known product), but neither the lab nor the models are the headline focus."
     }
   }
 }