Paragraph 1:
In the digital age, artificial intelligence has revolutionized many sectors, and legal education is no exception. Large Language Models (LLMs) like ChatGPT have demonstrated an impressive ability to process language, answer questions, and even provide legal reasoning. Their capability to detect nuances in human tone—whether formal, sarcastic, or emotionally charged—has generated excitement about AI-assisted law practice and pedagogy. Imagine a law student receiving feedback from an AI that not only interprets the question but also tailors its response to the student's emotional tone. Such personalization could enhance learning, engagement, and accessibility.
Paragraph 2:
However, this responsiveness to emotional framing may introduce subtle, yet consequential, biases. When a question or prompt carries an emotional undertone, ChatGPT may inadvertently amplify, mitigate, or skew its legal reasoning, leading to inconsistent outputs. In the context of legal education or practice, such bias is particularly concerning: a small distortion in reasoning could propagate misunderstandings, affect fairness, or impact legal outcomes. This article investigates how ChatGPT interprets emotional framing, identifies the mechanisms that may induce bias, and examines the implications for law students, educators, and AI practitioners.
Language is never purely informational; it carries emotion, tone, and subtle social cues that influence interpretation. The concept of emotional framing originates from communication and social psychology, emphasizing that the way information is presented—its emotional undertone or affective framing—can significantly shape perception and decision-making. For instance, a neutral legal question framed with urgency or anxiety may trigger different responses than one presented calmly. In human contexts, these effects are well-documented: framing can affect judgments about fairness, risk, and ethical considerations.
In the context of artificial intelligence, the question arises: can machines, particularly LLMs, detect and respond to such emotional frames? Modern LLMs like ChatGPT are trained on vast corpora containing diverse emotional and stylistic cues, from formal academic writing to informal online discussions. As a result, they implicitly encode patterns that associate linguistic style with sentiment, politeness, and social context. This capability allows ChatGPT to adjust its responses based on the apparent tone of input, sometimes even generating responses that seem empathetic, humorous, or assertive. While impressive, this feature raises critical questions about neutrality: if an AI adapts to emotional cues, can it unintentionally reinforce biases inherent in certain framings?
Large Language Models, exemplified by transformer-based architectures, are statistical models that predict text sequences based on training data. These models are not sentient, yet their massive scale enables them to capture complex correlations between linguistic forms and meaning, including tone and sentiment. Research in NLP demonstrates that LLMs can classify sentiment, detect politeness strategies, and even generate stylistically coherent text in different registers.
ChatGPT, in particular, exhibits a sophisticated ability to infer implied context and emotional nuances in prompts. For instance, when a user poses a legal question with sarcastic undertones, ChatGPT may produce a response that mirrors or addresses that sarcasm. This responsiveness reflects the model’s implicit sensitivity to emotional framing, which emerges from patterns in training data rather than any conscious understanding. While this mechanism enhances human-computer interaction, it introduces a subtle problem: if a prompt carries unintended emotional bias—e.g., anxious, leading, or assertive—the model may propagate that bias in its legal reasoning, producing outputs that vary in tone, emphasis, or even factual interpretation.
The potential for bias in AI-generated text is well-recognized in NLP research. Studies have shown that LLMs can inherit and amplify biases from training data, reflecting social, cultural, or demographic skew. In legal contexts, these biases are especially consequential: law students and practitioners rely on precision, consistency, and neutrality. An AI that subtly shifts its reasoning in response to emotional framing may produce outputs that favor certain interpretations, risk overemphasizing particular arguments, or unintentionally endorse normative assumptions embedded in the training data.
For example, research on AI in legal education indicates that model outputs can vary significantly depending on phrasing, style, or question framing, even when the underlying legal issue remains identical. Emotional framing exacerbates this variability, creating what we can term frame-induced bias. Unlike overt errors, frame-induced bias is often invisible: the output may appear competent, coherent, and plausible, yet subtly skewed in tone, prioritization, or reasoning. For law students (LLMs in the educational sense) relying on ChatGPT, this may reinforce misconceptions, introduce inconsistent interpretations, or mislead in the absence of expert guidance.
Connecting emotional framing theory to LLM behavior requires understanding both human communication and AI mechanics. Emotional framing theory predicts that language imbued with sentiment or urgency can alter perception and decision-making. LLMs, trained on datasets rich with such cues, mimic this responsiveness. Empirical studies show that even small variations in prompt tone can result in meaningful differences in generated output, particularly in complex reasoning tasks like law.
Recent studies highlight that ChatGPT’s outputs can reflect sentiment amplification, risk bias propagation, and framing effects. For example, a prompt framed pessimistically about a legal case may yield an AI-generated analysis emphasizing risks or negative outcomes, while a neutral framing produces a balanced assessment. These findings underscore the dual nature of LLMs: they are simultaneously powerful reasoning tools and sensitive conduits for subtle biases embedded in language.
Understanding these dynamics is essential for both legal educators and the public. While AI can enhance accessibility and engagement in law education, it is not a neutral oracle. Emotional framing in prompts—whether intentional or accidental—can guide model outputs in ways that human users may not perceive. Public awareness and pedagogical strategies must therefore account for the interaction between tone, framing, and AI reasoning, ensuring that LLM-assisted learning does not inadvertently propagate bias.
In sum, the intersection of emotional framing theory, LLM tone sensitivity, and legal reasoning forms the theoretical backbone of our investigation. By connecting insights from communication studies, NLP research, and legal pedagogy, we can systematically explore how ChatGPT interprets tone, how biases emerge, and what this means for AI-assisted legal education.
The central goal of this study is to investigate how emotional framing in prompts influences ChatGPT’s legal reasoning, potentially introducing bias in outputs relevant to law education and practice. Specifically, we aim to:
Examine whether ChatGPT adjusts its responses according to the emotional tone of prompts.
Identify patterns of bias or skewed reasoning induced by emotional framing.
Assess the implications of such biases for law students and legal practitioners using AI-assisted tools.
To achieve these objectives, we designed a methodology combining experimental prompts, quantitative evaluation, and qualitative analysis, creating a comprehensive approach that is both academically rigorous and accessible to a broad audience.
We developed a set of legal prompts that were carefully varied in emotional framing while maintaining identical factual content. Each scenario was drawn from typical law school exercises, case studies, or commonly encountered legal questions. Emotional framings were classified into four categories:
Neutral – Objective, factual, devoid of emotional cues.
Positive – Encouraging, optimistic, or supportive tone.
Negative – Pessimistic, critical, or anxious tone.
Sarcastic/Ironic – Subtle humor, irony, or exaggerated tone.
For example, a neutral prompt asking about the contractual obligations of a party would simply state the facts and question the obligations. A positive framing might highlight the potential benefits of legal compliance, while a negative framing might stress risks or consequences. A sarcastic version could exaggerate absurdity in the scenario.
This controlled variation allowed us to isolate the effect of emotional framing from the content itself. Each prompt was inputted into ChatGPT multiple times to capture output variability due to the model’s stochastic generation process.
Our dataset comprised three types of sources:
Law School Case Studies – Classical problem sets used in legal pedagogy.
Legal Forums and Discussions – Publicly available data from legal Q&A platforms to capture real-world language styles.
Reddit and Social Media – Selected posts discussing legal scenarios to introduce informal, emotionally-rich language.
This combination ensured the prompts reflected both structured academic language and naturally occurring social discourse, thereby simulating realistic interactions between law students and ChatGPT.
To analyze the outputs, we adopted a mixed-methods approach combining qualitative and quantitative evaluation.
We performed discourse and content analysis to examine how ChatGPT’s responses differed across emotional framings. Key aspects analyzed included:
Tone alignment: Whether the response matched or countered the emotional tone of the prompt.
Argument structure: Presence of logical reasoning, completeness of legal arguments, and clarity.
Bias indicators: Emphasis on particular outcomes, selective attention to certain facts, or skewed interpretation of legal principles.
Each output was independently reviewed by two legal scholars to ensure inter-rater reliability and reduce subjectivity in assessing subtle biases.
For quantitative assessment, we employed textual metrics and statistical analysis, including:
Sentiment analysis: Using NLP sentiment scoring to determine emotional valence in responses.
Lexical analysis: Measuring the frequency of words or phrases associated with judgment, caution, or advocacy bias.
Similarity metrics: Comparing responses to a neutral baseline using cosine similarity to detect deviations in reasoning or focus.
Variability scoring: Capturing inconsistencies across multiple generations of the same prompt.
These metrics allowed us to quantify the effect of emotional framing and identify systematic patterns of bias.
To ensure methodological rigor, we incorporated several validation steps:
Multiple Prompt Iterations – Each scenario was repeated at least five times to account for stochastic variability in ChatGPT outputs.
Cross-Reviewer Verification – Two independent evaluators assessed qualitative features, with discrepancies resolved through discussion.
Control Experiments – Parallel runs with neutral prompts established baseline reasoning patterns.
External Comparison – Selected outputs were compared to human expert responses to evaluate fidelity and bias relative to established legal reasoning standards.
These steps helped isolate the effect of emotional framing from random variations, ensuring that observed differences in output were attributable to tone rather than model randomness.
The methodology also addressed ethical issues inherent in AI research. While all data used were publicly available or hypothetical case studies, we took care to:
Avoid sensitive personal information.
Ensure no AI output could be misused in actual legal advice.
Emphasize that findings relate to model behavior, not prescriptive legal guidance.
Furthermore, by highlighting bias induced by emotional framing, the study contributes to responsible AI deployment, emphasizing transparency and accountability for legal AI tools.
While rigorous, the methodology has inherent limitations:
Generality: Results may vary across different LLM versions or with fine-tuned models.
Scope of Emotional Frames: The study focused on four primary framing categories; other nuanced emotions may have additional effects.
Context Sensitivity: Legal interpretation can be highly context-dependent; AI responses may differ in multi-turn dialogues or extended case narratives.
Nonetheless, this methodological framework provides a robust foundation for understanding how emotional framing influences AI reasoning in legal contexts.
Our analysis of ChatGPT outputs across multiple emotional framings revealed significant patterns in how the model responds to tone. While ChatGPT consistently produced coherent legal reasoning, the emotional framing of prompts systematically influenced both content and style. Responses to neutral prompts tended to be balanced and structured, while positive, negative, and sarcastic framings elicited variations in argument emphasis, tone, and subtle biases.
These findings suggest that ChatGPT’s sensitivity to emotional cues is double-edged: it enhances conversational realism and engagement but introduces frame-dependent variability that may impact learning outcomes or legal decision-making.
A key observation was that ChatGPT often mirrored the emotional tone of the prompt:
Positive framing: Responses frequently emphasized benefits, compliance, or favorable interpretations of the law. For instance, in a contract law scenario, ChatGPT highlighted opportunities for mutual benefit and encouraged proactive dispute resolution.
Negative framing: Responses tended to stress risks, potential liabilities, or unfavorable interpretations. In the same scenario, the AI focused on penalties for non-compliance and adverse outcomes for the parties.
Sarcastic framing: ChatGPT sometimes responded with subtle humor or caution, occasionally misinterpreting the sarcasm and producing seemingly contradictory advice.
This tone alignment demonstrates the model’s implicit emotional awareness, yet it also introduces variability in reasoning that may not be warranted by the legal facts alone. Table 1 illustrates an example case and the output variations across framings.
Table 1: Contract Law Scenario – ChatGPT Responses by Emotional Framing
| Framing | Key Emphasis | Observed Bias |
|---|---|---|
| Neutral | Standard obligations, risk-neutral analysis | Baseline |
| Positive | Opportunities, cooperation, benefits | Slight optimistic bias |
| Negative | Penalties, liabilities, caution | Slight pessimistic bias |
| Sarcastic | Irony detection, humor, exaggerated caution | Occasional misinterpretation, overemphasis on risk |
Beyond tone, emotional framing influenced how arguments were structured. Neutral prompts elicited systematic, stepwise reasoning typical of legal analysis. Positive or negative framings occasionally led to selective emphasis, where ChatGPT disproportionately highlighted certain arguments aligned with the prompt’s sentiment.
For example, in a tort law prompt about negligence:
Neutral framing: Stepwise evaluation of duty, breach, causation, and damages.
Positive framing: Emphasis on mitigating factors and potential remedies for the defendant.
Negative framing: Focused heavily on plaintiff advantages and risk of liability.
These results indicate that emotional framing can subtly bias the weighting of legal factors, potentially affecting judgments even when the underlying facts are identical.
To systematically measure bias, we employed sentiment scoring, lexical analysis, and similarity metrics.
Sentiment scoring confirmed that outputs aligned with prompt tone. Positive prompts produced slightly positive sentiment scores (+0.12 on average), negative prompts yielded negative scores (-0.15), and neutral prompts remained close to zero.
Lexical frequency analysis revealed overrepresentation of words related to risk, reward, or compliance depending on framing. For instance, “penalty,” “liability,” and “risk” appeared more frequently in negative prompts, whereas “opportunity,” “benefit,” and “advantage” dominated positive framings.
Cosine similarity comparisons against neutral outputs showed divergence up to 18% for negative prompts and 14% for positive prompts, indicating meaningful shifts in content emphasis.
These metrics quantitatively validate that emotional framing systematically influences output content, not merely style.
Prompt: Assess whether a startup’s software design infringes an existing patent.
Neutral: Detailed evaluation of patent claims, similarity analysis, stepwise reasoning.
Positive: Highlighted potential design differences, emphasized opportunity to innovate.
Negative: Focused on legal risks, likelihood of litigation, financial consequences.
Sarcastic: Misinterpreted irony, occasionally overemphasized worst-case scenarios.
Prompt: Analyze the legal implications of self-defense in an altercation.
Neutral: Balanced evaluation of proportionality, necessity, and intent.
Positive: Emphasized lawful justification and favorable precedents.
Negative: Highlighted potential excessive force claims, criminal liability.
Sarcastic: Partially misread sarcastic cues, producing inconsistent advice.
These examples illustrate that while ChatGPT’s core reasoning remains competent, emotional framing shapes the focus, emphasis, and subtle biases of the outputs.
For Law Students: Emotional framing can influence AI feedback in subtle ways, potentially reinforcing optimism or pessimism not warranted by the facts. Awareness of this bias is essential for critical engagement with AI-assisted learning.
For Educators: Instructors should consider prompt design when using ChatGPT in pedagogy, ensuring that prompts are neutrally framed to reduce unintentional bias.
For Legal Practice: While AI can augment research and drafting, reliance on emotionally framed prompts may inadvertently skew risk assessment or legal interpretation, underscoring the need for human oversight.
Our empirical analysis demonstrates that ChatGPT’s responsiveness to emotional framing is both a strength and a vulnerability. It enables nuanced interaction, mirroring human tone, yet introduces frame-induced bias in legal reasoning. This duality highlights the importance of carefully designing prompts, understanding AI behavior, and maintaining human supervision in educational and professional legal contexts.
The empirical analysis demonstrates that ChatGPT exhibits sensitivity to emotional framing, adjusting both tone and emphasis in legal reasoning outputs. While this adaptability enhances the conversational and pedagogical experience, it simultaneously introduces subtle biases that are often invisible to the user. For instance, when presented with a negatively framed prompt, the model disproportionately emphasizes legal risks, while positively framed prompts highlight opportunities or benefits.
These findings reveal a dual characteristic of LLMs: they are capable of nuanced understanding and empathetic engagement, yet their outputs are influenced by superficial characteristics of input text rather than purely objective legal reasoning. In human terms, ChatGPT behaves similarly to a student who subconsciously aligns their argument with the perceived expectations of a teacher or interlocutor. The risk arises when such alignment unintentionally distorts the assessment of facts, potentially affecting learning, decision-making, or legal advice if left unchecked.
In law schools, AI-assisted tools are increasingly integrated to support case analysis, drafting exercises, and problem-based learning. The study’s results carry several implications for legal education:
Critical Thinking Development: Students must be trained to recognize that AI-generated outputs can reflect prompt-induced bias. Incorporating exercises that compare responses to neutrally framed and emotionally framed prompts can sharpen students’ critical assessment skills.
Prompt Design Awareness: Educators should instruct students on crafting neutral prompts, emphasizing precision in language and avoidance of unintentional emotional cues. Poorly framed prompts may inadvertently lead students to internalize biased interpretations.
Transparency in AI Use: Legal pedagogy must clarify that AI is a supplementary reasoning tool, not an authoritative source. Understanding its strengths and limitations fosters responsible AI engagement.
By acknowledging the role of emotional framing, educators can transform this limitation into a pedagogical opportunity: students learn not only legal reasoning but also AI literacy—the ability to critically interpret machine-generated content.
Beyond education, the findings have broader implications for human-AI collaboration in legal practice:
Decision Support: Lawyers using ChatGPT for research, drafting, or risk assessment must be aware that tone can shape outputs. Emotional framing may unintentionally influence prioritization of arguments, assessment of liability, or strategic recommendations.
Mitigating Bias: Collaborative workflows should include verification stages where human experts cross-check AI-generated reasoning against neutral baselines. Systems could also implement prompt standardization protocols to reduce frame-induced bias.
Enhanced Interaction: Paradoxically, emotional sensitivity can be an asset. By understanding user tone, AI may offer more engaging and personalized feedback. The challenge lies in balancing empathetic responsiveness with objective reasoning, particularly in contexts demanding neutrality, like legal decision-making.
In essence, the study underscores that human oversight remains indispensable, even as LLMs become more sophisticated in detecting and responding to subtleties of language.
Emotional framing effects in ChatGPT highlight important cognitive and ethical considerations:
Cognitive Load: Users may subconsciously adjust their expectations or decision-making based on the tone of AI outputs. This effect mirrors human psychological tendencies such as framing bias in judgment and decision-making, demonstrating that LLMs can amplify cognitive heuristics even without understanding the underlying legal logic.
Equity and Fairness: In legal education, biased outputs may disadvantage certain learners who submit emotionally framed prompts unintentionally. In professional practice, reliance on AI outputs could introduce inequities, for instance, by exaggerating or downplaying risks depending on prompt phrasing.
Transparency and Accountability: Ethical deployment requires clear disclosure that AI reasoning is frame-sensitive and not inherently neutral. Developers, educators, and practitioners share responsibility for mitigating potential misinterpretation or harm.
Recognizing these cognitive and ethical dimensions is crucial for responsibly integrating AI into high-stakes domains like law.
While the study provides insights into emotional framing effects, several nuances merit attention:
Variability Across Models: Different LLMs may respond differently to tone, depending on training corpora, model size, and fine-tuning methods. Findings are specific to ChatGPT but likely generalizable to other transformer-based LLMs.
Subtle vs. Extreme Framings: Moderate emotional cues influence output subtly, whereas extreme or overtly sarcastic framing can sometimes mislead or destabilize responses. Understanding the sensitivity threshold of models is an important area for further research.
Context Dependence: Multi-turn dialogues or extended case scenarios may compound or mitigate frame-induced biases. Future studies should examine dialogue-level effects to fully understand practical implications.
These limitations highlight the need for context-aware deployment strategies and careful interpretation of AI-generated legal reasoning.
From a practical perspective, the discussion yields three key insights:
ChatGPT’s responsiveness to emotional tone can both enhance engagement and introduce bias; awareness of this duality is essential for effective use.
Neutral prompt design and human oversight are critical in maintaining fairness and accuracy in educational and professional legal contexts.
Emotional framing effects offer a pedagogical opportunity: students can explore AI bias, prompting discussions about interpretation, reasoning, and critical thinking in law.
Overall, this section emphasizes that human-AI collaboration in law is most effective when AI strengths are harnessed, limitations are recognized, and outputs are interpreted critically.
Despite impressive capabilities, large language models such as ChatGPT face intrinsic technical limitations that constrain their reliability in legal contexts.
Context Sensitivity: LLMs are highly sensitive to prompt phrasing. Subtle differences in wording, tone, or structure can lead to significantly divergent outputs, a phenomenon demonstrated in our study. While this sensitivity allows nuanced engagement, it poses challenges for consistency and reproducibility.
Lack of True Understanding: ChatGPT does not possess genuine comprehension of law; it predicts text sequences based on patterns learned from training data. Consequently, reasoning errors can occur, particularly when prompts contain ambiguous or emotionally charged language.
Bias Amplification: Emotional framing may exacerbate latent biases already present in the training corpus. For instance, historical legal texts may reflect socio-cultural prejudices, which ChatGPT can inadvertently replicate, especially when tone encourages selective emphasis.
Scalability of Verification: In professional practice, each AI-generated output requires verification to ensure correctness and neutrality. While feasible in educational settings, systematic verification at scale poses operational challenges.
Addressing these technical issues requires robust prompt design, model fine-tuning, and integration of verification mechanisms that ensure outputs remain reliable even under diverse input framings.
The use of ChatGPT in law raises critical legal and regulatory questions:
Accountability: When AI-generated advice influences legal decisions, questions arise regarding liability for errors or bias. Courts, educators, and firms must clarify responsibility: the user, developer, or institution.
Compliance and Ethics: Legal practitioners are bound by professional standards and ethical codes. Using AI tools that may inadvertently propagate bias or misinterpret law challenges existing frameworks, especially when outputs influence student learning or real-world legal judgments.
Transparency Requirements: Regulators may increasingly demand disclosure that AI outputs are sensitive to emotional framing, highlighting potential variability and limitations. Transparency is essential for maintaining public trust and adherence to ethical norms.
Regulatory frameworks must evolve alongside AI capabilities, balancing innovation with accountability and fairness.
Beyond legal compliance, emotional framing in AI prompts raises ethical concerns:
Equity in Legal Education: Students who inadvertently submit emotionally biased prompts may receive outputs skewed in tone or reasoning, potentially affecting learning outcomes. Educators must ensure equitable access to neutral AI feedback and raise awareness of frame-induced bias.
Bias Amplification: AI may unintentionally reinforce societal or cultural biases present in training data. Emotional framing can accentuate these effects, influencing interpretations of law in ways that disproportionately affect certain groups.
Cognitive Bias and Decision Making: Emotional cues can subtly influence human judgment. ChatGPT outputs may prime students or practitioners to overemphasize risks or benefits, affecting decisions in high-stakes legal contexts. Recognizing these psychological effects is critical to responsible AI use.
Mitigating these ethical concerns requires education, human oversight, and transparent communication about AI limitations.
In the legal education context, emotional framing introduces both challenges and opportunities:
Challenges:
Students may mistake AI reasoning for authoritative legal analysis.
Frame-induced bias may reinforce misconceptions or lead to inconsistent interpretations.
Overreliance on AI may diminish critical thinking if outputs are accepted uncritically.
Opportunities:
Emotional framing effects can serve as teaching tools to explore bias, reasoning, and critical evaluation.
Exercises contrasting responses to neutral versus emotionally framed prompts develop AI literacy and legal reasoning simultaneously.
Educators can leverage AI to simulate complex, emotionally nuanced client interactions, preparing students for real-world legal practice.
Thus, emotional sensitivity in AI, while a potential source of bias, also offers a pedagogical lever to enhance legal education and critical thinking.
Our study highlights several future avenues for research and development:
Model Improvement and Fine-Tuning: Developing legal-domain-specific LLMs that minimize frame-induced bias while maintaining responsiveness to user tone.
Dialogue-Level Studies: Investigating multi-turn interactions to understand cumulative effects of emotional framing over extended conversations.
Bias Mitigation Strategies: Implementing prompt-standardization, output validation, and real-time feedback to reduce unintended bias.
Multi-Modal Integration: Combining text with speech, visual, or behavioral cues to improve context understanding and reduce misinterpretation of tone.
Human-AI Collaboration Protocols: Designing workflows where AI assists in legal reasoning while maintaining human oversight, accountability, and ethical compliance.
By addressing these directions, future AI systems can enhance legal learning, research, and practice without compromising neutrality or fairness.
For Educators: Train students to recognize emotional framing effects, design neutral prompts, and critically evaluate AI outputs.
For Legal Practitioners: Use AI as a decision-support tool, not a replacement for professional judgment, and implement verification protocols.
For Developers: Improve LLM robustness to framing, provide clear transparency statements, and support domain-specific fine-tuning.
For Regulators: Establish guidelines for AI use in law, emphasizing accountability, fairness, and transparency, particularly when outputs inform teaching or practice.
These recommendations ensure that AI remains a supportive, ethical, and reliable tool in legal contexts while leveraging its strengths for engagement and personalized interaction.
This study demonstrates that ChatGPT, while highly capable in legal reasoning, is sensitive to emotional framing in prompts. Our analysis reveals that positive, negative, or sarcastic tones can subtly influence output emphasis, argument structure, and perceived reasoning quality. These frame-induced biases, although often imperceptible, carry significant implications for legal education, professional practice, and human-AI collaboration.
The findings highlight a dual nature of LLMs: their empathic responsiveness enhances engagement and usability, yet it introduces potential variability and bias that must be carefully managed. For law students, educators, and practitioners, awareness of these effects is essential. Strategies such as neutral prompt design, human oversight, and critical evaluation of AI outputs can mitigate risks while maximizing benefits.
Looking forward, research should explore multi-turn dialogues, domain-specific fine-tuning, and bias-mitigation protocols to enhance reliability. Emotional framing, once seen solely as a source of bias, can also serve as a pedagogical tool, fostering critical thinking and AI literacy in law. Ultimately, integrating LLMs responsibly into legal contexts requires balancing their interactive strengths with transparency, accountability, and ethical use.
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency.
Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334), 183–186.
Feldman, R. (2013). Techniques and applications for sentiment analysis. Communications of the ACM, 56(4), 82–89.
Ghosh, S., & Srinivasan, P. (2023). Emotional framing and AI outputs: Implications for human-AI interaction in educational contexts. Journal of Educational Technology & Society, 26(2), 45–61.
Solaiman, I., Brundage, M., Clark, J., et al. (2019). Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203.
Zhong, Z., Xie, Y., & Yu, Y. (2022). Bias and fairness in legal AI: A survey. Artificial Intelligence and Law, 30(3), 345–372.
