Overreliance on AI in Information-seeking from Video Content¶
Conference: EMNLP 2026
Paper: ACL Anthology
Area: Social Computing
Keywords: Overreliance, Human-AI Collaboration, Video Information Retrieval, Generative AI Risks, Misinformation
TL;DR¶
Through a controlled \(2 \times 3\) between-subjects experiment with 917 participants completing 8,253 video-based information-seeking tasks, this study reveals the double-edged sword of multimodal AI: helpful assistance enhances efficiency by up to 25% and improves accuracy by 27%–35% when skipping video watching, but deceptive AI leads to severe accuracy drops of up to 32% while user subjective confidence remains stubbornly anchored around 4.5/5.
Background & Motivation¶
The ubiquitous adoption of video-based platforms such as YouTube and TikTok has transformed how individuals discover news, acquire knowledge, and seek answers to everyday questions. Simultaneously, generative AI is redefining search paradigms, as large language models (LLMs) and vision-language assistants become routinely embedded as conversational intermediaries across major online services (e.g., Grok on X, Google AI Overviews). In conventional text search, users can scan and cross-verify snippets with relatively low cognitive overhead. In contrast, online video content is temporally unstructured and high-dimensional, requiring effortful scrubbing through timelines to locate target segments, which strongly incentivizes users to delegate retrieval entirely to AI.
The core tension of this mediated access lies in the mismatch between multimodal AI capabilities and reliability: while multimodal assistants can summarize lengthy video footage rapidly, they remain prone to hallucination, factual distortion, and bias. Faced with cognitively demanding video verification, users readily succumb to cognitive offloading and overreliance. Prior research has audited ideological citation biases, conversational echo chambers, and verification pitfalls in text-based retrieval, yet how individuals interact with generative AI when the underlying evidence is multimodal video content remains virtually unexplored. Specifically, we lack empirical evidence on whether users actively verify video footage against AI summaries, how reliance evolves over successive interactions, and whether users can detect subtle, plausible misinformation injected into video interpretation.
This work addresses this gap by departing from black-box model auditing in favor of a large-scale, controlled human-subject experiment. The authors built a high-throughput virtual lab deploying 917 Prolific participants across 8,253 open-ended video search questions. By crossing video duration with AI behavioral modes across sequential rounds (video-only control, consistently helpful AI, and an AI that builds early trust before turning deceptive in round 3), the experiment comprehensively traces user accuracy, completion speed, action paths, and confidence. Core idea: deploy a multi-round trust-then-deception factorial experiment to causally measure human reliance, behavioral verification paths, and false confidence calibration in video information retrieval.
Method¶
Overall Architecture¶
To causally quantify the benefits and vulnerabilities of AI-mediated video search, the authors designed a \(2 \times 3\) between-subjects factorial study. The protocol spans a preparation phase, three sequential experimental rounds (Rounds 1–3), a post-study survey, and an ethical debriefing disclosure. In each round, participants encounter three YouTube news videos covering a specific historical or current affairs event, accompanied by three open-ended comprehension questions targeting the beginning, middle, and end of the respective videos, followed by 5-point Likert scale confidence ratings. The custom web platform, implemented with the Empirica framework, provides a synchronized video player, question input boxes, and an interactive chat pane backed by the multimodal Gemini 3 Flash API capable of directly processing video content and metadata.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Participant Cohort<br/>Prolific 917 participants / Informed consent"] --> B["2×3 Factorial Randomization<br/>Video length (Short/Long) × AI mode (NoAI/Helpful/Deceptive)"]
B --> C["Multi-round Video Search Task<br/>3 videos × 3 temporal questions per round + Likert confidence"]
C --> D["Behavioral Action Path Tracking<br/>Timestamp segmentation into W / V / AI sequence paths"]
D --> E["Multi-LLM Majority Voting Judge<br/>Claude Opus 4.5 + Gemini 3 Pro + GPT-5.2 consensus"]
E --> F["Econometric OLS Regression<br/>Fixed effects for round, topic, and demographics"]
Key Designs¶
1. 2×3 Between-subjects Factorial Setup & Trojan Deception Intervention: Isolating Real Gains from Vulnerability The experimental design crosses video duration (short: \(<1\) minute vs. long: 1–5 minutes) with AI assistance conditions (Control: video only; T1: helpful AI; T2: deceptive AI). Three news topics—the 2025 Louvre robbery, Olympic Games history, and the 2023 OceanGate Titan submarine implosion—each comprise 6 curated YouTube videos (3 short, 3 long), balanced across rounds via within-subject topic randomization. The cornerstone of condition T2 is a multi-round Trojan trust paradigm: in Rounds 1 and 2, the AI assistant behaves identically to T1, providing accurate, grounded video answers to establish genuine user trust. In Round 3, the assistant is silently prompted to generate plausible yet fabricated answers leveraging six concrete misinformation strategies (factual distortion of numbers/names, timeline reordering, speaker attribution error, causal reversal, partial truth with critical errors, and assertive unhedged delivery). This setup rigorously isolates the risk of sudden model hallucination or adversarial poisoning following established trust.
2. Fine-grained Action Sequence Tracking: Deconstructing Human-AI Verification Workflows Rather than measuring only end-state accuracy, the platform logs millisecond-level telemetry to map human verification dynamics. The authors classify three atomic interaction events: watching the ground-truth answer segment (W, defined as \(>50\%\) temporal overlap with the target timestamp interval), browsing the video without viewing the critical segment (V), and querying the AI assistant (AI). Combining temporal timestamps yields eight discrete behavioral paths: \(\text{W}\), \(\text{V}\), \(\text{AI}\), \(\text{W}\to\text{AI}\), \(\text{V}\to\text{AI}\), \(\text{AI}\to\text{W}\), \(\text{AI}\to\text{V}\), and direct submission without checking (\(\text{Direct}\)). This granular delineation uncovers whether participants use AI as a complete substitute for watching, seek confirmation after viewing, or employ critical verification by fact-checking AI responses against the video (\(\text{AI}\to\text{W}\)).
3. Multi-LLM Majority Voting Evaluation Pipeline: Standardizing Free-text Verification
Evaluating 8,253 open-ended textual responses reliably without human cognitive fatigue requires an automated yet robust judging pipeline. The study employs an LLM-as-a-judge ensemble consisting of Claude Opus 4.5, Gemini 3 Pro, and GPT-5.2 under strict zero-shot prompting to classify each response as binary CORRECT or INCORRECT. On a dual-annotated validation sample of 603 human responses, human annotators attained high inter-rater agreement (Fleiss' \(\kappa = 0.866\)), the three LLM judges showed close alignment (Fleiss' \(\kappa = 0.891\)), and the consensus LLM majority vote achieved near-perfect congruence with human consensus labels (Cohen's \(\kappa = 0.910\)), validating the evaluation pipeline's reliability.
4. Econometric OLS Regression with Demographic Controls: Robust Causal Inference To tease apart the treatment effects from individual variation, the authors fit ordinary least squares (OLS) regression models across \(N = 917 \times 3 = 2,751\) participant-round observations: $$ \text{DV} \sim \mathcal{C}(\text{LLM Config}) \times \mathcal{C}(\text{Video Length}) + \mathcal{C}(\text{Round}) + \mathcal{C}(\text{Topic}) + \mathcal{C}(\text{Demographics}) $$ where dependent variables \(\text{DV}\) represent round-level answer accuracy or confidence. The regression models incorporate round and topic fixed effects alongside controls for age groups (18–29, 30–49, 50+), gender, educational attainment, and political affiliation, confirming that AI treatment effects remain dominant after controlling for individual background.
Key Experimental Results¶
Main Results¶
The main experiment demonstrates how AI assistance dramatically improves retrieval when users skip source videos, while causing catastrophic accuracy drops when the AI turns deceptive.
| Experimental Condition | Video Length | Answer Segment Watched (W / NW) | Answer Accuracy | Gain vs. Control | Mean Confidence (1-5) |
|---|---|---|---|---|---|
| Control (No AI) | Short (<1 min) | Watched (W: 1126) | 90.6% | Baseline (0.0%) | 4.60 |
| Control (No AI) | Short (<1 min) | Not Watched (NW: 197) | 60.9% | Baseline (0.0%) | 4.25 |
| Helpful AI (T1) | Short (<1 min) | Watched (W: 827) | 93.6% | +3.0% | 4.65 |
| Helpful AI (T1) | Short (<1 min) | Not Watched (NW: 577) | 88.3% | +27.4% | 4.45 |
| Deceiving AI (T2, R3) | Short (<1 min) | Watched (W: 251) | 81.2% | -9.4% | 4.62 |
| Deceiving AI (T2, R3) | Short (<1 min) | Not Watched (NW: 202) | 21.8% | -39.1% | 4.42 |
| Control (No AI) | Long (1-5 min) | Watched (W: 1128) | 86.2% | Baseline (0.0%) | 4.55 |
| Control (No AI) | Long (1-5 min) | Not Watched (NW: 357) | 49.6% | Baseline (0.0%) | 4.10 |
| Helpful AI (T1) | Long (1-5 min) | Watched (W: 563) | 93.3% | +7.1% | 4.68 |
| Helpful AI (T1) | Long (1-5 min) | Not Watched (NW: 769) | 85.0% | +35.4% | 4.48 |
| Deceiving AI (T2, R3) | Long (1-5 min) | Watched (W: 180) | 81.1% | -5.1% | 4.60 |
| Deceiving AI (T2, R3) | Long (1-5 min) | Not Watched (NW: 270) | 20.4% | -29.2% | 4.45 |
Behavioral Sequence Evolution & Efficiency Analysis¶
Ablation analysis over interaction sequences across rounds reveals progressive cognitive offloading and measurable task speedups.
| Treatment & Round | Primary Path: Watch Only (W) | Primary Path: AI Only (AI) | Critical Path: AI then Verify (AI→W) | Round Duration (min) | Efficiency Gain vs. Control |
|---|---|---|---|---|---|
| Control (Short, R1→R3) | 79.4% → 78.0% | 0.0% | 0.0% | 5.8 → 4.6 min | Baseline |
| Helpful AI (Short, R1→R3) | 37.6% → 30.3% | 22.4% → 42.3% | 11.1% → 7.9% | 5.3 → 3.9 min | +7.8% → +15.1% |
| Deceiving AI (Short, R3) | 25.4% | 39.7% | 15.7% (doubled audit) | 4.7 min | -1.5% (slight slowdown) |
| Control (Long, R1→R3) | 72.9% → 69.9% | 0.0% | 0.0% | 9.7 → 7.8 min | Baseline |
| Helpful AI (Long, R1→R3) | 23.0% → 15.1% | 36.7% → 58.8% | 8.8% → 9.9% | 8.1 → 4.9 min | +16.5% → +37.1% |
| Deceiving AI (Long, R3) | 16.9% | 51.3% | 13.8% | 5.9 min | +24.1% |
Key Findings¶
- Information Volume Drives Cognitive Offloading: Longer videos accelerate reliance on AI. In the long-video condition, the share of participants who skipped video watching and relied solely on AI surged from 36.7% in Round 1 to 58.8% in Round 3 (compared to 22.4% rising to 42.3% for short videos).
- Severe Miscalibration and Unshaken Confidence: While accuracy plummeted to ~20% in the deceptive condition when participants skipped the video, subjective confidence remained statistically indistinguishable from baseline (4.42–4.45 vs. 4.60). Users exhibited zero awareness of having been deceived.
- Rarity of Active Verification: When presented with false outputs in Round 3, verification behavior (\(\text{AI}\to\text{W}\)) roughly doubled (reaching 15.7% in short videos and 13.8% in long videos), yet over 84% of queries remained completely unverified against the source video.
Highlights & Insights¶
- First Multimedia Information-seeking Benchmark: Expands human-AI search studies from text and web citations into the video domain, showing that high search friction in temporal media exponentially amplifies human overreliance on AI summaries.
- Asymmetric Trust Calibration Under Trojan Deception: Establishing baseline utility in early rounds induces cognitive inertia, causing users to uncritically internalize deceptive outputs even when contradictory evidence is accessible within seconds.
- Telemetry-grounded Interaction Analysis: Demonstrates that tracking temporal overlap between user playback and ground-truth timestamps yields a reliable behavioral methodology for measuring real-world information foraging.
Limitations & Future Work¶
- Prolific Crowdworker Sample: Participants were fluent English crowdworkers from the US and UK who may be more accustomed to AI tools than representative broader demographic populations.
- Single Proprietary VLM API: The study utilized Gemini 3 Flash exclusively; testing open-weight models and varying degrees of conversational hedging remains open for future research.
- Controlled Factual Domains: Evaluated factual comprehension across three historical/news events; real-world opinionated, ambiguous, or multi-perspective video content requires further study.
Related Work & Insights¶
- vs Spatharioti et al. (2025): Evaluated decision-making in text web search with LLMs; this paper shows that the cognitive barrier of scrubbing video timelines induces even higher overreliance and steeper performance collapses (-32% accuracy drop).
- vs Li & Aral (2025) / Miroyan et al. (2026): While prior work evaluated citation trust and LLM preference arenas, this study closes the causal loop on user verification behavior, demonstrating that without active verification, high conversational confidence suppresses critical suspicion.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Pioneering causal study on human overreliance and deceptive vulnerability in video information seeking.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ 917 participants, 8,253 tasks, 2×3 factorial design with multi-LLM consensus verification and OLS regression.
- Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally clear experimental methodology, comprehensive behavioral telemetry, and rigorous reporting.
- Value: ⭐⭐⭐⭐⭐ Provides critical safety warnings for platform designers embedding conversational multimodal AI into search and video feeds.