RAISE-26 · Rutgers University · 2026

AI Cognitive Bias:
How Media Distorts
Our Understanding of AI

The narrative shift toward AI autonomy has already happened — we just haven't noticed. We analyzed 10,500 headlines to find out how.

10,500
AI Headlines
3,125
News Outlets
89
Days Analyzed
8
Bias Types
4
Novel Metrics
Scroll to explore
Project Resources

All Links in One Place

Everything you need to explore our research — code, presentation, profiles, and documentation.

The Project

Measuring How News Shapes Thinking About AI

Others analyze what media says. We analyze how it distorts thinking. Using NLP, cognitive psychology, and four original metrics, we quantify psychological manipulation in AI journalism.

Sadhana Vasanthakumar '28
Computer Science & Data Science — Statistics Track
Minor: Cognitive Science & Statistics
Krisha Patel '28
Data Science — Computer Science Track
Minor: Business Administration
Our Framework

Four Original Metrics

Cognitive Bias Detection

8 types of psychological biases in AI headlines — mental shortcuts that distort how readers understand AI.

Power Shift Index (PSI)

Quantifies who's framed as in control: humans or AI. PSI > 100 = AI framing dominates.

AI Anxiety Index (AAI)

Composite 6-dimension fear measure: negative sentiment, fear keywords, risk words, economic threat, health/safety, ethics signals.

The Invisible Human

Tracks deep human experiences — love, grief, spirituality, community — that AI coverage systematically ignores.

Finding #1

Over Half of Headlines Use Psychological Manipulation

News doesn't neutrally report on AI — it activates psychological triggers. 53.9% of headlines stack multiple biases simultaneously.

53.9%
Headlines with multiple biases
5,663
Out of 10,500 headlines
0.546
Avg bias intensity (0–1)
20.2%
Show extreme bias
Cognitive Bias Types in Headlines
Prevalence per bias type (n = 10,500)
Bias Load Distribution
Headlines with 0, 1, or multiple biases
Most prevalent: Anthropomorphism — treating AI as if it has human desires, feelings, or intentions (e.g. "AI wants to replace workers"). Found in 48.4% of high-bias headlines.
Bias Co-occurrence Network
Node size = prevalence · Line thickness = correlation strength · Hover for insight · Drag nodes

💡 Hover over nodes and connections for details · Drag to rearrange

Strongest link: Techno-Utopianism ↔ Control Loss (r = 0.85) — "AI will save us" routinely pairs with "We can't control AI," creating hope and helplessness simultaneously.
These aren't random word choices. Multiple emotional triggers are deliberately combined to maximize engagement — readers experience hope + fear + anxiety at once. This creates confusion, not clarity.
Fear Bias
Optimism Bias
Anthropomorphism
Techno-Utopianism
Economic Threat
Control Loss
Moral Panic
Agency Attribution
FEAR
"AI apocalypse threatens humanity" — triggers catastrophic thinking, activates fight-or-flight
OPTIMISM
"AI will solve all our problems" — uncritical enthusiasm, suppresses healthy skepticism
ANTHRO
"AI wants to replace workers" — frames AI as having human desires and intentions
CONTROL
"AI is spiraling beyond human control" — frames AI as unstoppable autonomous force
High Bias (AAI: 100)
"AI raises job security fears as 71% of Americans see risk of permanent losses"
Fear Bias: 1.0 · Economic Threat · Control Loss · Moral Panic
Neutral Reframe
"Study examines AI's potential effects on labor market participation"
Bias Score: 0.08 · Factual framing · No emotional triggers
Finding #2

AI Agency Already Dominates Media Narratives

The narrative shift isn't a future trend. It's already embedded in current coverage — for every headline showing human control, there are 1.45 showing AI agency.

145.5
Power Shift Index
12.7%
AI acts independently
5.5%
Humans control AI
1.45:1
AI vs human agency ratio
Agency Distribution
Who is portrayed as in control?
Power Shift Index Over Time
PSI = 100 is balance point · Our score: 145.5
A red flag: For every headline showing human control of AI, there are 1.45 showing AI agency. This primes the public to accept AI having independent decision-making power — and stops people from asking: WHO programmed it? WHO profits? WHO is accountable when it fails?
"We don't say electricity decides to power homes. But we DO say AI threatens jobs." — BiasHounds analysis, 2025 · The language of agency shapes the politics of accountability.
Finding #3

Anxiety Is Real, Moderate, and Precisely Targeted

AI coverage doesn't create mass panic — but it reliably channels fear toward specific, concentrated themes. The formula is: Jobs + Fear language + Loss of Control = Maximum Anxiety.

12.8
AI Anxiety Index (of 100)
13.2
Peak AAI (Aug 2025)
±13.9
Std deviation (wide spread)
0.46
Fear bias ↔ AAI correlation
What Predicts Higher AI Anxiety?
Pearson correlation coefficients (all significant at p < 0.001)
Predictorr with AAIVisual
Fear Bias0.456
Economic Threat0.285
Moral Panic0.249
Bias Intensity0.229
Control Loss0.173
Techno-Utopianism0.131
Optimism Bias0.097
Anthropomorphism0.074
Surprising finding: Power Shift Index ↔ Anxiety (r = 0.12). Portraying AI as powerful does not create anxiety. Anxiety comes from emotional language, not discussions of capability. We can talk about powerful AI without fear — if we use balanced framing.

The Formula: Jobs + Fear Language + Loss of Control = Maximum Anxiety

High-Anxiety Headlines

AAI: 45.2
"AI threatens 40% of jobs as automation accelerates"
AAI: 38.7
"Workers fear AI replacement as layoffs surge"
AAI: 36.1
"Experts warn: AI development spiraling out of control"

Low-Anxiety Headlines

AAI: 2.1
"Hospital implements AI scheduling system"
AAI: 1.8
"University offers new AI certificate program"
Weekend vs. weekday: No significant difference — anxiety is in the language, not the timing. It's structural, not cyclical.
Finding #4

The Invisible Human — What News Never Asks

Media reduces AI to Work + Money + Efficiency, systematically ignoring how AI affects love, loss, faith, community, and what it means to be human.

0.57%
Spirituality & Meaning
"What gives life meaning when AI does our work?"
0.80%
Embodied Experience
"What AI can't touch, smell, or feel"
1.14%
Grief & Caregiving
"Technology in end-of-life care"
1.71%
Intimacy & Relationships
"How AI affects romantic connections"
24.30%
Artistic Process
Most covered — AI art generates controversy & clicks
Coverage of Human Experience Domains
% of headlines touching each dimension of human life (NOTE: the low percentages ARE the finding)
The erasure: Out of 10,500 headlines — 10,494 said nothing about ethical dilemmas (99.94%). 10,440 said nothing about spirituality (99.43%). 10,416 said nothing about grief (99.20%). These aren't gaps. They're erasures.
"If AI truly transforms humanity, why aren't we asking how it changes grief, love, meaning, or faith?" — BiasHounds research question, 2025
Technical Documentation

Methodology & Formulas

All models are open-weights. No proprietary AI. Fully reproducible with seed=42.

Cognitive Bias Detection

BiasScore = SemanticSimilarity(headline, prototype)
+ (0.20 × KeywordCount)

Normalized to [0,1] via MinMaxScaler

Uses sentence-transformers/all-MiniLM-L6-v2 for semantic similarity to 8 prototype bias descriptions. Custom 84-keyword lexicons per bias category. VADER sentiment analysis as supporting signal.

Power Shift Index (PSI)

PSI = [(AI_actor + AI_threat) /
(Human_actor + AI_tool)] × 100

PSI = 145.48 (our result)

Adapted from computational media studies agency-framing research. PSI = 100 is balance; above 100 means AI agency framing dominates. Keyword sets validated via semantic embeddings.

AI Anxiety Index (AAI)

AAI = [(2.5×NS + 3.5×FI + 2.5×RI
+ 2.0×ET + 1.5×HS + 1.0×ES)
/ max] × 100

NS=Negative sentiment, FI=Fear intensity (highest weight: 3.5), RI=Risk intensity, ET=Economic threat, HS=Health/safety, ES=Ethics signals. Composite 6-dimension weighted score.

Invisible Human Framework

Coverage% = (Headlines_mentioning_domain
/ Total_headlines) × 100

8 human life domains tracked

Keyword-based domain detection across 8 human experience categories: artistic process, ethics, work/employment, community, intimacy, grief, embodiment, spirituality. Validated with semantic similarity scoring.

Statistical Validation

All major correlations: p < 0.001
n = 10,500
Pearson r with significance tests

sciPy Pearson r, p-values. At n = 10,500, findings are not due to chance. Bias intensity ↔ AAI: r=0.64. Economic threat ↔ work topics: r=0.78. Techno-utopianism ↔ control loss: r=0.85 (strongest).

Technical Stack

Python 3.10+ · Google Colab
sentence-transformers · VADER
scikit-learn · NetworkX · Plotly

Also: numpy, pandas, matplotlib, scipy, MinMaxScaler. All models open-weights. Random seeds set (seed=42). Single-click execution via "Runtime → Run All." Runtime: 8–12 min on CPU.

Real-World Impact

Who This Affects

Biased AI coverage doesn't just shape individual opinions — it reverberates across society, policy, and industry.

The Public

  • News shapes AI opinions unconsciously
  • Emotional headlines trigger real psychological responses
  • Views form on narrow, distorted framing
  • Critical media literacy is now essential

Business & Tech Leaders

  • Public fear affects AI adoption rates
  • Biased coverage influences policy and regulation
  • Employee anxiety needs thoughtful communication
  • Responsible AI needs better public engagement

Journalists & Media

  • Current framing creates confusion, not clarity
  • Jobs/risk focus misses bigger human questions
  • Balanced coverage requires bias awareness
  • Ask: "What are we NOT covering?"

Policymakers & Educators

  • Media anxiety affects public policy support
  • Students need skills to recognize biased framing
  • AI education must address ethics and meaning
  • Fund media literacy programs
Evidence-Based Action

What Should Change

Grounded in data. Targeting the specific mechanisms that create distorted AI coverage.

For Media Organizations

  • Create editorial guidelines on cognitive bias awareness
  • Train reporters to recognize fear/optimism triggers
  • Expand coverage beyond work & efficiency narratives
  • Cover ethics, relationships, meaning-making, community
  • Develop "bias auditing" tools for headline review

For Educators

  • Teach students to identify cognitive biases in headlines
  • Include media literacy in all AI education programs
  • Use "Invisible Human" framework in ethics courses
  • Encourage critical comparison across news sources

For Policymakers

  • Monitor Power Shift Index trends over time
  • Fund independent, balanced AI journalism
  • Support media literacy initiatives in schools
  • Commission regular AI public perception studies

For Everyone

  • Read AI news from 3+ sources with different perspectives
  • Notice when headlines use fear or optimism language
  • Ask: "What human experience is this article ignoring?"
  • Remember: AI is built by humans, for human purposes

Note on Credibility Scores: Bias intensity varies 2–3× across different news sources in our dataset. Our Credibility Score (Score = 100 − Bias Intensity × 100) ranged from 35 to 65 across outlets — outlet choice matters significantly for balanced AI information.

Research Integrity

Limitations & Future Work

Data Constraints

  • 89-day period (Jun–Sep 2025) — cannot claim long-term trends
  • English-language sources only (cultural bias toward US/UK/Australia)
  • Primarily mainstream news outlets
  • Headlines only — not full article body text

Methodological Notes

  • Bias detection uses AI + keyword matching — directional findings, not clinical diagnosis
  • Thresholds (0.5 for "multiple biases") involve judgment calls
  • Results are robust across reasonable threshold ranges
  • "Invisible Human" detects absence — low coverage may mean "not newsworthy" vs. intentionally ignored

Future Research

  • Extend to 12+ month periods for trend analysis
  • Compare across languages and cultures
  • Track PSI as AI capabilities evolve over time
  • Study reader response to different framing types

AI Tool Disclosure

  • Tool: Claude 3.5 Sonnet (Anthropic)
  • Purpose: Syntax verification, debugging, documentation structure
  • Extent: ~15% of development
  • Human work: 85% — all conceptualization, analysis, interpretation
  • 13 prompts fully disclosed in presentation appendix