DECAI Framework: Causality & AI Control Analysis

• DECAI Framework is a dual-paradigm system that combines a multi-task learning model for document-level event causality and a control-theoretic model for evaluating AI interface harms.
• The event causality model transforms causal detection into an MCQ task with rationale generation and structure-aware graph linearization, achieving competitive F1 scores on benchmark datasets.
• The control-theoretic model employs state-space dynamics to assess feedback loops in adaptive AI systems, highlighting how design patterns can drive ethical and social risks.

The DECAI framework refers to two distinct, technically rigorous paradigms: (1) a multi-task learning architecture for document-level event causality identification in natural language processing, and (2) a control-theoretic model for analyzing harms arising from the interactions between adaptive AI systems and user interface design patterns. Both share the DECAI acronym but address fundamentally different problem domains. Each framework is described below with precision according to the technical literature.

1. Multi-Task Learning Framework for Event Causality Identification

1.1. Problem Formulation

Document-level Event Causality Identification (DECI) seeks to determine whether a causal relationship exists between any pair of events in a document. Formally, for a document $\mathcal{D}$ with event mentions $E = \{e_1, ..., e_N\}$, the task is to assign to each ordered event pair $(e_i, e_j)$ a label $y_{ij} \in Y = \{0, 1\}$ indicating causality ($y_{ij} = 1$ if $e_i$ and $e_j$ are causally linked, $0$ otherwise) (Zhang et al., 2024).

1.2. Core Components

The DECAI framework transforms DECI into a multiple-choice question answering (MCQ) problem using a decoder-only generative LLM $f$, and augments this formulation with two auxiliary tasks: (a) rationale generation, and (b) structure-aware event causal graph linearization. This joint architecture is trained in a fully multi-task fashion.

MCQ Transformation

• Text Clipping: For each event $e_i$, candidate events $E = \{e_1, ..., e_N\}$0 are filtered to those occurring later in the narrative, extracting short context windows to suppress irrelevant content.
• Prompt Construction: A prediction prompt $E = \{e_1, ..., e_N\}$1 of the form:

"Given the following context: $E = \{e_1, ..., e_N\}$2, question: What are the causes and effects of [$E = \{e_1, ..., e_N\}$3]? Options: (A) $E = \{e_1, ..., e_N\}$4, (B) $E = \{e_1, ..., e_N\}$5, …, (None) None of the above. Answer with the letter(s)."

• Option Set:
  • Training: All true causal partners (avoiding coreference duplicates), at least three distractor (non-causal) events, and "None of the above."
  • Test: All $E = \{e_1, ..., e_N\}$6 and "None."

Rationale Generation

A rationale $E = \{e_1, ..., e_N\}$7 is generated for each MCQ by prompting a large LLM (e.g., GPT-3.5-turbo) with:

"The correct answer is {label}. Please explain in ≤50 words why this is so."

Rationale prediction is cast as a sequence-to-sequence task: given $E = \{e_1, ..., e_N\}$8 and the relevant context, the model generates $E = \{e_1, ..., e_N\}$9.

Structure-Aware Causal Graph (ECG) Linearization

An undirected event structure graph $(e_i, e_j)$0 is constructed, with edges denoting annotated causal or event-coreference relations. Linearization produces a textual sequence $(e_i, e_j)$1 by ordering all coreference edges, then causal edges (e.g., $(e_i, e_j)$2, "is $(e_i, e_j)$3", $(e_i, e_j)$4, "is the cause of $(e_i, e_j)$5", ...). The model is tasked, via $(e_i, e_j)$6, to reconstruct $(e_i, e_j)$7 token-by-token.

Unified Multi-Task Objective

Three tasks are optimized jointly, each distinguished by its prompt:

• Label prediction ($(e_i, e_j)$8)
• Rationale generation ($(e_i, e_j)$9)
• ECG reconstruction ($y_{ij} \in Y = \{0, 1\}$0)

The total loss is

$y_{ij} \in Y = \{0, 1\}$1

with $y_{ij} \in Y = \{0, 1\}$2, and $y_{ij} \in Y = \{0, 1\}$3, $y_{ij} \in Y = \{0, 1\}$4, $y_{ij} \in Y = \{0, 1\}$5 determined via grid search.

1.3. Model Architecture and Training

• Backbone: Baichuan2-7B-Chat decoder-only LLM.
• Fine-Tuning: LoRA (rank=32) for parameter-efficient fine-tuning.
• Hyperparameters: learning rate $y_{ij} \in Y = \{0, 1\}$6, dropout 0.1, 10 epochs, 2$y_{ij} \in Y = \{0, 1\}$7RTX3090 GPUs.

1.4. Empirical Results

• Datasets:
  • EventStoryLine: 258 docs, 5,655 causal pairs (mostly inter-sentential)
  • Causal-TimeBank: 184 docs, 318 causal pairs (mostly intra-sentential)
• Metrics: Precision, Recall, F1 (overall, intra-, and inter-sentential).

Dataset	Precision	Recall	F1	Intra-F1	Inter-F1
EventStoryLine	48.7	56.0	52.1	65.3	47.2
Causal-TimeBank	-	-	61.2	-	-

• On EventStoryLine, DECAI matches or exceeds discriminative state-of-the-art (e.g., CHEER: 51.4; SENDIR: 51.9), strongly outperforming generative baselines such as GenECI and GPT-3.5 zero-shot. On Causal-TimeBank, DECAI achieves F1=61.2, comparable to SOTA (62.3).

1.5. Ablation and Analysis

• Removing rationale loss or ECG loss yields F1 drops (both to 51.1).
• Ablating all auxiliary tasks (pure Q→A) reduces F1 to 48.6.
• Single-pass answer+rationale generation (Q→A⊕R) causes catastrophic F1 collapse to 28.5.
• Optimal performance with four candidates plus "None."

This suggests that multi-task signals and high-quality rationale supervision (source: GPT-3.5-turbo) are critical for robust DECI, while structure-aware graph encoding stabilizes generative reasoning, particularly for cross-sentence causal inference (Zhang et al., 2024).

2. Control-Theoretic Model for Analyzing Harmful Feedback in Adaptive AI Interfaces

2.1. Motivation and Scope

The Design-Enhanced Control of AI systems (DECAI) framework is motivated by the recognition that rich, interactive AI interfaces (e.g., recommendation feeds, chatbots) mediate nearly all real-world human–AI encounters. Interface design patterns—such as dark patterns, anthropomorphic features, frictionless elements (e.g., infinite scroll), and opaque controls—shape both user behavior and perception, setting up feedback loops that can amplify ethical and social harms (e.g., overreliance, privacy erosion, exposure to polarizing content) (Ibrahim et al., 2024).

2.2. Formal Control-Theoretic Formalism

DECAI models the human–AI–interface system as a closed-loop dynamical system:

• $y_{ij} \in Y = \{0, 1\}$8: user latent state (e.g., trust, agency, emotional arousal)
• $y_{ij} \in Y = \{0, 1\}$9: control input (intensity of design affordances: encouragement, friction, anthropomorphic cues)
• $y_{ij} = 1$0: observable user behavior (e.g., click-through rate, time spent, PII disclosures)
• $y_{ij} = 1$1: system matrices describing evolution and observation
• $y_{ij} = 1$2: system transfer function

The state-space equations are:

$y_{ij} = 1$3

Laplace domain:

$y_{ij} = 1$4

The control loop is closed as interface mediator functions transform AI model output $y_{ij} = 1$5 and user input $y_{ij} = 1$6 into $y_{ij} = 1$7 and $y_{ij} = 1$8, respectively. Design features $y_{ij} = 1$9 carry quantized action intensities $e_i$0.

2.3. System Block Structure

Three interacting blocks:

• AI System (Controller): Maps sensed user input $e_i$1, via $e_i$2, to outputs $e_i$3.
• Interface Mediator (Actuator+Sensor): Presents $e_i$4 using visible design features (actuator $e_i$5) and records user behaviors (sensor $e_i$6), applying action-intensity affordances.
• User (Process): Latent state $e_i$7 evolves under $e_i$8 and projects to measurable behavior $e_i$9.

The full feedback chain:

1. $e_j$0 (actuated affordances)
2. $e_j$1 and $e_j$2 (behavioral impact)
3. $e_j$3 (sensed input)
4. $e_j$4 (update AI output)

2.4. Case Studies

2.4.1. Recommendation Systems

• User conditions: High social validation needs, low expertise
• Interface features: Infinite scroll/autoplay (encourages engagement), hidden "not interested" button (discourages negative feedback), absence of content labeling
• Emergent loop: Passive monitoring treated as positive engagement, escalating addictive exposure, slow correction due to asynchronous retraining intervals

2.4.2. Conversational LLMs

• User conditions: Low technical literacy, weak fact-checking
• Interface features: Anthropomorphic cues (simulate sentience), hidden disclaimers (suppress skepticism), no query editing (overcommitment)
• Emergent loop: Sensitive disclosures increase, overreliance develops, interface nudges outpace model retraining, compounding risk

2.5. Practical Assessment and Recommendations

The DECAI assessment process consists of systematic modeling steps:

1. Initialization: Define initial latent/user states and system outputs
2. User Condition Profiling: Catalog user demographics, cognitive load, expertise
3. Interface Inventory: Enumerate features and map their action-intensity affordances
4. Single-Cycle Impact Modeling: Write and analyze cycle equations; test hypotheses with experiments or analytics
5. Temporal Evolution Analysis: Specify drift models (e.g., $e_j$5); design longitudinal studies
6. Update Rhythm Analysis: Compare actuator and sensor sampling rates; diagnose potential reinforcement instabilities

Application guidelines emphasize engineering friction at key points (interrupt infinite exposure), in-situ transparency, revision affordances, feedback monitoring, and the application of negative feedback when harmful drift is detected.

2.6. Advantages, Limitations, and Extensions

Advantages

• Dynamically captures feedback loops imperceptible to static audits
• Provides rigorous, parametric, hypothesis-driven evaluation ($e_j$6, $e_j$7, $e_j$8, $e_j$9)
• Merges affordance theory with control system formalism for quantitative model-based critique

Limitations

• Stochastic output (e.g., LLM nondeterminism) and multi-user settings are not directly modeled
• High-dimensional latent states $0$0 can challenge tractability
• Primarily models single-user–system interactions

Future Work

• Empirical time-series fitting for system identification ($0$1, $0$2, $0$3, $0$4)
• Extensions to stochastic (noise-injected) and multi-agent (networked users) formulations
• Deployment in varied domains and interface modalities (physical systems, voice, robotics)
• Aggregation to group-level dynamics for collective welfare optimization

These perspectives suggest DECAI transforms the critique of interface-driven AI harms into a quantitative discipline, enabling principled intervention and risk monitoring in adaptive AI deployments (Ibrahim et al., 2024).